diff --git a/.github/ISSUE_TEMPLATE/bug_report.md b/.github/ISSUE_TEMPLATE/bug_report.md index 5df12aaf5..eb0fb9662 100644 --- a/.github/ISSUE_TEMPLATE/bug_report.md +++ b/.github/ISSUE_TEMPLATE/bug_report.md @@ -64,7 +64,7 @@ Try the following: 1. `git clone https://github.com/abetlen/llama-cpp-python` 2. `cd llama-cpp-python` 3. `rm -rf _skbuild/` # delete any old builds -4. `python setup.py develop` +4. `python -m pip install .` 5. `cd ./vendor/llama.cpp` 6. Follow [llama.cpp's instructions](https://github.com/ggerganov/llama.cpp#build) to `cmake` llama.cpp 7. Run llama.cpp's `./main` with the same arguments you previously passed to llama-cpp-python and see if you can reproduce the issue. If you can, [log an issue with llama.cpp](https://github.com/ggerganov/llama.cpp/issues) diff --git a/.github/dependabot.yml b/.github/dependabot.yml index 91abb11fd..6bf90273a 100644 --- a/.github/dependabot.yml +++ b/.github/dependabot.yml @@ -8,4 +8,12 @@ updates: - package-ecosystem: "pip" # See documentation for possible values directory: "/" # Location of package manifests schedule: - interval: "weekly" + interval: "daily" + - package-ecosystem: "github-actions" + directory: "/" + schedule: + interval: "daily" + - package-ecosystem: "docker" + directory: "/" + schedule: + interval: "daily" diff --git a/.github/workflows/build-and-release.yaml b/.github/workflows/build-and-release.yaml index 61027ef99..7307c85ab 100644 --- a/.github/workflows/build-and-release.yaml +++ b/.github/workflows/build-and-release.yaml @@ -11,34 +11,77 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-latest, windows-latest, macOS-latest] + os: [ubuntu-20.04, windows-2019, macos-13] steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: - submodules: "true" + submodules: "recursive" # Used to host cibuildwheel - - uses: actions/setup-python@v3 + - uses: actions/setup-python@v5 with: - python-version: "3.8" + python-version: "3.9" - - name: Install cibuildwheel - run: python -m pip install cibuildwheel==2.12.1 + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash - - name: Install dependencies + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | python -m pip install --upgrade pip - python -m pip install -e .[all] + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd - name: Build wheels - run: python -m cibuildwheel --output-dir wheelhouse + uses: pypa/cibuildwheel@v2.22.0 env: # disable repair CIBW_REPAIR_WHEEL_COMMAND: "" + with: + package-dir: . + output-dir: wheelhouse + + - uses: actions/upload-artifact@v4 + with: + name: wheels-${{ matrix.os }} + path: ./wheelhouse/*.whl + + build_wheels_arm64: + name: Build arm64 wheels + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v4 + with: + submodules: "recursive" + + - name: Set up QEMU + uses: docker/setup-qemu-action@v3 + with: + platforms: linux/arm64 + + - name: Build wheels + uses: pypa/cibuildwheel@v2.22.0 + env: + CIBW_SKIP: "*musllinux* pp*" + CIBW_REPAIR_WHEEL_COMMAND: "" + CIBW_ARCHS: "aarch64" + CIBW_BUILD: "cp38-* cp39-* cp310-* cp311-* cp312-*" + with: + output-dir: wheelhouse - - uses: actions/upload-artifact@v3 + - name: Upload wheels as artifacts + uses: actions/upload-artifact@v4 with: + name: wheels_arm64 path: ./wheelhouse/*.whl build_sdist: @@ -46,35 +89,56 @@ jobs: runs-on: ubuntu-latest steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: - submodules: "true" - - uses: actions/setup-python@v3 + submodules: "recursive" + + - uses: actions/setup-python@v5 with: - python-version: "3.8" - - name: Install dependencies + python-version: "3.9" + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + python -m uv pip install build + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | - python -m pip install --upgrade pip build - python -m pip install -e .[all] + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + python -m uv pip install build + shell: cmd + - name: Build source distribution run: | python -m build --sdist - - uses: actions/upload-artifact@v3 + + - uses: actions/upload-artifact@v4 with: + name: sdist path: ./dist/*.tar.gz release: name: Release - needs: [build_wheels, build_sdist] + needs: [build_wheels, build_wheels_arm64, build_sdist] runs-on: ubuntu-latest steps: - - uses: actions/download-artifact@v3 + - uses: actions/download-artifact@v4 with: - name: artifact + merge-multiple: true path: dist - - uses: softprops/action-gh-release@v1 + + - uses: softprops/action-gh-release@v2 with: files: dist/* env: - GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} \ No newline at end of file + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} diff --git a/.github/workflows/build-docker.yaml b/.github/workflows/build-docker.yaml index 27a6b1e50..b5c7346db 100644 --- a/.github/workflows/build-docker.yaml +++ b/.github/workflows/build-docker.yaml @@ -12,18 +12,18 @@ jobs: runs-on: ubuntu-latest steps: - name: Checkout - uses: actions/checkout@v3 + uses: actions/checkout@v4 with: - submodules: "true" + submodules: "recursive" - name: Set up QEMU - uses: docker/setup-qemu-action@v2 + uses: docker/setup-qemu-action@v3 - name: Set up Docker Buildx - uses: docker/setup-buildx-action@v2 + uses: docker/setup-buildx-action@v3 - name: Login to GitHub Container Registry - uses: docker/login-action@v2 + uses: docker/login-action@v3 with: registry: ghcr.io username: ${{ github.repository_owner }} @@ -31,7 +31,7 @@ jobs: - name: Build and push id: docker_build - uses: docker/build-push-action@v4 + uses: docker/build-push-action@v6 with: context: . file: "docker/simple/Dockerfile" diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml new file mode 100644 index 000000000..745b2e602 --- /dev/null +++ b/.github/workflows/build-wheels-cuda.yaml @@ -0,0 +1,136 @@ +name: Build Wheels (CUDA) + +on: workflow_dispatch + +permissions: + contents: write + +jobs: + define_matrix: + name: Define Build Matrix + runs-on: ubuntu-latest + outputs: + matrix: ${{ steps.set-matrix.outputs.matrix }} + defaults: + run: + shell: pwsh + + steps: + - name: Define Job Output + id: set-matrix + run: | + $matrix = @{ + 'os' = @('ubuntu-latest', 'windows-2019') + 'pyver' = @("3.9", "3.10", "3.11", "3.12") + 'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1") #, "12.5.1", "12.6.1") + 'releasetag' = @("basic") + } + + $matrixOut = ConvertTo-Json $matrix -Compress + Write-Output ('matrix=' + $matrixOut) >> $env:GITHUB_OUTPUT + + build_wheels: + name: Build Wheel ${{ matrix.os }} ${{ matrix.pyver }} ${{ matrix.cuda }} ${{ matrix.releasetag == 'wheels' && 'AVX2' || matrix.releasetag }} + needs: define_matrix + runs-on: ${{ matrix.os }} + strategy: + matrix: ${{ fromJSON(needs.define_matrix.outputs.matrix) }} + defaults: + run: + shell: pwsh + env: + CUDAVER: ${{ matrix.cuda }} + AVXVER: ${{ matrix.releasetag }} + + steps: + - name: Add MSBuild to PATH + if: runner.os == 'Windows' + uses: microsoft/setup-msbuild@v2 + with: + vs-version: '[16.11,16.12)' + + - uses: actions/checkout@v4 + with: + submodules: "recursive" + + - uses: actions/setup-python@v5 + with: + python-version: ${{ matrix.pyver }} + cache: 'pip' + + - name: Setup Mamba + uses: conda-incubator/setup-miniconda@v3.1.0 + with: + activate-environment: "llamacpp" + python-version: ${{ matrix.pyver }} + miniforge-version: latest + add-pip-as-python-dependency: true + auto-activate-base: false + + - name: VS Integration Cache + id: vs-integration-cache + if: runner.os == 'Windows' + uses: actions/cache@v4 + with: + path: ./MSBuildExtensions + key: cuda-${{ matrix.cuda }}-vs-integration + + - name: Get Visual Studio Integration + if: runner.os == 'Windows' && steps.vs-integration-cache.outputs.cache-hit != 'true' + run: | + if ($env:CUDAVER -eq '12.1.1') {$x = '12.1.0'} else {$x = $env:CUDAVER} + $links = (Invoke-RestMethod 'https://raw.githubusercontent.com/Jimver/cuda-toolkit/master/src/links/windows-links.ts').Trim().split().where({$_ -ne ''}) + for ($i=$q=0;$i -lt $links.count -and $q -lt 2;$i++) {if ($links[$i] -eq "'$x',") {$q++}} + Invoke-RestMethod $links[$i].Trim("'") -OutFile 'cudainstaller.zip' + & 'C:\Program Files\7-Zip\7z.exe' e cudainstaller.zip -oMSBuildExtensions -r *\MSBuildExtensions\* > $null + Remove-Item 'cudainstaller.zip' + + - name: Install Visual Studio Integration + if: runner.os == 'Windows' + run: | + $y = (gi '.\MSBuildExtensions').fullname + '\*' + (gi 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\MSBuild\Microsoft\VC\*\BuildCustomizations').fullname.foreach({cp $y $_}) + $cupath = 'CUDA_PATH_V' + $env:CUDAVER.Remove($env:CUDAVER.LastIndexOf('.')).Replace('.','_') + echo "$cupath=$env:CONDA_PREFIX" >> $env:GITHUB_ENV + + - name: Install Dependencies + env: + MAMBA_DOWNLOAD_FAILFAST: "0" + MAMBA_NO_LOW_SPEED_LIMIT: "1" + run: | + $cudaVersion = $env:CUDAVER + mamba install -y 'cuda' -c nvidia/label/cuda-$cudaVersion + python -m pip install build wheel + + - name: Build Wheel + run: | + $cudaVersion = $env:CUDAVER.Remove($env:CUDAVER.LastIndexOf('.')).Replace('.','') + $env:CUDA_PATH = $env:CONDA_PREFIX + $env:CUDA_HOME = $env:CONDA_PREFIX + $env:CUDA_TOOLKIT_ROOT_DIR = $env:CONDA_PREFIX + if ($IsLinux) { + $env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH + } + $env:VERBOSE = '1' + $env:CMAKE_ARGS = '-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=all' + $env:CMAKE_ARGS = "-DGGML_CUDA_FORCE_MMQ=ON $env:CMAKE_ARGS" + # if ($env:AVXVER -eq 'AVX') { + $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DGGML_AVX2=off -DGGML_FMA=off -DGGML_F16C=off' + # } + # if ($env:AVXVER -eq 'AVX512') { + # $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DGGML_AVX512=on' + # } + # if ($env:AVXVER -eq 'basic') { + # $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DGGML_AVX=off -DGGML_AVX2=off -DGGML_FMA=off -DGGML_F16C=off' + # } + python -m build --wheel + # write the build tag to the output + Write-Output "CUDA_VERSION=$cudaVersion" >> $env:GITHUB_ENV + + - uses: softprops/action-gh-release@v2 + with: + files: dist/* + # Set tag_name to -cu + tag_name: ${{ github.ref_name }}-cu${{ env.CUDA_VERSION }} + env: + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} diff --git a/.github/workflows/build-wheels-metal.yaml b/.github/workflows/build-wheels-metal.yaml new file mode 100644 index 000000000..9b97bf2f5 --- /dev/null +++ b/.github/workflows/build-wheels-metal.yaml @@ -0,0 +1,79 @@ +name: Build Wheels (Metal) + +on: workflow_dispatch + +permissions: + contents: write + +jobs: + build_wheels: + name: Build wheels on ${{ matrix.os }} + runs-on: ${{ matrix.os }} + strategy: + matrix: + os: [macos-13, macos-14, macos-15] + + steps: + - uses: actions/checkout@v4 + with: + submodules: "recursive" + + # Used to host cibuildwheel + - uses: actions/setup-python@v5 + with: + python-version: "3.12" + cache: 'pip' + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace + run: | + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd + + - name: Build wheels + uses: pypa/cibuildwheel@v2.22.0 + env: + # disable repair + CIBW_REPAIR_WHEEL_COMMAND: "" + CIBW_ARCHS: "arm64" + CIBW_ENVIRONMENT: CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DGGML_METAL=on" + CIBW_BUILD: "cp39-* cp310-* cp311-* cp312-*" + with: + package-dir: . + output-dir: wheelhouse2 + + - uses: actions/upload-artifact@v4 + with: + name: wheels-mac_${{ matrix.os }} + path: ./wheelhouse2/*.whl + + release: + name: Release + needs: [build_wheels] + runs-on: ubuntu-latest + + steps: + - uses: actions/download-artifact@v4 + with: + merge-multiple: true + path: dist2 + + - uses: softprops/action-gh-release@v2 + with: + files: dist2/* + # set release name to -metal + tag_name: ${{ github.ref_name }}-metal + env: + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} diff --git a/.github/workflows/generate-index-from-release.yaml b/.github/workflows/generate-index-from-release.yaml new file mode 100644 index 000000000..255ee67d6 --- /dev/null +++ b/.github/workflows/generate-index-from-release.yaml @@ -0,0 +1,57 @@ +name: Wheels Index + +on: + # Trigger on new release + workflow_run: + workflows: ["Release", "Build Wheels (CUDA)", "Build Wheels (Metal)"] + types: + - completed + + # Allows you to run this workflow manually from the Actions tab + workflow_dispatch: + +# Sets permissions of the GITHUB_TOKEN to allow deployment to GitHub Pages +permissions: + contents: read + pages: write + id-token: write + +# Allow only one concurrent deployment, skipping runs queued between the run in-progress and latest queued. +# However, do NOT cancel in-progress runs as we want to allow these production deployments to complete. +concurrency: + group: "pages" + cancel-in-progress: false + +jobs: + # Single deploy job since we're just deploying + deploy: + environment: + name: github-pages + url: ${{ steps.deployment.outputs.page_url }} + runs-on: ubuntu-latest + steps: + - name: Checkout + uses: actions/checkout@v4 + - name: Setup Pages + uses: actions/configure-pages@v5 + - name: Build + env: + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} + run: | + ./scripts/get-releases.sh + ./scripts/releases-to-pep-503.sh index/whl/cpu '^[v]?[0-9]+\.[0-9]+\.[0-9]+$' + ./scripts/releases-to-pep-503.sh index/whl/cu121 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu121$' + ./scripts/releases-to-pep-503.sh index/whl/cu122 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu122$' + ./scripts/releases-to-pep-503.sh index/whl/cu123 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu123$' + ./scripts/releases-to-pep-503.sh index/whl/cu124 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' + # ./scripts/releases-to-pep-503.sh index/whl/cu125 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' + # ./scripts/releases-to-pep-503.sh index/whl/cu126 '^[v]?[0-9]+\.[0-9]+\.[0-9]+-cu124$' + ./scripts/releases-to-pep-503.sh index/whl/metal '^[v]?[0-9]+\.[0-9]+\.[0-9]+-metal$' + - name: Upload artifact + uses: actions/upload-pages-artifact@v3 + with: + # Upload entire repository + path: 'index' + - name: Deploy to GitHub Pages + id: deployment + uses: actions/deploy-pages@v4 diff --git a/.github/workflows/publish-to-test.yaml b/.github/workflows/publish-to-test.yaml index 9932d6130..de3ae42aa 100644 --- a/.github/workflows/publish-to-test.yaml +++ b/.github/workflows/publish-to-test.yaml @@ -16,28 +16,47 @@ jobs: runs-on: ubuntu-latest steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: - submodules: "true" + submodules: "recursive" + - name: Set up Python - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: - python-version: "3.8" + python-version: "3.11" + cache: 'pip' + - name: Append Dev Version to __version__ run: | DEV_VERSION=${{ github.event.inputs.dev_version }} CURRENT_VERSION=$(awk -F= '/__version__ =/ {print $2}' llama_cpp/__init__.py | tr -d ' "') NEW_VERSION="${CURRENT_VERSION}.dev${DEV_VERSION}" sed -i 's/__version__ = \".*\"/__version__ = \"'"${NEW_VERSION}"'\"/' llama_cpp/__init__.py - - name: Install dependencies + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' run: | - python3 -m pip install --upgrade pip build - python3 -m pip install -e .[all] + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace + run: | + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd + - name: Build source distribution run: | - python3 -m build --sdist + python -m build --sdist + - name: Publish to Test PyPI uses: pypa/gh-action-pypi-publish@release/v1 with: password: ${{ secrets.TEST_PYPI_API_TOKEN }} - repository-url: https://test.pypi.org/legacy/ \ No newline at end of file + repository-url: https://test.pypi.org/legacy/ diff --git a/.github/workflows/publish.yaml b/.github/workflows/publish.yaml index 7d6c97048..bb76f5394 100644 --- a/.github/workflows/publish.yaml +++ b/.github/workflows/publish.yaml @@ -10,20 +10,39 @@ jobs: runs-on: ubuntu-latest steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: - submodules: "true" + submodules: "recursive" + - name: Set up Python - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: - python-version: "3.8" - - name: Install dependencies + python-version: "3.9" + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install -e .[all] --verbose + python -m uv pip install build + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | - python3 -m pip install --upgrade pip build - python3 -m pip install -e .[all] + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + python -m uv pip install build + shell: cmd + - name: Build source distribution run: | - python3 -m build --sdist + python -m build --sdist + - name: Publish distribution to PyPI # TODO: move to tag based releases # if: startsWith(github.ref, 'refs/tags') diff --git a/.github/workflows/test-pypi.yaml b/.github/workflows/test-pypi.yaml index cc6a3a725..335033bba 100644 --- a/.github/workflows/test-pypi.yaml +++ b/.github/workflows/test-pypi.yaml @@ -8,57 +8,105 @@ jobs: runs-on: ubuntu-latest strategy: matrix: - python-version: ["3.7", "3.8", "3.9", "3.10", "3.11"] + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - - name: Install dependencies + cache: 'pip' + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install llama-cpp-python[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | - python3 -m pip install --upgrade pip - python3 -m pip install --verbose llama-cpp-python[all] + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install llama-cpp-python[all] --verbose + shell: cmd + - name: Test with pytest run: | - python3 -c "import llama_cpp" + python -c "import llama_cpp" build-windows: runs-on: windows-latest strategy: matrix: - python-version: ["3.7", "3.8", "3.9", "3.10", "3.11"] + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - - name: Install dependencies + cache: 'pip' + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install llama-cpp-python[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | - python3 -m pip install --upgrade pip - python3 -m pip install --verbose llama-cpp-python[all] + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install llama-cpp-python[all] --verbose + shell: cmd + - name: Test with pytest run: | - python3 -c "import llama_cpp" + python -c "import llama_cpp" build-macos: runs-on: macos-latest strategy: matrix: - python-version: ["3.7", "3.8", "3.9", "3.10", "3.11"] + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - - name: Install dependencies + cache: 'pip' + + - name: Install dependencies (Linux/MacOS) + if: runner.os != 'Windows' + run: | + python -m pip install --upgrade pip + python -m pip install uv + RUST_LOG=trace python -m uv pip install llama-cpp-python[all] --verbose + shell: bash + + - name: Install dependencies (Windows) + if: runner.os == 'Windows' + env: + RUST_LOG: trace run: | - python3 -m pip install --upgrade pip - python3 -m pip install --verbose llama-cpp-python[all] + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install llama-cpp-python[all] --verbose + shell: cmd + - name: Test with pytest run: | - python3 -c "import llama_cpp" \ No newline at end of file + python -c "import llama_cpp" diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 2cc6fb02d..95f6e5a27 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -1,5 +1,4 @@ name: Tests - on: pull_request: branches: @@ -8,119 +7,165 @@ on: branches: - main +env: + REPO_ID: Qwen/Qwen2-0.5B-Instruct-GGUF + MODEL_FILE: qwen2-0_5b-instruct-q8_0.gguf + jobs: - build-linux: + download-model: + runs-on: ubuntu-latest + steps: + - name: Set up Python + uses: actions/setup-python@v5 + with: + python-version: "3.9" + - name: Install huggingface-hub + run: pip install huggingface-hub + - name: Download model + run: huggingface-cli download ${{ env.REPO_ID }} ${{ env.MODEL_FILE }} + - name: Cache model + uses: actions/cache@v4 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} + build-linux: + needs: download-model runs-on: ubuntu-latest strategy: matrix: - python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] - + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - uses: actions/checkout@v4 with: - submodules: "true" + submodules: "recursive" + - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - - name: Install dependencies + cache: 'pip' + - name: Restore model cache + uses: actions/cache@v4 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} + - name: Install dependencies (Linux/MacOS) run: | - python3 -m pip install --upgrade pip - python3 -m pip install .[all] -v + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: bash - name: Test with pytest run: | - python3 -m pytest + python -m pytest build-windows: - + needs: download-model runs-on: windows-latest strategy: matrix: - python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] - + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: - submodules: "true" + submodules: "recursive" + - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - - name: Install dependencies + cache: 'pip' + + - name: Restore model cache + uses: actions/cache@v4 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} + + - name: Install dependencies (Windows) run: | - python3 -m pip install --upgrade pip - python3 -m pip install .[all] -v + python -m pip install --upgrade pip + python -m pip install uv + python -m uv pip install -e .[all] --verbose + shell: cmd + - name: Test with pytest run: | - python3 -m pytest + python -m pytest build-macos: - - runs-on: macos-latest + needs: download-model + runs-on: macos-13 strategy: matrix: - python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] - + python-version: ["3.9", "3.10", "3.11", "3.12"] steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: - submodules: "true" + submodules: "recursive" + - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 + uses: actions/setup-python@v5 with: python-version: ${{ matrix.python-version }} - - name: Install dependencies + cache: 'pip' + + - name: System Info + run: | + uname -a + sysctl -n machdep.cpu.brand_string + python3 -c "import platform; print(platform.machine(), platform.architecture())" + + - name: Restore model cache + uses: actions/cache@v4 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} + + - name: Install dependencies (Linux/MacOS) run: | python3 -m pip install --upgrade pip - python3 -m pip install .[all] --verbose + python3 -m pip install uv + python3 -m uv pip install -e .[all] --verbose + CMAKE_ARGS="-DLLAMA_METAL=off" python3 -m uv pip install .[all] --verbose + shell: bash + - name: Test with pytest run: | python3 -m pytest - # build-linux-opencl: - - # runs-on: ubuntu-latest - - # steps: - # - uses: actions/checkout@v3 - # with: - # submodules: "true" - # - name: Set up Python 3.8 - # uses: actions/setup-python@v4 - # with: - # python-version: "3.8" - # - name: Set up OpenCL & CLBlast - # run: | - # wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | sudo tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null - # echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneAPI.list - # sudo apt-get update - # sudo apt-get install -y --no-install-recommends llvm intel-oneapi-runtime-opencl intel-oneapi-runtime-compilers libclblast-dev - # - name: Install dependencies - # run: | - # python3 -m pip install --upgrade pip - # CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install .[all] --verbose - # - name: Test with pytest - # run: | - # python3 -m pytest - - build-macos-metal: - - runs-on: macos-latest - + needs: download-model + runs-on: macos-13 steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 with: - submodules: "true" - - name: Set up Python 3.8 - uses: actions/setup-python@v4 + submodules: "recursive" + + - name: Set up Python 3.9 + uses: actions/setup-python@v5 with: - python-version: "3.8" + python-version: "3.9" + + - name: System Info + run: | + uname -a + sysctl -n machdep.cpu.brand_string + python3 -c "import platform; print(platform.machine(), platform.architecture())" + + - name: Restore model cache + uses: actions/cache@v4 + with: + path: ~/.cache/huggingface/hub + key: ${{ runner.os }}-model-${{ env.REPO_ID }}-${{ env.MODEL_FILE }} + - name: Install dependencies run: | python3 -m pip install --upgrade pip CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install .[all] --verbose + shell: bash + - name: Test with pytest run: | python3 -m pytest diff --git a/.gitignore b/.gitignore index 51f357200..9d68dbcd9 100644 --- a/.gitignore +++ b/.gitignore @@ -1,3 +1,5 @@ +*.local + .python-version .vscode/ diff --git a/CHANGELOG.md b/CHANGELOG.md index 97404bf92..affbd5db7 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,506 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.9] + +- feat: Update llama.cpp to ggerganov/llama.cpp@8733e0cf6eefc7c7752297cc22d0836706f4222c + +## [0.3.8] + +- feat: Update llama.cpp to ggerganov/llama.cpp@7841fc723e059d1fd9640e5c0ef19050fcc7c698 + +## [0.3.7] + +- feat: Update llama.cpp to ggerganov/llama.cpp@794fe23f29fb40104975c91fe19f23798f7c726e +- fix(ci): Fix the CUDA workflow by @oobabooga in #1894 +- fix: error showing time spent in llama perf context print, adds `no_perf` flag to `Llama` class by @shakalaca in #1898 + +## [0.3.6] + +- feat: Update llama.cpp to ggerganov/llama.cpp@f7cd13301c2a88f97073fd119072b4cc92c08df1 +- fix(server): streaming resource lock by @gjpower in #1879 + +## [0.3.5] + +- feat: Update llama.cpp to ggerganov/llama.cpp@26a8406ba9198eb6fdd8329fa717555b4f77f05f +- fix(ci): Fix release by updating macos runner image to non-deprecated version by @abetlen in afedfc888462f9a6e809dc9455eb3b663764cc3f +- fix(server): add missing await statements for async exit_stack handling by @gjpower in #1858 + +## [0.3.4] + +- fix(ci): Build wheels for macos 13-15, cuda 12.1-12.4 by @abetlen in ca808028bd16b8327bd84128d48015a4b1304690 + +## [0.3.3] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ce8784bdb153ff7794dde5a50b0ebfa51baa6171 +- fix: chat API logprobs format by @domdomegg in #1788 +- feat: Add support for CUDA 12.6, fix CUDA 12.5 by @Smartappli in #1775 +- fix: Make content not required in ChatCompletionRequestAssistantMessage by @feloy in #1807 +- fix: Fix pickling of Llama class by setting seed from _seed member by @abetlen in 2523472c3eccb9ab9277117cc4ff705212b6888a +- fix: Fix logit-bias type hint by @ddh0 in #1802 +- fix(server): Avoid thread starvation on many concurrent requests by making use of asyncio to lock llama_proxy context by @gjpower in #1798 +- fix(server): Added missing exit_stack.close() to /v1/chat/completions by @Ian321 in #1796 +- fix(examples): Refactor Batching notebook to use new sampler chain API by @lukestanley in #1793 +- fix(docs): Update development instructions by @Florents-Tselai in #1833 +- fix(docs): Remove ref to llama_eval in llama_cpp.py docs by @richdougherty in #1819 + +## [0.3.2] + +- feat: Update llama.cpp to ggerganov/llama.cpp@74d73dc85cc2057446bf63cc37ff649ae7cebd80 + +## [0.3.1] + +- feat: Update llama.cpp to ggerganov/llama.cpp@c919d5db39c8a7fcb64737f008e4b105ee0acd20 +- feat: Expose libggml in internal APIs by @abetlen in #1761 +- fix: Fix speculative decoding by @abetlen in 9992c5084a3df2f533e265d10f81d4269b97a1e6 and e975dabf74b3ad85689c9a07719cbb181313139b +- misc: Rename all_text to remaining_text by @xu-song in #1658 + +## [0.3.0] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ea9c32be71b91b42ecc538bd902e93cbb5fb36cb +- feat: Enable detokenizing special tokens with special=True by @benniekiss in #1596 +- feat(ci): Speed up CI workflows using uv, add support for CUDA 12.5 wheels by @Smartappli in e529940f45d42ed8aa31334123b8d66bc67b0e78 +- feat: Add loading sharded GGUF files from HuggingFace with Llama.from_pretrained(additional_files=[...]) by @Gnurro in 84c092063e8f222758dd3d60bdb2d1d342ac292e +- feat: Add option to configure n_ubatch by @abetlen in 6c44a3f36b089239cb6396bb408116aad262c702 +- feat: Update sampling API for llama.cpp. Sampling now uses sampler chain by @abetlen in f8fcb3ea3424bcfba3a5437626a994771a02324b +- fix: Don't store scores internally unless logits_all=True. Reduces memory requirements for large context by @abetlen in 29afcfdff5e75d7df4c13bad0122c98661d251ab +- fix: Fix memory allocation of ndarray in by @xu-song in #1704 +- fix: Use system message in og qwen format by @abetlen in 98eb092d3c6e7c142c4ba2faaca6c091718abbb3 + + +## [0.2.90] + +- feat: Update llama.cpp to ggerganov/llama.cpp@1d1ccce67613674c75c9c7e3fa4c1e24e428ba48 +- feat: Add support for `MiniCPMv26ChatHandler` and `minicpm-v-26` in server by @abetlen in f70df824985d875226793b94dacc0c302a4256b2 + +## [0.2.89] + +- feat: Update llama.cpp to ggerganov/llama.cpp@cfac111e2b3953cdb6b0126e67a2487687646971 +- fix: Llama.close didn't free lora adapter by @jkawamoto in #1679 +- fix: missing dependencies for test by @jkawamoto in #1680 + +## [0.2.88] + +- feat: Update llama.cpp to ggerganov/llama.cpp@fc4ca27b25464a11b3b86c9dbb5b6ed6065965c2 +- fix: only print 'cache saved' in verbose mode by @lsorber in #1668 +- fix: Added back from_file method to LlamaGrammar by @ExtReMLapin in #1673 +- fix: grammar prints on each call by @abetlen in 0998ea0deea076a547d54bd598d6b413b588ee2b +- feat: Enable recursive search of HFFS.ls when using from_pretrained by @benHeidabetlen in #1656 +- feat: Add more detailed log for prefix-match by @xu-song in #1659 + +## [0.2.87] + +- feat: Update llama.cpp to ggerganov/llama.cpp@be55695eff44784a141a863f273661a6bce63dfc +- fix: Include all llama.cpp source files and subdirectories by @abetlen in 9cad5714ae6e7c250af8d0bbb179f631368c928b +- feat(ci): Re-build wheel index automatically when releases are created by @abetlen in 198f47dc1bd202fd2b71b29e041a9f33fe40bfad + +## [0.2.86] + +- feat: Update llama.cpp to ggerganov/llama.cpp@398ede5efeb07b9adf9fbda7ea63f630d476a792 +- feat: Ported back new grammar changes from C++ to Python implementation by @ExtReMLapin in (#1637) +- fix: llama_grammar_accept_token arg order by @tc-wolf in (#1649) + +## [0.2.85] + +- feat: Update llama.cpp to ggerganov/llama.cpp@398ede5efeb07b9adf9fbda7ea63f630d476a792 +- fix: Missing LoRA adapter after API change by @shamitv in #1630 +- fix(docker): Update Dockerfile BLAS options by @olivierdebauche in #1632 +- fix(docker): Fix GGML_CUDA param by @olivierdebauche in #1633 +- fix(docker): Update Dockerfile build options from `LLAMA_` to `GGML_` by @olivierdebauche in #1634 +- feat: FreeBSD compatibility by @yurivict in #1635 + +## [0.2.84] + +- feat: Update llama.cpp to ggerganov/llama.cpp@4730faca618ff9cee0780580145e3cbe86f24876 +- fix: fix: Correcting run.sh filepath in Simple Docker implementation by @mashuk999 in #1626 + +## [0.2.83] + +- feat: Update llama.cpp to ggerganov/llama.cpp@081fe431aa8fb6307145c4feb3eed4f48cab19f8 +- feat: Add 'required' literal to ChatCompletionToolChoiceOption by @mjschock in #1597 +- fix: Change repeat_penalty to 1.0 to match llama.cpp defaults by @ddh0 in #1590 +- fix(docs): Update README.md typo by @ericcurtin in #1589 +- fix(server): Use split_mode from model settings by @grider-withourai in #1594 +- feat(ci): Dockerfile update base images and post-install cleanup by @Smartappli in #1530 + +## [0.2.82] + +- feat: Update llama.cpp to ggerganov/llama.cpp@7fdb6f73e35605c8dbc39e9f19cd9ed84dbc87f2 + +## [0.2.81] + +- feat: Update llama.cpp to ggerganov/llama.cpp@968967376dc2c018d29f897c4883d335bbf384fb +- fix(ci): Fix CUDA wheels, use LLAMA_CUDA instead of removed LLAMA_CUBLAS by @abetlen in 4fb6fc12a02a68884c25dd9f6a421cacec7604c6 +- fix(ci): Fix MacOS release, use macos-12 image instead of removed macos-11 by @abetlen in 3a551eb5263fdbd24b36d7770856374c04e92788 + +## [0.2.80] + +- feat: Update llama.cpp to ggerganov/llama.cpp@023b8807e10bc3ade24a255f01c1ad2a01bb4228 +- fix(server): Fix bug in FastAPI streaming response where dependency was released before request completes causing SEGFAULT by @abetlen in 296304b60bb83689659883c9cc24f4c074dd88ff +- fix(server): Update default config value for embeddings to False to fix error in text generation where logits were not allocated by llama.cpp by @abetlen in bf5e0bb4b151f4ca2f5a21af68eb832a96a79d75 +- fix(ci): Fix the CUDA workflow by @oobabooga in #1551 +- docs: Update readme examples to use newer Qwen2 model by @jncraton in #1544 + +## [0.2.79] + +- feat: Update llama.cpp to ggerganov/llama.cpp@9c77ec1d74874ee22bdef8f110e8e8d41389abf2 +- feat(ci): Update workflows and pre-built wheels by @Smartappli in #1416 +- feat: Add .close() method to Llama class to explicitly free model from memory by @jkawamoto in #1513 +- feat: Support SPM infill by @CISC in #1492 + +## [0.2.78] + +- feat: Update llama.cpp to ggerganov/llama.cpp@fd5ea0f897ecb3659d6c269ef6f3d833e865ead7 +- fix: Avoid duplicate special tokens in chat formats by @CISC in #1439 +- fix: fix logprobs when BOS is not present by @ghorbani in #1471 +- feat: adding rpc_servers parameter to Llama class by @chraac in #1477 + +## [0.2.77] + +- feat: Update llama.cpp to ggerganov/llama.cpp@bde7cd3cd949c1a85d3a199498ac98e78039d46f +- fix: string value kv_overrides by @abetlen in df45a4b3fe46e72664bda87301b318210c6d4782 +- fix: Fix typo in Llama3VisionAlphaChatHandler by @abetlen in 165b4dc6c188f8fda2fc616154e111f710484eba +- fix: Use numpy recarray for candidates data, fixes bug with temp < 0 by @abetlen in af3ed503e9ce60fe6b5365031abad4176a3536b3 +fix: Disable Windows+CUDA workaround when compiling for HIPBLAS by Engininja2 in #1493 + +## [0.2.76] + +- feat: Update llama.cpp to ggerganov/llama.cpp@0df0aa8e43c3378975269a51f9b876c8692e70da +- feat: Improve Llama.eval performance by avoiding list conversion by @thoughtp0lice in #1476 +- example: LLM inference with Ray Serve by @rgerganov in #1465 + +## [0.2.75] + +- feat: Update llama.cpp to ggerganov/llama.cpp@13ad16af1231ab2d245d35df3295bcfa23de1305 +- fix: segfault for models without eos / bos tokens by @abetlen in d99a6ba607a4885fb00e63e967964aa41bdbbbcb +- feat: add MinTokensLogitProcessor and min_tokens argument to server by @twaka in #1333 +- misc: Remove unnecessary metadata lookups by @CISC in #1448 + +## [0.2.74] + +- feat: Update llama.cpp to ggerganov/llama.cpp@b228aba91ac2cd9eb90e9d423ba1d0d20e0117e2 +- fix: Enable CUDA backend for llava by @abetlen in 7f59856fa6f3e23f07e12fc15aeb9359dc6c3bb4 +- docs: Fix typo in README.md by @yupbank in #1444 + +## [0.2.73] + +- feat: Update llama.cpp to ggerganov/llama.cpp@25c6e82e7a1ad25a42b0894e87d9b5c557409516 +- fix: Clear kv cache at beginning of image chat formats to avoid bug when image is evaluated first by @abetlen in ac55d0a175115d1e719672ce1cb1bec776c738b1 + +## [0.2.72] + +- fix(security): Remote Code Execution by Server-Side Template Injection in Model Metadata by @retr0reg in b454f40a9a1787b2b5659cd2cb00819d983185df +- fix(security): Update remaining jinja chat templates to use immutable sandbox by @CISC in #1441 + +## [0.2.71] + +- feat: Update llama.cpp to ggerganov/llama.cpp@911b3900dded9a1cfe0f0e41b82c7a29baf3a217 +- fix: Make leading bos_token optional for image chat formats, fix nanollava system message by @abetlen in 77122638b4153e31d9f277b3d905c2900b536632 +- fix: free last image embed in llava chat handler by @abetlen in 3757328b703b2cd32dcbd5853271e3a8c8599fe7 + +## [0.2.70] + +- feat: Update llama.cpp to ggerganov/llama.cpp@c0e6fbf8c380718102bd25fcb8d2e55f8f9480d1 +- feat: fill-in-middle support by @CISC in #1386 +- fix: adding missing args in create_completion for functionary chat handler by @skalade in #1430 +- docs: update README.md @eltociear in #1432 +- fix: chat_format log where auto-detected format prints None by @balvisio in #1434 +- feat(server): Add support for setting root_path by @abetlen in 0318702cdc860999ee70f277425edbbfe0e60419 +- feat(ci): Add docker checks and check deps more frequently by @Smartappli in #1426 +- fix: detokenization case where first token does not start with a leading space by @noamgat in #1375 +- feat: Implement streaming for Functionary v2 + Bug fixes by @jeffrey-fong in #1419 +- fix: Use memmove to copy str_value kv_override by @abetlen in 9f7a85571ae80d3b6ddbd3e1bae407b9f1e3448a +- feat(server): Remove temperature bounds checks for server by @abetlen in 0a454bebe67d12a446981eb16028c168ca5faa81 +- fix(server): Propagate flash_attn to model load by @dthuerck in #1424 + +## [0.2.69] + +- feat: Update llama.cpp to ggerganov/llama.cpp@6ecf3189e00a1e8e737a78b6d10e1d7006e050a2 +- feat: Add llama-3-vision-alpha chat format by @abetlen in 31b1d95a6c19f5b615a3286069f181a415f872e8 +- fix: Change default verbose value of verbose in image chat format handlers to True to match Llama by @abetlen in 4f01c452b6c738dc56eacac3758119b12c57ea94 +- fix: Suppress all logs when verbose=False, use hardcoded fileno's to work in colab notebooks by @abetlen in f116175a5a7c84569c88cad231855c1e6e59ff6e +- fix: UTF-8 handling with grammars by @jsoma in #1415 + +## [0.2.68] + +- feat: Update llama.cpp to ggerganov/llama.cpp@77e15bec6217a39be59b9cc83d6b9afb6b0d8167 +- feat: Add option to enable flash_attn to Lllama params and ModelSettings by @abetlen in 22d77eefd2edaf0148f53374d0cac74d0e25d06e +- fix(ci): Fix build-and-release.yaml by @Smartappli in #1413 + +## [0.2.67] + +- fix: Ensure image renders before text in chat formats regardless of message content order by @abetlen in 3489ef09d3775f4a87fb7114f619e8ba9cb6b656 +- fix(ci): Fix bug in use of upload-artifact failing to merge multiple artifacts into a single release by @abetlen in d03f15bb73a1d520970357b702a9e7d4cc2a7a62 + +## [0.2.66] + +- feat: Update llama.cpp to ggerganov/llama.cpp@8843a98c2ba97a25e93319a104f9ddfaf83ce4c4 +- feat: Generic Chat Formats, Tool Calling, and Huggingface Pull Support for Multimodal Models (Obsidian, LLaVA1.6, Moondream) by @abetlen in #1147 +- ci(fix): Workflow actions updates and fix arm64 wheels not included in release by @Smartappli in #1392 +- ci: Add support for pre-built cuda 12.4.1 wheels by @Smartappli in #1388 +- feat: Add support for str type kv_overrides by @abetlen in a411612b385cef100d76145da1fbd02a7b7cc894 +- fix: Functionary bug fixes by @jeffrey-fong in #1385 +- examples: fix quantize example by @iyubondyrev in #1387 +- ci: Update dependabot.yml by @Smartappli in #1391 + +## [0.2.65] + +- feat: Update llama.cpp to ggerganov/llama.cpp@46e12c4692a37bdd31a0432fc5153d7d22bc7f72 +- feat: Allow for possibly non-pooled embeddings by @iamlemec in #1380 + +## [0.2.64] + +- feat: Update llama.cpp to ggerganov/llama.cpp@4e96a812b3ce7322a29a3008db2ed73d9087b176 +- feat: Add `llama-3` chat format by @andreabak in #1371 +- feat: Use new llama_token_is_eog in create_completions by @abetlen in d40a250ef3cfaa8224d12c83776a2f1de96ae3d1 +- feat(server): Provide ability to dynamically allocate all threads if desired using -1 by @sean-bailey in #1364 +- ci: Build arm64 wheels by @gaby in 611781f5319719a3d05fefccbbf0cc321742a026 +- fix: Update scikit-build-core build dependency avoid bug in 0.9.1 by @evelkey in #1370 + +## [0.2.63] + +- feat: Update llama.cpp to ggerganov/llama.cpp@0e4802b2ecbaab04b4f829fde4a3096ca19c84b5 +- feat: Add stopping_criteria to ChatFormatter, allow stopping on arbitrary token ids, fixes llama3 instruct by @abetlen in cc81afebf04d26ca1ac3cf72f23f18da6ab58588 + +## [0.2.62] + +- feat: Update llama.cpp to ggerganov/llama.cpp@3b8f1ec4b18770531d0b1d792f3edf08254e4f0c +- feat: update grammar schema converter to match llama.cpp by @themrzmaster in #1353 +- feat: add disable_ping_events flag by @khimaros in #1257 +- feat: Make saved state more compact on-disk by @tc-wolf in #1296 +- feat: Use all available CPUs for batch processing by @ddh0 in #1345 + +## [0.2.61] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ba5e134e073ec6837078c874aba44a702944a676 +- fix: pass correct type to chat handlers for chat completion logprobs by @abetlen in bb65b4d76411112c6fb0bf759efd746f99ef3c6b +- feat: Add support for yaml based server configs by @abetlen in 060bfa64d529ade2af9b1f4e207a3937bbc4138f +- feat: Add typechecking for ctypes structure attributes by @abetlen in 1347e1d050fc5a9a32ffe0bb3e22858da28003bd + +## [0.2.60] + +- feat: Update llama.cpp to ggerganov/llama.cpp@75cd4c77292034ecec587ecb401366f57338f7c0 +- fix: Always embed metal library by @abetlen in b3bfea6dbfb6ed9ce18f9a2723e0a9e4bd1da7ad +- fix: missing logprobs in response, incorrect response type for functionary by @abetlen in 1ae3abbcc3af7f4a25a3ffc40b246f18039565e8 +- fix(docs): incorrect tool_choice example by @CISC in #1330 + +## [0.2.59] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ba0c7c70ab5b15f1f2be7fb0dfbe0366dda30d6c +- feat: Binary wheels for CPU, CUDA (12.1 - 12.3), Metal by @abetlen, @jllllll, and @oobabooga in #1247 +- fix: segfault when logits_all=False by @abetlen in 8649d7671bd1a7c0d9cc6a5ad91c6ca286512ab3 +- fix: last tokens passing to sample_repetition_penalties function by @ymikhailov in #1295 + +## [0.2.58] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ba0c7c70ab5b15f1f2be7fb0dfbe0366dda30d6c +- feat: add support for KV cache quantization options by @Limour-dev in #1307 +- feat: Add logprobs support to chat completions by @windspirit95 in #1311 +- fix: set LLAMA_METAL_EMBED_LIBRARY=on on MacOS arm64 by @bretello in #1289 +- feat: Add tools/functions variables to Jinja2ChatFormatter, add function response formatting for all simple chat formats by @CISC in #1273 +- fix: Changed local API doc references to hosted by by @lawfordp2017 in #1317 + +## [0.2.57] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ac9ee6a4ad740bc1ee484ede43e9f92b5af244c1 +- fix: set default embedding pooling type to unspecified by @abetlen in 4084aabe867b8ec2aba1b22659e59c9318b0d1f3 +- fix: Fix and optimize functionary chat handler by @jeffrey-fong in #1282 +- fix: json mode for basic chat formats by @abetlen in 20e6815252d0efd9f015f7adbf108faaf36e3f3c + +## [0.2.56] + +- feat: Update llama.cpp to ggerganov/llama.cpp@c2101a2e909ac7c08976d414e64e96c90ee5fa9e +- feat(server): Add endpoints for tokenize, detokenize and count tokens by @felipelo in #1136 +- feat: Switch embed to llama_get_embeddings_seq by @iamlemec in #1263 +- fix: Fixed json strings grammar by blacklisting character control set by @ExtReMLapin in d02a9cf16ff88ad011e2eb1ce29f4d9400f13cd1 +- fix: Check for existence of clip model path by @kejcao in #1264 + +## [0.2.55] + +- feat: Update llama.cpp to ggerganov/llama.cpp@9731134296af3a6839cd682e51d9c2109a871de5 +- docs: fix small typo in README: 'model know how' -> 'model knows how' by @boegel in #1244 + +## [0.2.54] + +- feat: Update llama.cpp to ggerganov/llama.cpp@cb49e0f8c906e5da49e9f6d64a57742a9a241c6a +- docs: fix typo in README.md embeddings example by @iamlemec in #1232 + +## [0.2.53] + +- feat: Update llama.cpp to ggerganov/llama.cpp@cb49e0f8c906e5da49e9f6d64a57742a9a241c6a +- fix: eos/bos_token set correctly for Jinja2ChatFormatter and automatic chat formatter by @CISC in #1230 + +## [0.2.52] + +- feat: Update llama.cpp to ggerganov/llama.cpp@a33e6a0d2a66104ea9a906bdbf8a94d050189d91 +- fix: Llava15ChatHandler (this function takes at least 4 arguments) by @abetlen in 8383a9e5620f5df5a88f62da16813eac200dd706 + +## [0.2.51] + +- feat: Update llama.cpp to ggerganov/llama.cpp@c39373398803c669056304090050fe3f44b41bf9 +- fix: Restore type hints for low-level api by @abetlen in 19234aa0dbd0c3c87656e65dd2b064665371925b + +## [0.2.50] + +- docs: Update Functionary OpenAI Server Readme by @jeffrey-fong in #1193 +- fix: LlamaHFTokenizer now receives pre_tokens by @abetlen in 47bad30dd716443652275099fa3851811168ff4a + +## [0.2.49] + +- fix: module 'llama_cpp.llama_cpp' has no attribute 'c_uint8' in Llama.save_state by @abetlen in db776a885cd4c20811f22f8bd1a27ecc71dba927 +- feat: Auto detect Mixtral's slightly different format by @lukestanley in #1214 + +## [0.2.48] + +- feat: Update llama.cpp to ggerganov/llama.cpp@15499eb94227401bdc8875da6eb85c15d37068f7 +- feat: Add Google's Gemma formatting via chat_format="gemma" by @alvarobartt in #1210 +- feat: support minItems/maxItems in JSON grammar converter by @nopperl in 3921e10770996d95a9eb22c8248bacef39f69365 +- fix: Update from_pretrained defaults to match hf_hub_download and pull to local cache folder by @abetlen in e6d6260a91b7831733f7d1f73c7af46a3e8185ed +- fix: Raise exceptions when llama model or context fails to load by @abetlen in dd22010e85265ae840c76ec835d67a29ed852722 +- docs: Update README.md to fix pip install llama cpp server by @audip in #1187 + +## [0.2.47] + +- feat: Update llama.cpp to ggerganov/llama.cpp@973053d8b0d04809836b3339a50f68d9c842de90 + +## [0.2.46] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ba2135ccae7462470b3865c6e41d2e1d734eac05 +- feat: Pull models directly from huggingface by @abetlen in #1206 +- feat(low-level-api): Improve API static type-safety and performance. Low level api functions are positional args only now. by @abetlen in #1205 + +## [0.2.45] + +- feat: Update llama.cpp to ggerganov/llama.cpp@89febfed9322c8849520dc63c93ee4f5fd72556e + +## [0.2.44] + +- feat: Update llama.cpp to ggerganov/llama.cpp@4524290e87b8e107cc2b56e1251751546f4b9051 +- fix: create_embedding broken response for input type str by @abetlen in 0ce66bc080fe537590b05b24bf442480bf2dd045 +- fix: Use '\n' seperator for EventSourceResponse by @khimaros in #1188 +- fix: Incorporate embedding pooling layer fixes by @iamlemec in #1194 + +## [0.2.43] + +- feat: Update llama.cpp to ggerganov/llama.cpp@8084d554406b767d36b3250b3b787462d5dd626f +- feat: Support batch embeddings by @iamlemec in #1186 +- fix: submodule kompute is not included in sdist by @abetlen in 7dbbfdecadebe7750be650d9409959640ff9a460 +- fix: fix: Update openbuddy prompt format by @abetlen in 07a783779a62a4aac0b11161c7e0eb983ff215f8 + +## [0.2.42] + +- feat: Update llama.cpp to ggerganov/llama.cpp@ea9c8e11436ad50719987fa23a289c74b7b40d40 +- fix: sample idx off-by-one error for logit_processors by @lapp0 in #1179 +- fix: chat formatting bugs in `chatml-function-calling` by @abetlen in 4b0e3320bd8c2c209e29978d0b21e2e471cc9ee3 and 68fb71b6a26a1e57331868f959b47ab4b87851e1 + +## [0.2.41] + +- feat: Update llama.cpp to ggerganov/llama.cpp@895407f31b358e3d9335e847d13f033491ec8a5b +- fix: Don't change order of json schema object properties in generated grammar unless prop_order is passed by @abetlen in d1822fed6b706f38bd1ff0de4dec5baaa3cf84fa + +## [0.2.40] + +- feat: Update llama.cpp to ggerganov/llama.cpp@3bdc4cd0f595a6096cca4a64aa75ffa8a3503465 +- feat: Generic chatml Function Calling using chat_format="chatml-function-calling"` by @abetlen in #957 +- fix: Circular dependancy preventing early Llama object free by @notwa in #1176 +- docs: Set the correct command for compiling with syscl support by @akarshanbiswas in #1172 +- feat: use gpu backend for clip if available by @iamlemec in #1175 + +## [0.2.39] + +- feat: Update llama.cpp to ggerganov/llama.cpp@b08f22c882a1443e6b97081f3ce718a4d1a741f8 +- fix: Fix destructor logging bugs by using llama_log_callback to avoid suppress_stdout_stderr by @abetlen in 59760c85eddc72dfcc1839f43760ef72c23d6874 + +## [0.2.38] + +- feat: Update llama.cpp to ggerganov/llama.cpp@1cfb5372cf5707c8ec6dde7c874f4a44a6c4c915 +- feat: Add speculative decoding by @abetlen in #1120 +- fix: Pass raise_exception and add_generation_prompt to jinja2 chat template by @abetlen in 078cca0361bf5a94d2cf52ed04980d20e32d6f95 + +## [0.2.37] + +- feat: Update llama.cpp to ggerganov/llama.cpp@fea4fd4ba7f6b754ac795387b275e1a014a77bde +- feat: Automatically set chat format from gguf by @abetlen in #1110 + +## [0.2.36] + +- feat: Update llama.cpp to ggerganov/llama.cpp@2aed77eb06a329f0d82bb1c467f4244904d4073f +- feat: Add mistral instruct chat format as "mistral-instruct" by @Rafaelblsilva in #799 + +## [0.2.35] + +- feat: Update llama.cpp to ggerganov/llama.cpp@d2f650cb5b04ee2726663e79b47da5efe196ce00 + +## [0.2.34] + +- feat: Update llama.cpp to ggerganov/llama.cpp@6db2b41a76ee78d5efdd5c3cddd5d7ad3f646855 +- feat: Add json schema mode by @abetlen in #1122 + +## [0.2.33] + +- feat: Update llama.cpp to ggerganov/llama.cpp@faa3526a1eba458120987ed8269e5616385a76f4 +- feat(server): include llama-cpp-python version in openapi spec by @abetlen in cde7514c3d28e6d52f272614e9957208c344dde5 +- fix: use both eos and bos tokens as stop sequences for hf-tokenizer-config chat format. by @abetlen in 5b982d0f8c6f35242c8862ffdce00e17cea0b44f +- fix: GGUF metadata KV overrides, re #1011 by @phiharri in #1116 +- fix: llama_log_set should be able to accept null pointer by @abetlen in c970d41a85381fd55235136f123422df0bf0c7e7 + +## [0.2.32] + +- feat: Update llama.cpp to ggerganov/llama.cpp@504dc37be8446fb09b1ede70300250ad41be32a2 +- fix: from_json_schema oneof/anyof bug by @jndiogo in d3f5528ca8bcb9d69d4f27e21631e911f1fb9bfe +- fix: pass chat handler not chat formatter for huggingface autotokenizer and tokenizer_config formats by @abetlen in 24f39454e91cf5dddbc4b6041aead4accc7c7a2d +- feat: Add add_generation_prompt option for jinja2chatformatter by @abetlen in 7f3209b1eb4ad3260ba063801fab80a8c25a2f4c +- feat: Add Jinja2ChatFormatter by @abetlen in be09318c26add8674ce494ae7cc480cce72a4146 +- feat: Expose gguf model metadata in metadata property by @abetlen in 5a34c57e5479e50c99aba9b38218cc48e6560b81 + +## [0.2.31] + +- feat: Update llama.cpp to ggerganov/llama.cpp@a5cacb22b2114fd9adf61c00cbb237384d86bced +- fix: Mirostat sampling now passes correct type to ctypes and tracks state during generation by @abetlen in 3babe3512cb95743108f2b595210c38ed6f1b904 +- fix: Python3.8 support in server by @abetlen in 141293a75b564a8699e0acba1da24d9aa1cf0ab1 + +## [0.2.30] + +- feat: Update llama.cpp to ggerganov/llama.cpp@57e2a7a52a819883f40dada8a2edc24ecf48186b +- feat(server): Add ability to load chat format from huggingface autotokenizer or tokenizer_config.json files by @abetlen in b8fc1c7d83ad4a9207c707ba1d954fe580286a01 +- feat: Integration of Jinja2 Templating for chat formats by @teleprint-me in #875 +- fix: Offload KQV by default by @abetlen in 48c3b77e6f558a9899de0e1155c7dc0c7958d8e8 +- fix: Support Accept text/event-stream in chat and completion endpoints, resolves #1083 by @aniljava in #1088 +- fix(cli): allow passing n_ctx=0 to openAI API server args to use model n_ctx_train field per #1015 by @K-Mistele in #1093 + +## [0.2.29] + +- feat: Update llama.cpp to ggerganov/llama.cpp@4483396751c79dea540808b9cb9238245d06da2b +- feat: Add split_mode option by @abetlen in 84615adbc6855c8384807c42f0130f9a1763f99d +- feat: Implement GGUF metadata KV overrides by @phiharri in #1011 +- fix: Avoid "LookupError: unknown encoding: ascii" when open() called in a destructor by @yieldthought in #1012 +- fix: Fix low_level_api_chat_cpp example to match current API by @aniljava in #1086 +- fix: Fix Pydantic model parsing by @DeNeutoy in #1087 + +## [0.2.28] + +- feat: Update llama.cpp to ggerganov/llama.cpp@6efb8eb30e7025b168f3fda3ff83b9b386428ad6 +- feat: Add ability to pass in penalize_nl param by @shankinson in #1068 +- fix: print_grammar to stderr by @turian in #1052 + +## [0.2.27] + +- feat: Update llama.cpp to ggerganov/llama.cpp@b3a7c20b5c035250257d2b62851c379b159c899a +- feat: Add `saiga` chat format by @femoiseev in #1050 +- feat: Added `chatglm3` chat format by @xaviviro in #1059 +- fix: Correct typo in README.md by @qeleb in (#1058) + +## [0.2.26] + +- feat: Update llama.cpp to ggerganov/llama.cpp@f6793491b5af6da75edad34d6f503ef86d31b09f + +## [0.2.25] + +- feat(server): Multi model support by @D4ve-R in #931 +- feat(server): Support none defaulting to infinity for completions by @swg in #111 +- feat(server): Implement openai api compatible authentication by @docmeth2 in #1010 +- fix: text_offset of multi-token characters by @twaka in #1037 +- fix: ctypes bindings for kv override by @phiharri in #1011 - fix: ctypes definitions of llama_kv_cache_view_update and llama_kv_cache_view_free. by @e-c-d in #1028 ## [0.2.24] @@ -127,7 +627,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [0.2.11] -- Fix bug in `llama_model_params` object has no attribute `logits_all` by @abetlen in d696251fbe40015e8616ea7a7d7ad5257fd1b896 +- Fix bug in `llama_model_params` object has no attribute `logits_all` by @abetlen in d696251fbe40015e8616ea7a7d7ad5257fd1b896 ## [0.2.10] @@ -315,7 +815,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [0.1.60] -NOTE: This release was deleted due to a bug with the packaging system that caused pip installations to fail. +NOTE: This release was deleted due to a bug with the packaging system that caused pip installations to fail. - Truncate max_tokens in create_completion so requested tokens doesn't exceed context size. - Temporarily disable cache for completion requests @@ -339,4 +839,4 @@ NOTE: This release was deleted due to a bug with the packaging system that caus - (misc) Added first version of the changelog - (server) Use async routes - (python-api) Use numpy for internal buffers to reduce memory usage and improve performance. -- (python-api) Performance bug in stop sequence check slowing down streaming. \ No newline at end of file +- (python-api) Performance bug in stop sequence check slowing down streaming. diff --git a/CMakeLists.txt b/CMakeLists.txt index 795dad726..b9178e856 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -5,70 +5,176 @@ project(llama_cpp) option(LLAMA_BUILD "Build llama.cpp shared library and install alongside python package" ON) option(LLAVA_BUILD "Build llava shared library and install alongside python package" ON) -if (LLAMA_BUILD) - set(BUILD_SHARED_LIBS "On") - - # Building llama - if (APPLE AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm64") - # Need to disable these llama.cpp flags on Apple x86_64, - # otherwise users may encounter invalid instruction errors - set(LLAMA_AVX "Off" CACHE BOOL "llama: enable AVX" FORCE) - set(LLAMA_AVX2 "Off" CACHE BOOL "llama: enable AVX2" FORCE) - set(LLAMA_FMA "Off" CACHE BOOL "llama: enable FMA" FORCE) - set(LLAMA_F16C "Off" CACHE BOOL "llama: enable F16C" FORCE) +function(llama_cpp_python_install_target target) + if(NOT TARGET ${target}) + return() endif() - add_subdirectory(vendor/llama.cpp) + install( - TARGETS llama - LIBRARY DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - RUNTIME DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - ARCHIVE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - FRAMEWORK DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - RESOURCE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp + TARGETS ${target} + LIBRARY DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + RUNTIME DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + ARCHIVE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + FRAMEWORK DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + RESOURCE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib ) - # Temporary fix for https://github.com/scikit-build/scikit-build-core/issues/374 install( - TARGETS llama - LIBRARY DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - RUNTIME DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - ARCHIVE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - FRAMEWORK DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - RESOURCE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp + TARGETS ${target} + LIBRARY DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + RUNTIME DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + ARCHIVE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + FRAMEWORK DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + RESOURCE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib ) - # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563 - install( - FILES $ - DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - ) - install( - FILES $ - DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp + set_target_properties(${target} PROPERTIES + INSTALL_RPATH "$ORIGIN" + BUILD_WITH_INSTALL_RPATH TRUE ) + if(UNIX) + if(APPLE) + set_target_properties(${target} PROPERTIES + INSTALL_RPATH "@loader_path" + BUILD_WITH_INSTALL_RPATH TRUE + ) + else() + set_target_properties(${target} PROPERTIES + INSTALL_RPATH "$ORIGIN" + BUILD_WITH_INSTALL_RPATH TRUE + ) + endif() + endif() +endfunction() + +if (LLAMA_BUILD) + set(BUILD_SHARED_LIBS "On") + + set(CMAKE_SKIP_BUILD_RPATH FALSE) + + # When building, don't use the install RPATH already + # (but later on when installing) + set(CMAKE_BUILD_WITH_INSTALL_RPATH FALSE) + + # Add the automatically determined parts of the RPATH + # which point to directories outside the build tree to the install RPATH + set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE) + set(CMAKE_SKIP_RPATH FALSE) + + # Enable building of the common library + set(LLAMA_BUILD_COMMON ON CACHE BOOL "Build llama.cpp common library" FORCE) + + # Disable building curl support + set(LLAMA_CURL OFF CACHE BOOL "llama.cpp: enable curl" FORCE) + + # Architecture detection and settings for Apple platforms + if (APPLE) + # Get the target architecture + execute_process( + COMMAND uname -m + OUTPUT_VARIABLE HOST_ARCH + OUTPUT_STRIP_TRAILING_WHITESPACE + ) + + # If CMAKE_OSX_ARCHITECTURES is not set, use the host architecture + if(NOT CMAKE_OSX_ARCHITECTURES) + set(CMAKE_OSX_ARCHITECTURES ${HOST_ARCH} CACHE STRING "Build architecture for macOS" FORCE) + endif() + + message(STATUS "Host architecture: ${HOST_ARCH}") + message(STATUS "Target architecture: ${CMAKE_OSX_ARCHITECTURES}") + + # Configure based on target architecture + if(CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64") + # Intel Mac settings + set(GGML_AVX "OFF" CACHE BOOL "ggml: enable AVX" FORCE) + set(GGML_AVX2 "OFF" CACHE BOOL "ggml: enable AVX2" FORCE) + set(GGML_FMA "OFF" CACHE BOOL "ggml: enable FMA" FORCE) + set(GGML_F16C "OFF" CACHE BOOL "ggml: enable F16C" FORCE) + endif() + + # Metal settings (enable for both architectures) + set(GGML_METAL "ON" CACHE BOOL "ggml: enable Metal" FORCE) + set(GGML_METAL_EMBED_LIBRARY "ON" CACHE BOOL "ggml: embed metal library" FORCE) + endif() + + add_subdirectory(vendor/llama.cpp) + llama_cpp_python_install_target(llama) + llama_cpp_python_install_target(ggml) + + llama_cpp_python_install_target(ggml-base) + + llama_cpp_python_install_target(ggml-amx) + llama_cpp_python_install_target(ggml-blas) + llama_cpp_python_install_target(ggml-can) + llama_cpp_python_install_target(ggml-cpu) + llama_cpp_python_install_target(ggml-cuda) + llama_cpp_python_install_target(ggml-hip) + llama_cpp_python_install_target(ggml-kompute) + llama_cpp_python_install_target(ggml-metal) + llama_cpp_python_install_target(ggml-musa) + llama_cpp_python_install_target(ggml-rpc) + llama_cpp_python_install_target(ggml-sycl) + llama_cpp_python_install_target(ggml-vulkan) + + # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563 + if (WIN32) + install( + FILES $ + DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + ) + install( + FILES $ + DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + ) + install( + FILES $ + DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + ) + install( + FILES $ + DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + ) + endif() if (LLAVA_BUILD) + if (LLAMA_CUBLAS OR LLAMA_CUDA) + add_compile_definitions(GGML_USE_CUBLAS) + add_compile_definitions(GGML_USE_CUDA) + endif() + + if (LLAMA_METAL) + add_compile_definitions(GGML_USE_METAL) + endif() + # Building llava - add_subdirectory(vendor/llama.cpp/examples/llava) + add_subdirectory(vendor/llama.cpp/tools/mtmd) set_target_properties(llava_shared PROPERTIES OUTPUT_NAME "llava") - # Set CUDA_ARCHITECTURES to OFF on windows + if (WIN32) set_target_properties(llava_shared PROPERTIES CUDA_ARCHITECTURES OFF) endif() - install( - TARGETS llava_shared - LIBRARY DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - RUNTIME DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - ARCHIVE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - FRAMEWORK DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - RESOURCE DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp - ) - # Temporary fix for https://github.com/scikit-build/scikit-build-core/issues/374 - install( - TARGETS llava_shared - LIBRARY DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - RUNTIME DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - ARCHIVE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - FRAMEWORK DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - RESOURCE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp - ) + llama_cpp_python_install_target(llava_shared) + if (WIN32) + install( + FILES $ + DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + ) + install( + FILES $ + DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + ) + endif() + + # Fix for llava build: Add include directory for llama.h + # Move these commands after the add_subdirectory call + target_include_directories(llava PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + target_include_directories(llava PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/ggml/include) + + if (BUILD_SHARED_LIBS) + target_include_directories(llava_shared PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + target_include_directories(llava_shared PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/ggml/include) + endif() + + target_include_directories(llama-llava-cli PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + target_include_directories(llama-minicpmv-cli PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) endif() endif() diff --git a/Makefile b/Makefile index e930609ff..26ddf2c7a 100644 --- a/Makefile +++ b/Makefile @@ -10,25 +10,52 @@ deps: python3 -m pip install -e ".[all]" build: - python3 -m pip install -e . + python3 -m pip install --verbose -e . + +build.debug: + python3 -m pip install \ + --verbose \ + --config-settings=cmake.verbose=true \ + --config-settings=logging.level=INFO \ + --config-settings=install.strip=false \ + --config-settings=cmake.args="-DCMAKE_BUILD_TYPE=Debug;-DCMAKE_C_FLAGS='-ggdb -O0';-DCMAKE_CXX_FLAGS='-ggdb -O0'" \ + --editable . + +build.debug.extra: + python3 -m pip install \ + --verbose \ + --config-settings=cmake.verbose=true \ + --config-settings=logging.level=INFO \ + --config-settings=install.strip=false \ + --config-settings=cmake.args="-DCMAKE_BUILD_TYPE=Debug;-DCMAKE_C_FLAGS='-fsanitize=address -ggdb -O0';-DCMAKE_CXX_FLAGS='-fsanitize=address -ggdb -O0'" \ + --editable . build.cuda: - CMAKE_ARGS="-DLLAMA_CUBLAS=on" python3 -m pip install -e . - -build.opencl: - CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install -e . + CMAKE_ARGS="-DGGML_CUDA=on" python3 -m pip install --verbose -e . build.openblas: - CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install -e . + CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" python3 -m pip install --verbose -e . build.blis: - CMAKE_ARGS="-DLLAMA_OPENBLAS=on -DLLAMA_OPENBLAS_VENDOR=blis" python3 -m pip install -e . + CMAKE_ARGS="-DGGML_BLAS=on -DGGML_BLAS_VENDOR=FLAME" python3 -m pip install --verbose -e . build.metal: - CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install -e . + CMAKE_ARGS="-DGGML_METAL=on" python3 -m pip install --verbose -e . + +build.vulkan: + CMAKE_ARGS="-DGGML_VULKAN=on" python3 -m pip install --verbose -e . + +build.kompute: + CMAKE_ARGS="-DGGML_KOMPUTE=on" python3 -m pip install --verbose -e . + +build.sycl: + CMAKE_ARGS="-DGGML_SYCL=on" python3 -m pip install --verbose -e . + +build.rpc: + CMAKE_ARGS="-DGGML_RPC=on" python3 -m pip install --verbose -e . build.sdist: - python3 -m build --sdist + python3 -m build --sdist --verbose deploy.pypi: python3 -m twine upload dist/* @@ -38,23 +65,23 @@ deploy.gh-docs: mkdocs gh-deploy test: - python3 -m pytest + python3 -m pytest --full-trace -v docker: docker build -t llama-cpp-python:latest -f docker/simple/Dockerfile . run-server: - uvicorn --factory llama.server:app --host ${HOST} --port ${PORT} + python3 -m llama_cpp.server --model ${MODEL} clean: - cd vendor/llama.cpp && make clean - cd vendor/llama.cpp && rm libllama.so - rm -rf _skbuild - - rm llama_cpp/*.so - - rm llama_cpp/*.dylib - - rm llama_cpp/*.metal - - rm llama_cpp/*.dll - - rm llama_cpp/*.lib + - rm llama_cpp/lib/*.so + - rm llama_cpp/lib/*.dylib + - rm llama_cpp/lib/*.metal + - rm llama_cpp/lib/*.dll + - rm llama_cpp/lib/*.lib .PHONY: \ update \ @@ -67,4 +94,4 @@ clean: deploy.pypi \ deploy.gh-docs \ docker \ - clean \ No newline at end of file + clean diff --git a/README.md b/README.md index 2f413db16..e00456580 100644 --- a/README.md +++ b/README.md @@ -1,4 +1,8 @@ -# 🦙 Python Bindings for [`llama.cpp`](https://github.com/ggerganov/llama.cpp) +

+ +

+ +# Python Bindings for [`llama.cpp`](https://github.com/ggerganov/llama.cpp) [![Documentation Status](https://readthedocs.org/projects/llama-cpp-python/badge/?version=latest)](https://llama-cpp-python.readthedocs.io/en/latest/?badge=latest) [![Tests](https://github.com/abetlen/llama-cpp-python/actions/workflows/test.yaml/badge.svg?branch=main)](https://github.com/abetlen/llama-cpp-python/actions/workflows/test.yaml) @@ -6,6 +10,7 @@ [![PyPI - Python Version](https://img.shields.io/pypi/pyversions/llama-cpp-python)](https://pypi.org/project/llama-cpp-python/) [![PyPI - License](https://img.shields.io/pypi/l/llama-cpp-python)](https://pypi.org/project/llama-cpp-python/) [![PyPI - Downloads](https://img.shields.io/pypi/dm/llama-cpp-python)](https://pypi.org/project/llama-cpp-python/) +[![Github All Releases](https://img.shields.io/github/downloads/abetlen/llama-cpp-python/total.svg?label=Github%20Downloads)]() Simple Python bindings for **@ggerganov's** [`llama.cpp`](https://github.com/ggerganov/llama.cpp) library. This package provides: @@ -14,120 +19,249 @@ This package provides: - High-level Python API for text completion - OpenAI-like API - [LangChain compatibility](https://python.langchain.com/docs/integrations/llms/llamacpp) + - [LlamaIndex compatibility](https://docs.llamaindex.ai/en/stable/examples/llm/llama_2_llama_cpp.html) - OpenAI compatible web server - [Local Copilot replacement](https://llama-cpp-python.readthedocs.io/en/latest/server/#code-completion) - [Function Calling support](https://llama-cpp-python.readthedocs.io/en/latest/server/#function-calling) - [Vision API support](https://llama-cpp-python.readthedocs.io/en/latest/server/#multimodal-models) + - [Multiple Models](https://llama-cpp-python.readthedocs.io/en/latest/server/#configuration-and-multi-model-support) Documentation is available at [https://llama-cpp-python.readthedocs.io/en/latest](https://llama-cpp-python.readthedocs.io/en/latest). +## Installation +Requirements: -## Installation + - Python 3.8+ + - C compiler + - Linux: gcc or clang + - Windows: Visual Studio or MinGW + - MacOS: Xcode -`llama-cpp-python` can be installed directly from PyPI as a source distribution by running: +To install the package, run: ```bash pip install llama-cpp-python ``` -This will build `llama.cpp` from source using cmake and your system's c compiler (required) and install the library alongside this python package. +This will also build `llama.cpp` from source and install it alongside this python package. -If you run into issues during installation add the `--verbose` flag to the `pip install` command to see the full cmake build log. +If this fails, add `--verbose` to the `pip install` see the full cmake build log. +**Pre-built Wheel (New)** -### Installation with Specific Hardware Acceleration (BLAS, CUDA, Metal, etc) +It is also possible to install a pre-built wheel with basic CPU support. + +```bash +pip install llama-cpp-python \ + --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu +``` -The default pip install behaviour is to build `llama.cpp` for CPU only on Linux and Windows and use Metal on MacOS. +### Installation Configuration -`llama.cpp` supports a number of hardware acceleration backends depending including OpenBLAS, cuBLAS, CLBlast, HIPBLAS, and Metal. -See the [llama.cpp README](https://github.com/ggerganov/llama.cpp#build) for a full list of supported backends. +`llama.cpp` supports a number of hardware acceleration backends to speed up inference as well as backend specific options. See the [llama.cpp README](https://github.com/ggerganov/llama.cpp#build) for a full list. -All of these backends are supported by `llama-cpp-python` and can be enabled by setting the `CMAKE_ARGS` environment variable before installing. +All `llama.cpp` cmake build options can be set via the `CMAKE_ARGS` environment variable or via the `--config-settings / -C` cli flag during installation. -On Linux and Mac you set the `CMAKE_ARGS` like this: +
+Environment Variables ```bash -CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python +# Linux and Mac +CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" \ + pip install llama-cpp-python ``` -On Windows you can set the `CMAKE_ARGS` like this: - -```ps -$env:CMAKE_ARGS = "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" +```powershell +# Windows +$env:CMAKE_ARGS = "-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python ``` +
+ +
+CLI / requirements.txt + +They can also be set via `pip install -C / --config-settings` command and saved to a `requirements.txt` file: + +```bash +pip install --upgrade pip # ensure pip is up to date +pip install llama-cpp-python \ + -C cmake.args="-DGGML_BLAS=ON;-DGGML_BLAS_VENDOR=OpenBLAS" +``` + +```txt +# requirements.txt + +llama-cpp-python -C cmake.args="-DGGML_BLAS=ON;-DGGML_BLAS_VENDOR=OpenBLAS" +``` + +
+ +### Supported Backends + +Below are some common backends, their build commands and any additional environment variables required. + +
+OpenBLAS (CPU) + +To install with OpenBLAS, set the `GGML_BLAS` and `GGML_BLAS_VENDOR` environment variables before installing: + +```bash +CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python +``` +
+ +
+CUDA + +To install with CUDA support, set the `GGML_CUDA=on` environment variable before installing: + +```bash +CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python +``` -#### OpenBLAS +**Pre-built Wheel (New)** -To install with OpenBLAS, set the `LLAMA_BLAS and LLAMA_BLAS_VENDOR` environment variables before installing: +It is also possible to install a pre-built wheel with CUDA support. As long as your system meets some requirements: + +- CUDA Version is 12.1, 12.2, 12.3, 12.4 or 12.5 +- Python Version is 3.10, 3.11 or 3.12 ```bash -CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python +pip install llama-cpp-python \ + --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/ ``` -#### cuBLAS +Where `` is one of the following: +- `cu121`: CUDA 12.1 +- `cu122`: CUDA 12.2 +- `cu123`: CUDA 12.3 +- `cu124`: CUDA 12.4 +- `cu125`: CUDA 12.5 -To install with cuBLAS, set the `LLAMA_CUBLAS=1` environment variable before installing: +For example, to install the CUDA 12.1 wheel: ```bash -CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python +pip install llama-cpp-python \ + --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121 ``` -#### Metal +
+ +
+Metal -To install with Metal (MPS), set the `LLAMA_METAL=on` environment variable before installing: +To install with Metal (MPS), set the `GGML_METAL=on` environment variable before installing: ```bash -CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_METAL=on" pip install llama-cpp-python ``` -#### CLBlast +**Pre-built Wheel (New)** + +It is also possible to install a pre-built wheel with Metal support. As long as your system meets some requirements: -To install with CLBlast, set the `LLAMA_CLBLAST=1` environment variable before installing: +- MacOS Version is 11.0 or later +- Python Version is 3.10, 3.11 or 3.12 ```bash -CMAKE_ARGS="-DLLAMA_CLBLAST=on" pip install llama-cpp-python +pip install llama-cpp-python \ + --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/metal ``` -#### hipBLAS +
-To install with hipBLAS / ROCm support for AMD cards, set the `LLAMA_HIPBLAS=on` environment variable before installing: +
+hipBLAS (ROCm) + +To install with hipBLAS / ROCm support for AMD cards, set the `GGML_HIPBLAS=on` environment variable before installing: ```bash -CMAKE_ARGS="-DLLAMA_HIPBLAS=on" pip install llama-cpp-python +CMAKE_ARGS="-DGGML_HIPBLAS=on" pip install llama-cpp-python ``` +
+ +
+Vulkan + +To install with Vulkan support, set the `GGML_VULKAN=on` environment variable before installing: + +```bash +CMAKE_ARGS="-DGGML_VULKAN=on" pip install llama-cpp-python +``` + +
+ +
+SYCL + +To install with SYCL support, set the `GGML_SYCL=on` environment variable before installing: + +```bash +source /opt/intel/oneapi/setvars.sh +CMAKE_ARGS="-DGGML_SYCL=on -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx" pip install llama-cpp-python +``` +
+ +
+RPC + +To install with RPC support, set the `GGML_RPC=on` environment variable before installing: + +```bash +source /opt/intel/oneapi/setvars.sh +CMAKE_ARGS="-DGGML_RPC=on" pip install llama-cpp-python +``` +
+ + ### Windows Notes +
+Error: Can't find 'nmake' or 'CMAKE_C_COMPILER' + If you run into issues where it complains it can't find `'nmake'` `'?'` or CMAKE_C_COMPILER, you can extract w64devkit as [mentioned in llama.cpp repo](https://github.com/ggerganov/llama.cpp#openblas) and add those manually to CMAKE_ARGS before running `pip` install: + ```ps $env:CMAKE_GENERATOR = "MinGW Makefiles" -$env:CMAKE_ARGS = "-DLLAMA_OPENBLAS=on -DCMAKE_C_COMPILER=C:/w64devkit/bin/gcc.exe -DCMAKE_CXX_COMPILER=C:/w64devkit/bin/g++.exe" +$env:CMAKE_ARGS = "-DGGML_OPENBLAS=on -DCMAKE_C_COMPILER=C:/w64devkit/bin/gcc.exe -DCMAKE_CXX_COMPILER=C:/w64devkit/bin/g++.exe" ``` See the above instructions and set `CMAKE_ARGS` to the BLAS backend you want to use. +
### MacOS Notes +Detailed MacOS Metal GPU install documentation is available at [docs/install/macos.md](https://llama-cpp-python.readthedocs.io/en/latest/install/macos/) + +
+M1 Mac Performance Issue + Note: If you are using Apple Silicon (M1) Mac, make sure you have installed a version of Python that supports arm64 architecture. For example: -``` + +```bash wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-MacOSX-arm64.sh bash Miniforge3-MacOSX-arm64.sh ``` -Otherwise, while installing it will build the llama.cpp x86 version which will be 10x slower on Apple Silicon (M1) Mac. -Detailed MacOS Metal GPU install documentation is available at [docs/install/macos.md](https://llama-cpp-python.readthedocs.io/en/latest/install/macos/) +Otherwise, while installing it will build the llama.cpp x86 version which will be 10x slower on Apple Silicon (M1) Mac. +
-### Upgrading and Reinstalling +
+M Series Mac Error: `(mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64'))` -To upgrade or rebuild `llama-cpp-python` add the following flags to ensure that the package is rebuilt correctly: +Try installing with ```bash -pip install llama-cpp-python --upgrade --force-reinstall --no-cache-dir +CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DGGML_METAL=on" pip install --upgrade --verbose --force-reinstall --no-cache-dir llama-cpp-python ``` +
-This will ensure that all source files are re-built with the most recently set `CMAKE_ARGS` flags. +### Upgrading and Reinstalling + +To upgrade and rebuild `llama-cpp-python` add `--upgrade --force-reinstall --no-cache-dir` flags to the `pip install` command to ensure the package is rebuilt from source. ## High-level API @@ -138,15 +272,26 @@ The high-level API provides a simple managed interface through the [`Llama`](htt Below is a short example demonstrating how to use the high-level API to for basic text completion: ```python ->>> from llama_cpp import Llama ->>> llm = Llama(model_path="./models/7B/llama-model.gguf") ->>> output = llm( +from llama_cpp import Llama + +llm = Llama( + model_path="./models/7B/llama-model.gguf", + # n_gpu_layers=-1, # Uncomment to use GPU acceleration + # seed=1337, # Uncomment to set a specific seed + # n_ctx=2048, # Uncomment to increase the context window +) +output = llm( "Q: Name the planets in the solar system? A: ", # Prompt - max_tokens=32, # Generate up to 32 tokens + max_tokens=32, # Generate up to 32 tokens, set to None to generate up to the end of the context window stop=["Q:", "\n"], # Stop generating just before the model would generate a new question echo=True # Echo the prompt back in the output ) # Generate a completion, can also call create_completion ->>> print(output) +print(output) +``` + +By default `llama-cpp-python` generates completions in an OpenAI compatible format: + +```python { "id": "cmpl-xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx", "object": "text_completion", @@ -170,16 +315,43 @@ Below is a short example demonstrating how to use the high-level API to for basi Text completion is available through the [`__call__`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.__call__) and [`create_completion`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_completion) methods of the [`Llama`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama) class. +### Pulling models from Hugging Face Hub + +You can download `Llama` models in `gguf` format directly from Hugging Face using the [`from_pretrained`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.from_pretrained) method. +You'll need to install the `huggingface-hub` package to use this feature (`pip install huggingface-hub`). + +```python +llm = Llama.from_pretrained( + repo_id="Qwen/Qwen2-0.5B-Instruct-GGUF", + filename="*q8_0.gguf", + verbose=False +) +``` + +By default [`from_pretrained`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.from_pretrained) will download the model to the huggingface cache directory, you can then manage installed model files with the [`huggingface-cli`](https://huggingface.co/docs/huggingface_hub/en/guides/cli) tool. + ### Chat Completion The high-level API also provides a simple interface for chat completion. -Note that `chat_format` option must be set for the particular model you are using. +Chat completion requires that the model knows how to format the messages into a single prompt. +The `Llama` class does this using pre-registered chat formats (ie. `chatml`, `llama-2`, `gemma`, etc) or by providing a custom chat handler object. + +The model will will format the messages into a single prompt using the following order of precedence: + - Use the `chat_handler` if provided + - Use the `chat_format` if provided + - Use the `tokenizer.chat_template` from the `gguf` model's metadata (should work for most new models, older models may not have this) + - else, fallback to the `llama-2` chat format + +Set `verbose=True` to see the selected chat format. ```python ->>> from llama_cpp import Llama ->>> llm = Llama(model_path="path/to/llama-2/llama-model.gguf", chat_format="llama-2") ->>> llm.create_chat_completion( +from llama_cpp import Llama +llm = Llama( + model_path="path/to/llama-2/llama-model.gguf", + chat_format="llama-2" +) +llm.create_chat_completion( messages = [ {"role": "system", "content": "You are an assistant who perfectly describes images."}, { @@ -192,23 +364,75 @@ Note that `chat_format` option must be set for the particular model you are usin Chat completion is available through the [`create_chat_completion`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_chat_completion) method of the [`Llama`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama) class. -### Function Calling +For OpenAI API v1 compatibility, you use the [`create_chat_completion_openai_v1`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_chat_completion_openai_v1) method which will return pydantic models instead of dicts. + + +### JSON and JSON Schema Mode + +To constrain chat responses to only valid JSON or a specific JSON Schema use the `response_format` argument in [`create_chat_completion`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_chat_completion). + +#### JSON Mode + +The following example will constrain the response to valid JSON strings only. + +```python +from llama_cpp import Llama +llm = Llama(model_path="path/to/model.gguf", chat_format="chatml") +llm.create_chat_completion( + messages=[ + { + "role": "system", + "content": "You are a helpful assistant that outputs in JSON.", + }, + {"role": "user", "content": "Who won the world series in 2020"}, + ], + response_format={ + "type": "json_object", + }, + temperature=0.7, +) +``` + +#### JSON Schema Mode -The high-level API also provides a simple interface for function calling. +To constrain the response further to a specific JSON Schema add the schema to the `schema` property of the `response_format` argument. + +```python +from llama_cpp import Llama +llm = Llama(model_path="path/to/model.gguf", chat_format="chatml") +llm.create_chat_completion( + messages=[ + { + "role": "system", + "content": "You are a helpful assistant that outputs in JSON.", + }, + {"role": "user", "content": "Who won the world series in 2020"}, + ], + response_format={ + "type": "json_object", + "schema": { + "type": "object", + "properties": {"team_name": {"type": "string"}}, + "required": ["team_name"], + }, + }, + temperature=0.7, +) +``` -Note that the only model that supports full function calling at this time is "functionary". -The gguf-converted files for this model can be found here: [functionary-7b-v1](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) +### Function Calling +The high-level API supports OpenAI compatible function and tool calling. This is possible through the `functionary` pre-trained models chat format or through the generic `chatml-function-calling` chat format. ```python ->>> from llama_cpp import Llama ->>> llm = Llama(model_path="path/to/functionary/llama-model.gguf", chat_format="functionary") ->>> llm.create_chat_completion( +from llama_cpp import Llama +llm = Llama(model_path="path/to/chatml/llama-model.gguf", chat_format="chatml-function-calling") +llm.create_chat_completion( messages = [ { "role": "system", "content": "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant calls functions with appropriate input when necessary" - + }, { "role": "user", @@ -236,53 +460,184 @@ The gguf-converted files for this model can be found here: [functionary-7b-v1](h } } }], - tool_choices=[{ + tool_choice={ "type": "function", "function": { "name": "UserDetail" } - }] + } ) ``` -### Multi-modal Models +
+Functionary v2 +The various gguf-converted files for this set of models can be found [here](https://huggingface.co/meetkai). Functionary is able to intelligently call functions and also analyze any provided function outputs to generate coherent responses. All v2 models of functionary supports **parallel function calling**. You can provide either `functionary-v1` or `functionary-v2` for the `chat_format` when initializing the Llama class. -`llama-cpp-python` supports the llava1.5 family of multi-modal models which allow the language model to -read information from both text and images. +Due to discrepancies between llama.cpp and HuggingFace's tokenizers, it is required to provide HF Tokenizer for functionary. The `LlamaHFTokenizer` class can be initialized and passed into the Llama class. This will override the default llama.cpp tokenizer used in Llama class. The tokenizer files are already included in the respective HF repositories hosting the gguf files. -You'll first need to download one of the available multi-modal models in GGUF format: +```python +from llama_cpp import Llama +from llama_cpp.llama_tokenizer import LlamaHFTokenizer +llm = Llama.from_pretrained( + repo_id="meetkai/functionary-small-v2.2-GGUF", + filename="functionary-small-v2.2.q4_0.gguf", + chat_format="functionary-v2", + tokenizer=LlamaHFTokenizer.from_pretrained("meetkai/functionary-small-v2.2-GGUF") +) +``` -- [llava-v1.5-7b](https://huggingface.co/mys/ggml_llava-v1.5-7b) -- [llava-v1.5-13b](https://huggingface.co/mys/ggml_llava-v1.5-13b) -- [bakllava-1-7b](https://huggingface.co/mys/ggml_bakllava-1) +**NOTE**: There is no need to provide the default system messages used in Functionary as they are added automatically in the Functionary chat handler. Thus, the messages should contain just the chat messages and/or system messages that provide additional context for the model (e.g.: datetime, etc.). +
+ +### Multi-modal Models + +`llama-cpp-python` supports such as llava1.5 which allow the language model to read information from both text and images. + +Below are the supported multi-modal models and their respective chat handlers (Python API) and chat formats (Server API). + +| Model | `LlamaChatHandler` | `chat_format` | +|:--- |:--- |:--- | +| [llava-v1.5-7b](https://huggingface.co/mys/ggml_llava-v1.5-7b) | `Llava15ChatHandler` | `llava-1-5` | +| [llava-v1.5-13b](https://huggingface.co/mys/ggml_llava-v1.5-13b) | `Llava15ChatHandler` | `llava-1-5` | +| [llava-v1.6-34b](https://huggingface.co/cjpais/llava-v1.6-34B-gguf) | `Llava16ChatHandler` | `llava-1-6` | +| [moondream2](https://huggingface.co/vikhyatk/moondream2) | `MoondreamChatHandler` | `moondream2` | +| [nanollava](https://huggingface.co/abetlen/nanollava-gguf) | `NanollavaChatHandler` | `nanollava` | +| [llama-3-vision-alpha](https://huggingface.co/abetlen/llama-3-vision-alpha-gguf) | `Llama3VisionAlphaChatHandler` | `llama-3-vision-alpha` | +| [minicpm-v-2.6](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) | `MiniCPMv26ChatHandler` | `minicpm-v-2.6` | Then you'll need to use a custom chat handler to load the clip model and process the chat messages and images. ```python ->>> from llama_cpp import Llama ->>> from llama_cpp.llama_chat_format import Llava15ChatHandler ->>> chat_handler = Llava15ChatHandler(clip_model_path="path/to/llava/mmproj.bin") ->>> llm = Llama( +from llama_cpp import Llama +from llama_cpp.llama_chat_format import Llava15ChatHandler +chat_handler = Llava15ChatHandler(clip_model_path="path/to/llava/mmproj.bin") +llm = Llama( model_path="./path/to/llava/llama-model.gguf", chat_handler=chat_handler, - n_ctx=2048, # n_ctx should be increased to accomodate the image embedding - logits_all=True,# needed to make llava work + n_ctx=2048, # n_ctx should be increased to accommodate the image embedding ) ->>> llm.create_chat_completion( +llm.create_chat_completion( messages = [ {"role": "system", "content": "You are an assistant who perfectly describes images."}, { "role": "user", "content": [ - {"type": "image_url", "image_url": {"url": "https://.../image.png"}}, - {"type" : "text", "text": "Describe this image in detail please."} + {"type" : "text", "text": "What's in this image?"}, + {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" } } + ] + } + ] +) +``` + +You can also pull the model from the Hugging Face Hub using the `from_pretrained` method. + +```python +from llama_cpp import Llama +from llama_cpp.llama_chat_format import MoondreamChatHandler + +chat_handler = MoondreamChatHandler.from_pretrained( + repo_id="vikhyatk/moondream2", + filename="*mmproj*", +) + +llm = Llama.from_pretrained( + repo_id="vikhyatk/moondream2", + filename="*text-model*", + chat_handler=chat_handler, + n_ctx=2048, # n_ctx should be increased to accommodate the image embedding +) + +response = llm.create_chat_completion( + messages = [ + { + "role": "user", + "content": [ + {"type" : "text", "text": "What's in this image?"}, + {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" } } + ] } ] ) +print(response["choices"][0]["text"]) ``` +**Note**: Multi-modal models also support tool calling and JSON mode. + +
+Loading a Local Image + +Images can be passed as base64 encoded data URIs. The following example demonstrates how to do this. + +```python +import base64 + +def image_to_base64_data_uri(file_path): + with open(file_path, "rb") as img_file: + base64_data = base64.b64encode(img_file.read()).decode('utf-8') + return f"data:image/png;base64,{base64_data}" + +# Replace 'file_path.png' with the actual path to your PNG file +file_path = 'file_path.png' +data_uri = image_to_base64_data_uri(file_path) + +messages = [ + {"role": "system", "content": "You are an assistant who perfectly describes images."}, + { + "role": "user", + "content": [ + {"type": "image_url", "image_url": {"url": data_uri }}, + {"type" : "text", "text": "Describe this image in detail please."} + ] + } +] + +``` + +
+ +### Speculative Decoding + +`llama-cpp-python` supports speculative decoding which allows the model to generate completions based on a draft model. + +The fastest way to use speculative decoding is through the `LlamaPromptLookupDecoding` class. + +Just pass this as a draft model to the `Llama` class during initialization. + +```python +from llama_cpp import Llama +from llama_cpp.llama_speculative import LlamaPromptLookupDecoding + +llama = Llama( + model_path="path/to/model.gguf", + draft_model=LlamaPromptLookupDecoding(num_pred_tokens=10) # num_pred_tokens is the number of tokens to predict 10 is the default and generally good for gpu, 2 performs better for cpu-only machines. +) +``` + +### Embeddings + +To generate text embeddings use [`create_embedding`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.create_embedding) or [`embed`](https://llama-cpp-python.readthedocs.io/en/latest/api-reference/#llama_cpp.Llama.embed). Note that you must pass `embedding=True` to the constructor upon model creation for these to work properly. + +```python +import llama_cpp + +llm = llama_cpp.Llama(model_path="path/to/model.gguf", embedding=True) + +embeddings = llm.create_embedding("Hello, world!") + +# or create multiple embeddings at once + +embeddings = llm.create_embedding(["Hello, world!", "Goodbye, world!"]) +``` + +There are two primary notions of embeddings in a Transformer-style model: *token level* and *sequence level*. Sequence level embeddings are produced by "pooling" token level embeddings together, usually by averaging them or using the first token. + +Models that are explicitly geared towards embeddings will usually return sequence level embeddings by default, one for each input string. Non-embedding models such as those designed for text generation will typically return only token level embeddings, one for each token in each sequence. Thus the dimensionality of the return type will be one higher for token level embeddings. + +It is possible to control pooling behavior in some cases using the `pooling_type` flag on model creation. You can ensure token level embeddings from any model using `LLAMA_POOLING_TYPE_NONE`. The reverse, getting a generation oriented model to yield sequence level embeddings is currently not possible, but you can always do the pooling manually. + ### Adjusting the Context Window The context window of the Llama models determines the maximum number of tokens that can be processed at once. By default, this is set to 512 tokens, but can be adjusted based on your requirements. @@ -293,7 +648,6 @@ For instance, if you want to work with larger contexts, you can expand the conte llm = Llama(model_path="./models/7B/llama-model.gguf", n_ctx=2048) ``` - ## OpenAI Compatible Web Server `llama-cpp-python` offers a web server which aims to act as a drop-in replacement for the OpenAI API. @@ -302,14 +656,14 @@ This allows you to use llama.cpp compatible models with any OpenAI compatible cl To install the server package and get started: ```bash -pip install llama-cpp-python[server] +pip install 'llama-cpp-python[server]' python3 -m llama_cpp.server --model models/7B/llama-model.gguf ``` Similar to Hardware Acceleration section above, you can also install with GPU (cuBLAS) support like this: ```bash -CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python[server] +CMAKE_ARGS="-DGGML_CUDA=on" FORCE_CMAKE=1 pip install 'llama-cpp-python[server]' python3 -m llama_cpp.server --model models/7B/llama-model.gguf --n_gpu_layers 35 ``` @@ -327,11 +681,18 @@ python3 -m llama_cpp.server --model models/7B/llama-model.gguf --chat_format cha That will format the prompt according to how model expects it. You can find the prompt format in the model card. For possible options, see [llama_cpp/llama_chat_format.py](llama_cpp/llama_chat_format.py) and look for lines starting with "@register_chat_format". +If you have `huggingface-hub` installed, you can also use the `--hf_model_repo_id` flag to load a model from the Hugging Face Hub. + +```bash +python3 -m llama_cpp.server --hf_model_repo_id Qwen/Qwen2-0.5B-Instruct-GGUF --model '*q8_0.gguf' +``` + ### Web Server Features - [Local Copilot replacement](https://llama-cpp-python.readthedocs.io/en/latest/server/#code-completion) - [Function Calling support](https://llama-cpp-python.readthedocs.io/en/latest/server/#function-calling) - [Vision API support](https://llama-cpp-python.readthedocs.io/en/latest/server/#multimodal-models) +- [Multiple Models](https://llama-cpp-python.readthedocs.io/en/latest/server/#configuration-and-multi-model-support) ## Docker image @@ -340,7 +701,8 @@ A Docker image is available on [GHCR](https://ghcr.io/abetlen/llama-cpp-python). ```bash docker run --rm -it -p 8000:8000 -v /path/to/models:/models -e MODEL=/models/llama-model.gguf ghcr.io/abetlen/llama-cpp-python:latest ``` -[Docker on termux (requires root)](https://gist.github.com/FreddieOliveira/efe850df7ff3951cb62d74bd770dce27) is currently the only known way to run this on phones, see [termux support issue](https://github.com/abetlen/llama-cpp-python/issues/389) + +[Docker on termux (requires root)](https://gist.github.com/FreddieOliveira/efe850df7ff3951cb62d74bd770dce27) is currently the only known way to run this on phones, see [termux support issue](https://github.com/abetlen/llama-cpp-python/issues/389) ## Low-level API @@ -352,23 +714,22 @@ The entire low-level API can be found in [llama_cpp/llama_cpp.py](https://github Below is a short example demonstrating how to use the low-level API to tokenize a prompt: ```python ->>> import llama_cpp ->>> import ctypes ->>> llama_cpp.llama_backend_init(numa=False) # Must be called once at the start of each program ->>> params = llama_cpp.llama_context_default_params() +import llama_cpp +import ctypes +llama_cpp.llama_backend_init(False) # Must be called once at the start of each program +params = llama_cpp.llama_context_default_params() # use bytes for char * params ->>> model = llama_cpp.llama_load_model_from_file(b"./models/7b/llama-model.gguf", params) ->>> ctx = llama_cpp.llama_new_context_with_model(model, params) ->>> max_tokens = params.n_ctx +model = llama_cpp.llama_load_model_from_file(b"./models/7b/llama-model.gguf", params) +ctx = llama_cpp.llama_new_context_with_model(model, params) +max_tokens = params.n_ctx # use ctypes arrays for array params ->>> tokens = (llama_cpp.llama_token * int(max_tokens))() ->>> n_tokens = llama_cpp.llama_tokenize(ctx, b"Q: Name the planets in the solar system? A: ", tokens, max_tokens, add_bos=llama_cpp.c_bool(True)) ->>> llama_cpp.llama_free(ctx) +tokens = (llama_cpp.llama_token * int(max_tokens))() +n_tokens = llama_cpp.llama_tokenize(ctx, b"Q: Name the planets in the solar system? A: ", tokens, max_tokens, llama_cpp.c_bool(True)) +llama_cpp.llama_free(ctx) ``` Check out the [examples folder](examples/low_level_api) for more examples of using the low-level API. - ## Documentation Documentation is available via [https://llama-cpp-python.readthedocs.io/](https://llama-cpp-python.readthedocs.io/). @@ -391,15 +752,32 @@ pip install --upgrade pip pip install -e . # if you want to use the fastapi / openapi server -pip install -e .[server] +pip install -e '.[server]' # to install all optional dependencies -pip install -e .[all] +pip install -e '.[all]' # to clear the local build cache make clean ``` +Now try running the tests + +```bash +pytest +``` + +There's a `Makefile` available with useful targets. +A typical workflow would look like this: + +```bash +make build +make test +``` + +You can also test out specific commits of `llama.cpp` by checking out the desired commit in the `vendor/llama.cpp` submodule and then running `make clean` and `pip install -e .` again. Any changes in the `llama.h` API will require +changes to the `llama_cpp/llama_cpp.py` file to match the new API (additional changes may be required elsewhere). + ## FAQ ### Are there pre-built binaries / binary wheels available? diff --git a/docker/cuda_simple/Dockerfile b/docker/cuda_simple/Dockerfile index a9e51cdc1..0bbf20ffe 100644 --- a/docker/cuda_simple/Dockerfile +++ b/docker/cuda_simple/Dockerfile @@ -1,4 +1,4 @@ -ARG CUDA_IMAGE="12.1.1-devel-ubuntu22.04" +ARG CUDA_IMAGE="12.5.0-devel-ubuntu22.04" FROM nvidia/cuda:${CUDA_IMAGE} # We need to set the host to 0.0.0.0 to allow outside access @@ -15,13 +15,13 @@ COPY . . # setting build related env vars ENV CUDA_DOCKER_ARCH=all -ENV LLAMA_CUBLAS=1 +ENV GGML_CUDA=1 # Install depencencies RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context # Install llama-cpp-python (build with cuda) -RUN CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python +RUN CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python # Run the server CMD python3 -m llama_cpp.server diff --git a/docker/open_llama/Dockerfile b/docker/open_llama/Dockerfile index 23e37d4d4..f05e66ef2 100644 --- a/docker/open_llama/Dockerfile +++ b/docker/open_llama/Dockerfile @@ -1,5 +1,5 @@ # Define the image argument and provide a default value -ARG IMAGE=python:3-slim-bullseye +ARG IMAGE=python:3-slim-bookworm # Use the image as specified FROM ${IMAGE} @@ -12,19 +12,21 @@ RUN apt-get update && apt-get upgrade -y && apt-get install -y --no-install-reco python3 \ python3-pip \ ninja-build \ - build-essential + build-essential \ + && apt-get clean \ + && rm -rf /var/lib/apt/lists/* RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context # Perform the conditional installations based on the image RUN echo "Image: ${IMAGE}" && \ - if [ "${IMAGE}" = "python:3-slim-bullseye" ] ; then \ + if [ "${IMAGE}" = "python:3-slim-bookworm" ] ; then \ echo "OpenBLAS install:" && \ apt-get install -y --no-install-recommends libopenblas-dev && \ - LLAMA_OPENBLAS=1 pip install llama-cpp-python --verbose; \ + CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install llama-cpp-python --verbose; \ else \ echo "CuBLAS install:" && \ - LLAMA_CUBLAS=1 pip install llama-cpp-python --verbose; \ + CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --verbose; \ fi # Clean up apt cache diff --git a/docker/openblas_simple/Dockerfile b/docker/openblas_simple/Dockerfile index e213518b9..5ee667dc0 100644 --- a/docker/openblas_simple/Dockerfile +++ b/docker/openblas_simple/Dockerfile @@ -1,4 +1,4 @@ -FROM python:3-slim-bullseye +FROM python:3-slim-bookworm # We need to set the host to 0.0.0.0 to allow outside access ENV HOST 0.0.0.0 @@ -6,10 +6,13 @@ ENV HOST 0.0.0.0 COPY . . # Install the package -RUN apt update && apt install -y libopenblas-dev ninja-build build-essential pkg-config +RUN apt update && apt install -y libopenblas-dev ninja-build build-essential pkg-config \ + && apt-get clean \ + && rm -rf /var/lib/apt/lists/* /tmp/* + RUN python -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context -RUN CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS" pip install llama_cpp_python --verbose +RUN CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install llama_cpp_python --verbose # Run the server CMD python3 -m llama_cpp.server diff --git a/docker/simple/Dockerfile b/docker/simple/Dockerfile index 41cc68ea2..3594df1a5 100644 --- a/docker/simple/Dockerfile +++ b/docker/simple/Dockerfile @@ -1,5 +1,5 @@ # Define the image argument and provide a default value -ARG IMAGE=python:3-slim-bullseye +ARG IMAGE=python:3-slim-bookworm # Use the image as specified FROM ${IMAGE} @@ -13,7 +13,9 @@ RUN apt-get update && apt-get upgrade -y && apt-get install -y --no-install-reco python3-pip \ ninja-build \ libopenblas-dev \ - build-essential + build-essential \ + && apt-get clean \ + && rm -rf /var/lib/apt/lists/* /tmp/* RUN mkdir /app WORKDIR /app @@ -21,7 +23,9 @@ COPY . /app RUN python3 -m pip install --upgrade pip -RUN make deps && make build && make clean +RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette pydantic-settings starlette-context + +RUN pip install llama-cpp-python --verbose; # Set environment variable for the host ENV HOST=0.0.0.0 diff --git a/docs/api-reference.md b/docs/api-reference.md index 562410fe1..ab51ef754 100644 --- a/docs/api-reference.md +++ b/docs/api-reference.md @@ -21,11 +21,13 @@ High-level Python bindings for llama.cpp. - create_completion - __call__ - create_chat_completion + - create_chat_completion_openai_v1 - set_cache - save_state - load_state - token_bos - token_eos + - from_pretrained show_root_heading: true ::: llama_cpp.LlamaGrammar diff --git a/docs/icon.svg b/docs/icon.svg new file mode 100644 index 000000000..9f2d08753 --- /dev/null +++ b/docs/icon.svg @@ -0,0 +1,5 @@ + + + + + diff --git a/docs/install/macos.md b/docs/install/macos.md index 240422832..e006fc0a3 100644 --- a/docs/install/macos.md +++ b/docs/install/macos.md @@ -30,7 +30,7 @@ conda activate llama *(you needed xcode installed in order pip to build/compile the C++ code)* ``` pip uninstall llama-cpp-python -y -CMAKE_ARGS="-DLLAMA_METAL=on" pip install -U llama-cpp-python --no-cache-dir +CMAKE_ARGS="-DGGML_METAL=on" pip install -U llama-cpp-python --no-cache-dir pip install 'llama-cpp-python[server]' # you should now have llama-cpp-python v0.1.62 or higher installed diff --git a/docs/server.md b/docs/server.md index 4e1e56228..cd6f86c51 100644 --- a/docs/server.md +++ b/docs/server.md @@ -32,6 +32,12 @@ python3 -m llama_cpp.server --help NOTE: All server options are also available as environment variables. For example, `--model` can be set by setting the `MODEL` environment variable. +Check out the server config reference below settings for more information on the available options. +CLI arguments and environment variables are available for all of the fields defined in [`ServerSettings`](#llama_cpp.server.settings.ServerSettings) and [`ModelSettings`](#llama_cpp.server.settings.ModelSettings) + +Additionally the server supports configuration check out the [configuration section](#configuration-and-multi-model-support) for more information and examples. + + ## Guides ### Code Completion @@ -70,12 +76,14 @@ Function calling is completely compatible with the OpenAI function calling API a You'll first need to download one of the available function calling models in GGUF format: -- [functionary-7b-v1](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) +- [functionary](https://huggingface.co/meetkai) + +Then when you run the server you'll need to also specify either `functionary-v1` or `functionary-v2` chat_format. -Then when you run the server you'll need to also specify the `functionary` chat_format +Note that since functionary requires a HF Tokenizer due to discrepancies between llama.cpp and HuggingFace's tokenizers as mentioned [here](https://github.com/abetlen/llama-cpp-python/blob/main?tab=readme-ov-file#function-calling), you will need to pass in the path to the tokenizer too. The tokenizer files are already included in the respective HF repositories hosting the gguf files. ```bash -python3 -m llama_cpp.server --model --chat_format functionary +python3 -m llama_cpp.server --model --chat_format functionary-v2 --hf_pretrained_model_name_or_path ``` Check out this [example notebook](https://github.com/abetlen/llama-cpp-python/blob/main/examples/notebooks/Functions.ipynb) for a walkthrough of some interesting use cases for function calling. @@ -90,6 +98,8 @@ You'll first need to download one of the available multi-modal models in GGUF fo - [llava-v1.5-7b](https://huggingface.co/mys/ggml_llava-v1.5-7b) - [llava-v1.5-13b](https://huggingface.co/mys/ggml_llava-v1.5-13b) - [bakllava-1-7b](https://huggingface.co/mys/ggml_bakllava-1) +- [llava-v1.6-34b](https://huggingface.co/cjpais/llava-v1.6-34B-gguf) +- [moondream2](https://huggingface.co/vikhyatk/moondream2) Then when you run the server you'll need to also specify the path to the clip model used for image embedding and the `llava-1-5` chat_format @@ -121,4 +131,92 @@ response = client.chat.completions.create( ], ) print(response) -``` \ No newline at end of file +``` + +## Configuration and Multi-Model Support + +The server supports configuration via a JSON config file that can be passed using the `--config_file` parameter or the `CONFIG_FILE` environment variable. + +```bash +python3 -m llama_cpp.server --config_file +``` + +Config files support all of the server and model options supported by the cli and environment variables however instead of only a single model the config file can specify multiple models. + +The server supports routing requests to multiple models based on the `model` parameter in the request which matches against the `model_alias` in the config file. + +At the moment only a single model is loaded into memory at, the server will automatically load and unload models as needed. + +```json +{ + "host": "0.0.0.0", + "port": 8080, + "models": [ + { + "model": "models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf", + "model_alias": "gpt-3.5-turbo", + "chat_format": "chatml", + "n_gpu_layers": -1, + "offload_kqv": true, + "n_threads": 12, + "n_batch": 512, + "n_ctx": 2048 + }, + { + "model": "models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf", + "model_alias": "gpt-4", + "chat_format": "chatml", + "n_gpu_layers": -1, + "offload_kqv": true, + "n_threads": 12, + "n_batch": 512, + "n_ctx": 2048 + }, + { + "model": "models/ggml_llava-v1.5-7b/ggml-model-q4_k.gguf", + "model_alias": "gpt-4-vision-preview", + "chat_format": "llava-1-5", + "clip_model_path": "models/ggml_llava-v1.5-7b/mmproj-model-f16.gguf", + "n_gpu_layers": -1, + "offload_kqv": true, + "n_threads": 12, + "n_batch": 512, + "n_ctx": 2048 + }, + { + "model": "models/mistral-7b-v0.1-GGUF/ggml-model-Q4_K.gguf", + "model_alias": "text-davinci-003", + "n_gpu_layers": -1, + "offload_kqv": true, + "n_threads": 12, + "n_batch": 512, + "n_ctx": 2048 + }, + { + "model": "models/replit-code-v1_5-3b-GGUF/replit-code-v1_5-3b.Q4_0.gguf", + "model_alias": "copilot-codex", + "n_gpu_layers": -1, + "offload_kqv": true, + "n_threads": 12, + "n_batch": 1024, + "n_ctx": 9216 + } + ] +} +``` + +The config file format is defined by the [`ConfigFileSettings`](#llama_cpp.server.settings.ConfigFileSettings) class. + +## Server Options Reference + +::: llama_cpp.server.settings.ConfigFileSettings + options: + show_if_no_docstring: true + +::: llama_cpp.server.settings.ServerSettings + options: + show_if_no_docstring: true + +::: llama_cpp.server.settings.ModelSettings + options: + show_if_no_docstring: true diff --git a/examples/batch-processing/server.py b/examples/batch-processing/server.py new file mode 100644 index 000000000..0b36746f9 --- /dev/null +++ b/examples/batch-processing/server.py @@ -0,0 +1,31 @@ +"""llama-cpp-python server from scratch in a single file. +""" + +# import llama_cpp + +# path = b"../../models/Qwen1.5-0.5B-Chat-GGUF/qwen1_5-0_5b-chat-q8_0.gguf" + +# model_params = llama_cpp.llama_model_default_params() +# model = llama_cpp.llama_load_model_from_file(path, model_params) + +# if model is None: +# raise RuntimeError(f"Failed to load model from file: {path}") + + +# ctx_params = llama_cpp.llama_context_default_params() +# ctx = llama_cpp.llama_new_context_with_model(model, ctx_params) + +# if ctx is None: +# raise RuntimeError("Failed to create context") + + +from fastapi import FastAPI + +app = FastAPI() + +import openai.types.chat as types + + +@app.post("/v1/chat/completions") +def create_chat_completions(): + return {"message": "Hello World"} diff --git a/examples/gradio_chat/local.py b/examples/gradio_chat/local.py new file mode 100644 index 000000000..e16bf234a --- /dev/null +++ b/examples/gradio_chat/local.py @@ -0,0 +1,67 @@ +import llama_cpp +import llama_cpp.llama_tokenizer + +import gradio as gr + +llama = llama_cpp.Llama.from_pretrained( + repo_id="Qwen/Qwen1.5-0.5B-Chat-GGUF", + filename="*q8_0.gguf", + tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained( + "Qwen/Qwen1.5-0.5B" + ), + verbose=False, +) + +model = "gpt-3.5-turbo" + + +def predict(message, history): + messages = [] + + for user_message, assistant_message in history: + messages.append({"role": "user", "content": user_message}) + messages.append({"role": "assistant", "content": assistant_message}) + + messages.append({"role": "user", "content": message}) + + response = llama.create_chat_completion_openai_v1( + model=model, messages=messages, stream=True + ) + + text = "" + for chunk in response: + content = chunk.choices[0].delta.content + if content: + text += content + yield text + + +js = """function () { + gradioURL = window.location.href + if (!gradioURL.endsWith('?__theme=dark')) { + window.location.replace(gradioURL + '?__theme=dark'); + } +}""" + +css = """ +footer { + visibility: hidden; +} +full-height { + height: 100%; +} +""" + +with gr.Blocks(theme=gr.themes.Soft(), js=js, css=css, fill_height=True) as demo: + gr.ChatInterface( + predict, + fill_height=True, + examples=[ + "What is the capital of France?", + "Who was the first person on the moon?", + ], + ) + + +if __name__ == "__main__": + demo.launch() diff --git a/examples/gradio_chat/server.py b/examples/gradio_chat/server.py new file mode 100644 index 000000000..52061bea6 --- /dev/null +++ b/examples/gradio_chat/server.py @@ -0,0 +1,59 @@ +import gradio as gr + +from openai import OpenAI + +client = OpenAI(base_url="http://localhost:8000/v1", api_key="llama.cpp") + +model = "gpt-3.5-turbo" + + +def predict(message, history): + messages = [] + + for user_message, assistant_message in history: + messages.append({"role": "user", "content": user_message}) + messages.append({"role": "assistant", "content": assistant_message}) + + messages.append({"role": "user", "content": message}) + + response = client.chat.completions.create( + model=model, messages=messages, stream=True + ) + + text = "" + for chunk in response: + content = chunk.choices[0].delta.content + if content: + text += content + yield text + + +js = """function () { + gradioURL = window.location.href + if (!gradioURL.endsWith('?__theme=dark')) { + window.location.replace(gradioURL + '?__theme=dark'); + } +}""" + +css = """ +footer { + visibility: hidden; +} +full-height { + height: 100%; +} +""" + +with gr.Blocks(theme=gr.themes.Soft(), js=js, css=css, fill_height=True) as demo: + gr.ChatInterface( + predict, + fill_height=True, + examples=[ + "What is the capital of France?", + "Who was the first person on the moon?", + ], + ) + + +if __name__ == "__main__": + demo.launch() diff --git a/examples/hf_pull/main.py b/examples/hf_pull/main.py new file mode 100644 index 000000000..dfed17516 --- /dev/null +++ b/examples/hf_pull/main.py @@ -0,0 +1,36 @@ +import llama_cpp +import llama_cpp.llama_tokenizer + + +llama = llama_cpp.Llama.from_pretrained( + repo_id="Qwen/Qwen1.5-0.5B-Chat-GGUF", + filename="*q8_0.gguf", + tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained( + "Qwen/Qwen1.5-0.5B" + ), + verbose=False, +) + +response = llama.create_chat_completion( + messages=[{"role": "user", "content": "What is the capital of France?"}], + response_format={ + "type": "json_object", + "schema": { + "type": "object", + "properties": { + "country": {"type": "string"}, + "capital": {"type": "string"}, + }, + "required": ["country", "capital"], + }, + }, + stream=True, +) + +for chunk in response: + delta = chunk["choices"][0]["delta"] + if "content" not in delta: + continue + print(delta["content"], end="", flush=True) + +print() diff --git a/examples/high_level_api/fastapi_server.py b/examples/high_level_api/fastapi_server.py index 4b3189dd1..ee59767d6 100644 --- a/examples/high_level_api/fastapi_server.py +++ b/examples/high_level_api/fastapi_server.py @@ -9,7 +9,7 @@ Then run: ``` -uvicorn llama_cpp.server.app:app --reload +uvicorn --factory llama_cpp.server.app:create_app --reload ``` or @@ -24,6 +24,7 @@ To actually see the implementation of the server, see llama_cpp/server/app.py """ + import os import uvicorn diff --git a/examples/high_level_api/high_level_api_infill.py b/examples/high_level_api/high_level_api_infill.py new file mode 100644 index 000000000..282333e5a --- /dev/null +++ b/examples/high_level_api/high_level_api_infill.py @@ -0,0 +1,37 @@ +import argparse + +from llama_cpp import Llama + +parser = argparse.ArgumentParser() +parser.add_argument("-m", "--model", type=str, default="../models/7B/ggml-models.bin") +parser.add_argument("-p", "--prompt", type=str, default="def add(") +parser.add_argument("-s", "--suffix", type=str, default="\n return sum\n\n") +parser.add_argument("-i", "--spm-infill", action="store_true") +args = parser.parse_args() + +llm = Llama(model_path=args.model, n_gpu_layers=-1, spm_infill=args.spm_infill) + +output = llm.create_completion( + temperature=0.0, + repeat_penalty=1.0, + prompt=args.prompt, + suffix=args.suffix, +) + +# Models sometimes repeat suffix in response, attempt to filter that +response = output["choices"][0]["text"] +response_stripped = response.rstrip() +unwanted_response_suffix = args.suffix.rstrip() +unwanted_response_length = len(unwanted_response_suffix) + +filtered = False +if ( + unwanted_response_suffix + and response_stripped[-unwanted_response_length:] == unwanted_response_suffix +): + response = response_stripped[:-unwanted_response_length] + filtered = True + +print( + f"Fill-in-Middle completion{' (filtered)' if filtered else ''}:\n\n{args.prompt}\033[32m{response}\033[{'33' if filtered else '0'}m{args.suffix}\033[0m" +) diff --git a/examples/low_level_api/Chat.py b/examples/low_level_api/Chat.py index fcef8cd80..a755089b2 100644 --- a/examples/low_level_api/Chat.py +++ b/examples/low_level_api/Chat.py @@ -3,10 +3,12 @@ from common import GptParams from low_level_api_chat_cpp import LLaMAInteract + def env_or_def(env, default): - if (env in os.environ): - return os.environ[env] - return default + if env in os.environ: + return os.environ[env] + return default + AI_NAME = env_or_def("AI_NAME", "ChatLLaMa") MODEL = env_or_def("MODEL", "./models/llama-13B/ggml-model.bin") @@ -15,10 +17,10 @@ def env_or_def(env, default): N_THREAD = int(env_or_def("N_THREAD", "8")) today = datetime.datetime.today() -DATE_YEAR=today.strftime("%Y") -DATE_TIME=today.strftime("%H:%M") +DATE_YEAR = today.strftime("%Y") +DATE_TIME = today.strftime("%H:%M") -prompt=f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}. +prompt = f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}. {AI_NAME} is helpful, kind, honest, friendly, good at writing and never fails to answer {USER_NAME}'s requests immediately and with details and precision. There are no annotations like (30 seconds passed...) or (to himself), just what {USER_NAME} and {AI_NAME} say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. @@ -45,27 +47,29 @@ def env_or_def(env, default): {AI_NAME}: Blue. {USER_NAME}: What time is it? {AI_NAME}: It is {DATE_TIME}. -{USER_NAME}:""" + " ".join(sys.argv[1:]) +{USER_NAME}:""" + " ".join( + sys.argv[1:] +) print("Loading model...") params = GptParams( - n_ctx=2048, - temp=0.7, - top_k=40, - top_p=0.5, - repeat_last_n=256, - n_batch=1024, - repeat_penalty=1.17647, - model=MODEL, - n_threads=N_THREAD, - n_predict=N_PREDICTS, - use_color=True, - interactive=True, - antiprompt=[f"{USER_NAME}:"], - input_prefix=" ", - input_suffix=f"{AI_NAME}:", - prompt=prompt, + n_ctx=2048, + temp=0.7, + top_k=40, + top_p=0.5, + repeat_last_n=256, + n_batch=1024, + repeat_penalty=1.17647, + model=MODEL, + n_threads=N_THREAD, + n_predict=N_PREDICTS, + use_color=True, + interactive=True, + antiprompt=[f"{USER_NAME}:"], + input_prefix=" ", + input_suffix=f"{AI_NAME}:", + prompt=prompt, ) with LLaMAInteract(params) as m: - m.interact() + m.interact() diff --git a/examples/low_level_api/Miku.py b/examples/low_level_api/Miku.py index eb9a2cfa9..e072ab1b1 100644 --- a/examples/low_level_api/Miku.py +++ b/examples/low_level_api/Miku.py @@ -3,10 +3,12 @@ from common import GptParams from low_level_api_chat_cpp import LLaMAInteract + def env_or_def(env, default): - if (env in os.environ): - return os.environ[env] - return default + if env in os.environ: + return os.environ[env] + return default + AI_NAME = env_or_def("AI_NAME", "Miku") MODEL = env_or_def("MODEL", "./models/llama-13B/ggml-model.bin") @@ -14,7 +16,7 @@ def env_or_def(env, default): N_PREDICTS = int(env_or_def("N_PREDICTS", "4096")) N_THREAD = int(env_or_def("N_THREAD", "0")) -prompt=f"""This is a transcript of a 1000 page, never ending conversation between {USER_NAME} and the cute and helpful AI assistant {AI_NAME}. {AI_NAME} is a girl who is an AI running on the users computer. +prompt = f"""This is a transcript of a 1000 page, never ending conversation between {USER_NAME} and the cute and helpful AI assistant {AI_NAME}. {AI_NAME} is a girl who is an AI running on the users computer. {AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next. {AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct she will ask the user for help. {AI_NAME} is a very helpful AI and will help the user with anything they need, she is also very friendly and will try to make the user feel better if they are sad. @@ -32,28 +34,30 @@ def env_or_def(env, default): {AI_NAME}: /think It sounds like {USER_NAME} is happy to have me as their assistant! I'm so happy too! ^_^ Glad that whole emotion thing didn't scare him off! {AI_NAME}: /think I wonder what {USER_NAME} likes to do in his free time? I should ask him about that! {AI_NAME}: What do you like to do in your free time? ^_^ -{USER_NAME}:""" + " ".join(sys.argv[1:]) +{USER_NAME}:""" + " ".join( + sys.argv[1:] +) print("Loading model...") params = GptParams( - n_batch=1024, - n_ctx=2048, - n_keep=-1, - repeat_last_n=256, - repeat_penalty=1.17647, - temp=0.7, - top_k=40, - top_p=0.5, - model=MODEL, - n_predict=N_PREDICTS, - use_color=True, - interactive=True, - antiprompt=[f"{USER_NAME}:"], - prompt=prompt, + n_batch=1024, + n_ctx=2048, + n_keep=-1, + repeat_last_n=256, + repeat_penalty=1.17647, + temp=0.7, + top_k=40, + top_p=0.5, + model=MODEL, + n_predict=N_PREDICTS, + use_color=True, + interactive=True, + antiprompt=[f"{USER_NAME}:"], + prompt=prompt, ) if N_THREAD > 0: - params.n_threads = N_THREAD + params.n_threads = N_THREAD with LLaMAInteract(params) as m: - m.interact() + m.interact() diff --git a/examples/low_level_api/ReasonAct.py b/examples/low_level_api/ReasonAct.py index 82e5c4487..1f2c59017 100644 --- a/examples/low_level_api/ReasonAct.py +++ b/examples/low_level_api/ReasonAct.py @@ -3,14 +3,16 @@ from common import GptParams from low_level_api_chat_cpp import LLaMAInteract + def env_or_def(env, default): - if (env in os.environ): - return os.environ[env] - return default + if env in os.environ: + return os.environ[env] + return default + MODEL = env_or_def("MODEL", "./models/llama-13B/ggml-model.bin") -prompt=f"""You run in a loop of Thought, Action, Observation. +prompt = f"""You run in a loop of Thought, Action, Observation. At the end of the loop either Answer or restate your Thought and Action. Use Thought to describe your thoughts about the question you have been asked. Use Action to run one of these actions available to you: @@ -27,23 +29,25 @@ def env_or_def(env, default): Question: What is capital of france? Thought: Do I need to use an action? No, I know the answer Answer: Paris is the capital of France -Question:""" + " ".join(sys.argv[1:]) +Question:""" + " ".join( + sys.argv[1:] +) print("Loading model...") params = GptParams( - interactive=True, - interactive_start=True, - top_k=10000, - temp=0.2, - repeat_penalty=1, - n_threads=7, - n_ctx=2048, - antiprompt=["Question:","Observation:"], - model=MODEL, - input_prefix=" ", - n_predict=-1, - prompt=prompt, + interactive=True, + interactive_start=True, + top_k=10000, + temp=0.2, + repeat_penalty=1, + n_threads=7, + n_ctx=2048, + antiprompt=["Question:", "Observation:"], + model=MODEL, + input_prefix=" ", + n_predict=-1, + prompt=prompt, ) with LLaMAInteract(params) as m: - m.interact() + m.interact() diff --git a/examples/low_level_api/common.py b/examples/low_level_api/common.py index 55d08db5f..a0212ff0d 100644 --- a/examples/low_level_api/common.py +++ b/examples/low_level_api/common.py @@ -65,82 +65,242 @@ class GptParams: # Set to "\nUser:" etc. # This is an alternative to input_prefix which always adds it, so it potentially duplicates "User:"" fix_prefix: str = "" - input_echo: bool = True, + input_echo: bool = (True,) # Default instructions for Alpaca # switch to "Human" and "Assistant" for Vicuna. # TODO: TBD how they are gonna handle this upstream - instruct_inp_prefix: str="\n\n### Instruction:\n\n" - instruct_inp_suffix: str="\n\n### Response:\n\n" + instruct_inp_prefix: str = "\n\n### Instruction:\n\n" + instruct_inp_suffix: str = "\n\n### Response:\n\n" -def gpt_params_parse(argv = None): - parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter) - parser.add_argument("-s", "--seed", type=int, default=-1, help="RNG seed (use random seed for <= 0)",dest="seed") - parser.add_argument("-t", "--threads", type=int, default=min(4, os.cpu_count() or 1), help="number of threads to use during computation",dest="n_threads") - parser.add_argument("-n", "--n_predict", type=int, default=128, help="number of tokens to predict (-1 = infinity)",dest="n_predict") - parser.add_argument("--n_parts", type=int, default=-1, help="number of model parts", dest="n_parts") - parser.add_argument("-c", "--ctx_size", type=int, default=512, help="size of the prompt context",dest="n_ctx") - parser.add_argument("-b", "--batch_size", type=int, default=8, help="batch size for prompt processing",dest="n_batch") - parser.add_argument("--keep", type=int, default=0, help="number of tokens to keep from the initial prompt",dest="n_keep") +def gpt_params_parse(argv=None): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + parser.add_argument( + "-s", + "--seed", + type=int, + default=-1, + help="RNG seed (use random seed for <= 0)", + dest="seed", + ) + parser.add_argument( + "-t", + "--threads", + type=int, + default=min(4, os.cpu_count() or 1), + help="number of threads to use during computation", + dest="n_threads", + ) + parser.add_argument( + "-n", + "--n_predict", + type=int, + default=128, + help="number of tokens to predict (-1 = infinity)", + dest="n_predict", + ) + parser.add_argument( + "--n_parts", type=int, default=-1, help="number of model parts", dest="n_parts" + ) + parser.add_argument( + "-c", + "--ctx_size", + type=int, + default=512, + help="size of the prompt context", + dest="n_ctx", + ) + parser.add_argument( + "-b", + "--batch_size", + type=int, + default=8, + help="batch size for prompt processing", + dest="n_batch", + ) + parser.add_argument( + "--keep", + type=int, + default=0, + help="number of tokens to keep from the initial prompt", + dest="n_keep", + ) parser.add_argument( "-l", "--logit-bias", type=str, - action='append', + action="append", help="--logit-bias TOKEN_ID(+/-)BIAS", - dest="logit_bias_str" - ) - parser.add_argument("--ignore-eos", action="store_true", help="ignore end of stream token and continue generating", dest="ignore_eos") - parser.add_argument("--top_k", type=int, default=40, help="top-k sampling",dest="top_k") - parser.add_argument("--top_p", type=float, default=0.95, help="top-p samplin",dest="top_p") - parser.add_argument("--tfs", type=float, default=1.0, help="tail free sampling, parameter z (1.0 = disabled)",dest="tfs_z") - parser.add_argument("--temp", type=float, default=0.80, help="temperature",dest="temp") - parser.add_argument("--repeat_penalty", type=float, default=1.10, help="penalize repeat sequence of tokens",dest="repeat_penalty") - parser.add_argument("--repeat_last_n", type=int, default=64, help="last n tokens to consider for penalize ",dest="repeat_last_n") - parser.add_argument("--frequency_penalty", type=float, default=0.0, help="repeat alpha frequency penalty (0.0 = disabled)",dest="tfs_z") - parser.add_argument("--presence_penalty", type=float, default=0.0, help="repeat alpha presence penalty (0.0 = disabled)",dest="presence_penalty") - parser.add_argument("--mirostat", type=float, default=1.0, help="use Mirostat sampling.",dest="mirostat") - parser.add_argument("--mirostat_ent", type=float, default=5.0, help="Mirostat target entropy, parameter tau represents the average surprise value",dest="mirostat_tau") - parser.add_argument("--mirostat_lr", type=float, default=0.1, help="Mirostat learning rate, parameter eta",dest="mirostat_eta") - - parser.add_argument("-m", "--model", type=str, default="./models/llama-7B/ggml-model.bin", help="model path",dest="model") - parser.add_argument("-p", "--prompt", type=str, default="", help="initial prompt",dest="prompt") - parser.add_argument("-f", "--file", type=str, default=None, help="file containing initial prompt to load",dest="file") - parser.add_argument("--session", type=str, default=None, help="file to cache model state in (may be large!)",dest="path_session") - parser.add_argument("--in-prefix", type=str, default="", help="string to prefix user inputs with", dest="input_prefix") - parser.add_argument("--in-suffix", type=str, default="", help="append to input", dest="input_suffix") + dest="logit_bias_str", + ) + parser.add_argument( + "--ignore-eos", + action="store_true", + help="ignore end of stream token and continue generating", + dest="ignore_eos", + ) + parser.add_argument( + "--top_k", type=int, default=40, help="top-k sampling", dest="top_k" + ) + parser.add_argument( + "--top_p", type=float, default=0.95, help="top-p samplin", dest="top_p" + ) + parser.add_argument( + "--tfs", + type=float, + default=1.0, + help="tail free sampling, parameter z (1.0 = disabled)", + dest="tfs_z", + ) + parser.add_argument( + "--temp", type=float, default=0.80, help="temperature", dest="temp" + ) + parser.add_argument( + "--repeat_penalty", + type=float, + default=1.10, + help="penalize repeat sequence of tokens", + dest="repeat_penalty", + ) + parser.add_argument( + "--repeat_last_n", + type=int, + default=64, + help="last n tokens to consider for penalize ", + dest="repeat_last_n", + ) + parser.add_argument( + "--frequency_penalty", + type=float, + default=0.0, + help="repeat alpha frequency penalty (0.0 = disabled)", + dest="tfs_z", + ) + parser.add_argument( + "--presence_penalty", + type=float, + default=0.0, + help="repeat alpha presence penalty (0.0 = disabled)", + dest="presence_penalty", + ) + parser.add_argument( + "--mirostat", + type=float, + default=1.0, + help="use Mirostat sampling.", + dest="mirostat", + ) + parser.add_argument( + "--mirostat_ent", + type=float, + default=5.0, + help="Mirostat target entropy, parameter tau represents the average surprise value", + dest="mirostat_tau", + ) + parser.add_argument( + "--mirostat_lr", + type=float, + default=0.1, + help="Mirostat learning rate, parameter eta", + dest="mirostat_eta", + ) + + parser.add_argument( + "-m", + "--model", + type=str, + default="./models/llama-7B/ggml-model.bin", + help="model path", + dest="model", + ) + parser.add_argument( + "-p", "--prompt", type=str, default=None, help="initial prompt", dest="prompt" + ) + parser.add_argument( + "-f", + "--file", + type=str, + default=None, + help="file containing initial prompt to load", + dest="file", + ) + parser.add_argument( + "--session", + type=str, + default=None, + help="file to cache model state in (may be large!)", + dest="path_session", + ) + parser.add_argument( + "--in-prefix", + type=str, + default="", + help="string to prefix user inputs with", + dest="input_prefix", + ) + parser.add_argument( + "--in-suffix", type=str, default="", help="append to input", dest="input_suffix" + ) parser.add_argument( "-r", "--reverse-prompt", type=str, - action='append', + action="append", help="poll user input upon seeing PROMPT (can be\nspecified more than once for multiple prompts).", - dest="antiprompt" + dest="antiprompt", + ) + + parser.add_argument( + "--lora", + type=str, + default="", + help="apply LoRA adapter (implies --no-mmap)", + dest="lora_adapter", + ) + parser.add_argument( + "--lora-base", + type=str, + default="", + help="optional model to use as a base for the layers modified by the LoRA adapter", + dest="lora_base", ) - - parser.add_argument("--lora", type=str, default="", help="apply LoRA adapter (implies --no-mmap)", dest="lora_adapter") - parser.add_argument("--lora-base", type=str, default="", help="optional model to use as a base for the layers modified by the LoRA adapter", dest="lora_base") - parser.add_argument("--memory_f32", action="store_false", help="use f32 instead of f16 for memory key+value",dest="memory_f16") - parser.add_argument("--random-prompt", action="store_true", help="start with a randomized prompt.", dest="random_prompt") + parser.add_argument( + "--memory_f32", + action="store_false", + help="use f32 instead of f16 for memory key+value", + dest="memory_f16", + ) + parser.add_argument( + "--random-prompt", + action="store_true", + help="start with a randomized prompt.", + dest="random_prompt", + ) parser.add_argument( "--color", action="store_true", help="colorise output to distinguish prompt and user input from generations", - dest="use_color" + dest="use_color", ) parser.add_argument( - "-i", "--interactive", action="store_true", help="run in interactive mode", dest="interactive" + "-i", + "--interactive", + action="store_true", + help="run in interactive mode", + dest="interactive", ) - + parser.add_argument("--embedding", action="store_true", help="", dest="embedding") parser.add_argument( "--interactive-first", action="store_true", help="run in interactive mode and wait for input right away", - dest="interactive_start" + dest="interactive_start", ) parser.add_argument( @@ -148,42 +308,84 @@ def gpt_params_parse(argv = None): "--instruct", action="store_true", help="run in instruction mode (use with Alpaca or Vicuna models)", - dest="instruct" + dest="instruct", + ) + parser.add_argument( + "--no-penalize-nl", + action="store_false", + help="do not penalize newline token", + dest="penalize_nl", + ) + parser.add_argument( + "--perplexity", + action="store_true", + help="compute perplexity over the prompt", + dest="perplexity", + ) + parser.add_argument( + "--no-mmap", + action="store_false", + help="do not memory-map model (slower load but may reduce pageouts if not using mlock)", + dest="use_mmap", + ) + parser.add_argument( + "--mlock", + action="store_true", + help="force system to keep model in RAM rather than swapping or compressing", + dest="use_mlock", + ) + parser.add_argument( + "--mtest", + action="store_true", + help="compute maximum memory usage", + dest="mem_test", + ) + parser.add_argument( + "--verbose-prompt", + action="store_true", + help="print prompt before generation", + dest="verbose_prompt", ) - parser.add_argument("--no-penalize-nl", action="store_false", help="do not penalize newline token", dest="penalize_nl") - parser.add_argument("--perplexity", action="store_true", help="compute perplexity over the prompt", dest="perplexity") - parser.add_argument("--no-mmap", action="store_false",help="do not memory-map model (slower load but may reduce pageouts if not using mlock)",dest="use_mmap") - parser.add_argument("--mlock", action="store_true",help="force system to keep model in RAM rather than swapping or compressing",dest="use_mlock") - parser.add_argument("--mtest", action="store_true",help="compute maximum memory usage",dest="mem_test") - parser.add_argument("--verbose-prompt", action="store_true",help="print prompt before generation",dest="verbose_prompt") - #Custom args - parser.add_argument("--fix-prefix", type=str, default="", help="append to input when generated n_predict tokens", dest="fix_prefix") - parser.add_argument("--input-noecho", action="store_false", help="dont output the input", dest="input_echo") + # Custom args + parser.add_argument( + "--fix-prefix", + type=str, + default="", + help="append to input when generated n_predict tokens", + dest="fix_prefix", + ) + parser.add_argument( + "--input-noecho", + action="store_false", + help="dont output the input", + dest="input_echo", + ) parser.add_argument( "--interactive-start", action="store_true", help="run in interactive mode", - dest="interactive" + dest="interactive", ) args = parser.parse_args(argv) - + logit_bias_str = args.logit_bias_str delattr(args, "logit_bias_str") params = GptParams(**vars(args)) - if (params.lora_adapter): + if params.lora_adapter: params.use_mmap = False - if (logit_bias_str != None): + if logit_bias_str != None: for i in logit_bias_str: - if (m := re.match(r"(\d+)([-+]\d+)", i)): + if m := re.match(r"(\d+)([-+]\d+)", i): params.logit_bias[int(m.group(1))] = float(m.group(2)) return params + def gpt_random_prompt(rng): return [ "So", @@ -198,5 +400,6 @@ def gpt_random_prompt(rng): "They", ][rng % 10] + if __name__ == "__main__": print(gpt_params_parse()) diff --git a/examples/low_level_api/low_level_api_chat_cpp.py b/examples/low_level_api/low_level_api_chat_cpp.py index 44b6d4a35..39081be17 100644 --- a/examples/low_level_api/low_level_api_chat_cpp.py +++ b/examples/low_level_api/low_level_api_chat_cpp.py @@ -10,6 +10,7 @@ You should also still be feeding the model with a "primer" prompt that shows it the expected format. """ + import ctypes import sys from time import time @@ -19,194 +20,257 @@ from common import GptParams, gpt_params_parse, gpt_random_prompt import util + # A LLaMA interactive session class LLaMAInteract: - def __init__(self, params: GptParams) -> None: - # input args - self.params = params - if self.params.path_session is None: - self.params.path_session = "" - if self.params.antiprompt is None: - self.params.antiprompt = "" - - if (self.params.perplexity): - raise NotImplementedError("""************ + def __init__(self, params: GptParams) -> None: + # input args + self.params = params + if self.params.path_session is None: + self.params.path_session = "" + if self.params.antiprompt is None: + self.params.antiprompt = "" + + if self.params.perplexity: + raise NotImplementedError( + """************ please use the 'perplexity' tool for perplexity calculations -************""") +************""" + ) - if (self.params.embedding): - raise NotImplementedError("""************ + if self.params.embedding: + raise NotImplementedError( + """************ please use the 'embedding' tool for embedding calculations -************""") +************""" + ) - if (self.params.n_ctx > 2048): - print(f"""warning: model does not support \ + if self.params.n_ctx > 2048: + print( + f"""warning: model does not support \ context sizes greater than 2048 tokens ({self.params.n_ctx} \ -specified) expect poor results""", file=sys.stderr) - - if (self.params.seed <= 0): - self.params.seed = int(time()) - - print(f"seed = {self.params.seed}", file=sys.stderr) - - if (self.params.random_prompt): - self.params.prompt = gpt_random_prompt(self.params.seed) - - # runtime args - self.input_consumed = 0 - self.n_past = 0 - self.n_session_consumed = 0 - self.first_antiprompt = [] - self.remaining_tokens = self.params.n_predict - self.output_echo = self.params.input_echo - self.multibyte_fix = [] - - # model load - self.lparams = llama_cpp.llama_context_default_params() - self.lparams.n_ctx = self.params.n_ctx - self.lparams.n_parts = self.params.n_parts - self.lparams.seed = self.params.seed - self.lparams.memory_f16 = self.params.memory_f16 - self.lparams.use_mlock = self.params.use_mlock - self.lparams.use_mmap = self.params.use_mmap - - self.model = llama_cpp.llama_load_model_from_file( - self.params.model.encode("utf8"), self.lparams) - self.ctx = llama_cpp.llama_new_context_with_model(self.model, self.lparams) - if (not self.ctx): - raise RuntimeError(f"error: failed to load model '{self.params.model}'") - - if (self.params.ignore_eos): - self.params.logit_bias[llama_cpp.llama_token_eos()] = -float("inf") - - if (len(self.params.lora_adapter) > 0): - if (llama_cpp.llama_apply_lora_from_file( - self.ctx, - self.params.lora_adapter.encode("utf8"), - self.params.lora_base.encode("utf8") if len(self.params.lora_base) > 0 else None, - self.params.n_threads - ) != 0): - print("error: failed to apply lora adapter") - return - - print(file=sys.stderr) - print(f"system_info: n_threads = {self.params.n_threads} / {cpu_count()} \ -| {llama_cpp.llama_print_system_info().decode('utf8')}", file=sys.stderr) - - # determine the required inference memory per token: - if (self.params.mem_test): - tmp = [0, 1, 2, 3] - llama_cpp.llama_eval(self.ctx, (llama_cpp.c_int * len(tmp))(*tmp), len(tmp), 0, self.n_threads) - llama_cpp.llama_print_timings(self.ctx) - self.exit() - return - - # create internal context - self.n_ctx = llama_cpp.llama_n_ctx(self.ctx) - - # Add a space in front of the first character to match OG llama tokenizer behavior - self.params.prompt = " " + self.params.prompt - - # Load prompt file - if (self.params.file): - with open(self.params.file) as f: - self.params.prompt = f.read() - - self.session_tokens: list[llama_cpp.llama_token] = [] - if (len(self.params.path_session) > 0): - print(f"attempting to load saved session from '{self.params.path_session}'", file=sys.stderr) - - if (path.exists(self.params.path_session)): - _session_tokens = (llama_cpp.llama_token * (self.params.n_ctx))() - _n_token_count_out = llama_cpp.c_size_t() - if (llama_cpp.llama_load_session_file( - self.ctx, - self.params.path_session.encode("utf8"), - _session_tokens, - self.params.n_ctx, - ctypes.byref(_n_token_count_out) - ) != 1): - print(f"error: failed to load session file '{self.params.path_session}'", file=sys.stderr) - return - _n_token_count_out = _n_token_count_out.value - self.session_tokens = _session_tokens[:_n_token_count_out] - print(f"loaded a session with prompt size of {_n_token_count_out} tokens", file=sys.stderr) - else: - print(f"session file does not exist, will create", file=sys.stderr) - - # tokenize the prompt - self.embd = [] - self.embd_inp = self._tokenize(self.params.prompt) - - if (len(self.embd_inp) > self.n_ctx - 4): - raise RuntimeError(f"error: prompt is too long ({len(self.embd_inp)} tokens, max {self.params.n_ctx - 4})") - - # debug message about similarity of saved session, if applicable - self.n_matching_session_tokens = 0 - if len(self.session_tokens) > 0: - for id in self.session_tokens: - if self.n_matching_session_tokens >= len(self.embd_inp) or id != self.embd_inp[self.n_matching_session_tokens]: - break - self.n_matching_session_tokens += 1 - - if self.n_matching_session_tokens >= len(self.embd_inp): - print(f"session file has exact match for prompt!") - elif self.n_matching_session_tokens < (len(self.embd_inp) / 2): - print(f"warning: session file has low similarity to prompt ({self.n_matching_session_tokens} / {len(self.embd_inp)} tokens); will mostly be reevaluated") - else: - print(f"session file matches {self.n_matching_session_tokens} / {len(self.embd_inp)} tokens of prompt") - - self.need_to_save_session = len(self.params.path_session) > 0 and self.n_matching_session_tokens < (len(self.embd_inp) * 3 / 4) - - # number of tokens to keep when resetting context - if (self.params.n_keep < 0 or self.params.n_keep > len(self.embd_inp) or self.params.instruct): - self.params.n_keep = len(self.embd_inp) - - self.inp_prefix = self._tokenize(self.params.instruct_inp_prefix) - self.inp_suffix = self._tokenize(self.params.instruct_inp_suffix, False) - - # in instruct mode, we inject a prefix and a suffix to each input by the user - self.antiecho = None - if (self.params.instruct): - self.params.interactive_start = True - _ptn = self._tokenize(self.params.instruct_inp_prefix.strip(), False) - self.first_antiprompt.append(_ptn) - self.antiecho = util.IterSearch(_ptn) - - # enable interactive mode if reverse prompt or interactive start is specified - if (len(self.params.antiprompt) != 0 or self.params.interactive_start): - self.params.interactive = True - - # determine newline token - self.llama_token_newline = self._tokenize("\n", False) - self.llama_token_eot = self._tokenize(" [end of text]\n", False) - - if (self.params.verbose_prompt): - print(f""" +specified) expect poor results""", + file=sys.stderr, + ) + + if self.params.seed <= 0: + self.params.seed = int(time()) + + print(f"seed = {self.params.seed}", file=sys.stderr) + + if self.params.random_prompt: + self.params.prompt = gpt_random_prompt(self.params.seed) + + # runtime args + self.input_consumed = 0 + self.n_past = 0 + self.n_session_consumed = 0 + self.first_antiprompt = [] + self.remaining_tokens = self.params.n_predict + self.output_echo = self.params.input_echo + self.multibyte_fix = [] + + # model load + self.lparams = llama_cpp.llama_model_default_params() + self.lparams.n_ctx = self.params.n_ctx + self.lparams.n_parts = self.params.n_parts + self.lparams.seed = self.params.seed + self.lparams.memory_f16 = self.params.memory_f16 + self.lparams.use_mlock = self.params.use_mlock + self.lparams.use_mmap = self.params.use_mmap + + self.model = llama_cpp.llama_load_model_from_file( + self.params.model.encode("utf8"), self.lparams + ) + + # Context Params. + self.cparams = llama_cpp.llama_context_default_params() + + self.ctx = llama_cpp.llama_new_context_with_model(self.model, self.cparams) + if not self.ctx: + raise RuntimeError(f"error: failed to load model '{self.params.model}'") + + if self.params.ignore_eos: + self.params.logit_bias[llama_cpp.llama_token_eos()] = -float("inf") + + if len(self.params.lora_adapter) > 0: + if ( + llama_cpp.llama_apply_lora_from_file( + self.ctx, + self.params.lora_adapter.encode("utf8"), + ( + self.params.lora_base.encode("utf8") + if len(self.params.lora_base) > 0 + else None + ), + self.params.n_threads, + ) + != 0 + ): + print("error: failed to apply lora adapter") + return + + print(file=sys.stderr) + print( + f"system_info: n_threads = {self.params.n_threads} / {cpu_count()} \ +| {llama_cpp.llama_print_system_info().decode('utf8')}", + file=sys.stderr, + ) + + # determine the required inference memory per token: + if self.params.mem_test: + tmp = [0, 1, 2, 3] + llama_cpp.llama_eval( + self.ctx, + (llama_cpp.c_int * len(tmp))(*tmp), + len(tmp), + 0, + self.n_threads, + ) + llama_cpp.llama_print_timings(self.ctx) + self.exit() + return + + # create internal context + self.n_ctx = llama_cpp.llama_n_ctx(self.ctx) + + # Add a space in front of the first character to match OG llama tokenizer behavior + self.params.prompt = " " + self.params.prompt + + # Load prompt file + if self.params.file: + with open(self.params.file) as f: + self.params.prompt = f.read() + + self.session_tokens: list[llama_cpp.llama_token] = [] + if len(self.params.path_session) > 0: + print( + f"attempting to load saved session from '{self.params.path_session}'", + file=sys.stderr, + ) + + if path.exists(self.params.path_session): + _session_tokens = (llama_cpp.llama_token * (self.params.n_ctx))() + _n_token_count_out = llama_cpp.c_size_t() + if ( + llama_cpp.llama_load_session_file( + self.ctx, + self.params.path_session.encode("utf8"), + _session_tokens, + self.params.n_ctx, + ctypes.byref(_n_token_count_out), + ) + != 1 + ): + print( + f"error: failed to load session file '{self.params.path_session}'", + file=sys.stderr, + ) + return + _n_token_count_out = _n_token_count_out.value + self.session_tokens = _session_tokens[:_n_token_count_out] + print( + f"loaded a session with prompt size of {_n_token_count_out} tokens", + file=sys.stderr, + ) + else: + print(f"session file does not exist, will create", file=sys.stderr) + + # tokenize the prompt + self.embd = [] + self.embd_inp = self._tokenize(self.params.prompt) + + if len(self.embd_inp) > self.n_ctx - 4: + raise RuntimeError( + f"error: prompt is too long ({len(self.embd_inp)} tokens, max {self.params.n_ctx - 4})" + ) + + # debug message about similarity of saved session, if applicable + self.n_matching_session_tokens = 0 + if len(self.session_tokens) > 0: + for id in self.session_tokens: + if ( + self.n_matching_session_tokens >= len(self.embd_inp) + or id != self.embd_inp[self.n_matching_session_tokens] + ): + break + self.n_matching_session_tokens += 1 + + if self.n_matching_session_tokens >= len(self.embd_inp): + print(f"session file has exact match for prompt!") + elif self.n_matching_session_tokens < (len(self.embd_inp) / 2): + print( + f"warning: session file has low similarity to prompt ({self.n_matching_session_tokens} / {len(self.embd_inp)} tokens); will mostly be reevaluated" + ) + else: + print( + f"session file matches {self.n_matching_session_tokens} / {len(self.embd_inp)} tokens of prompt" + ) + + self.need_to_save_session = len( + self.params.path_session + ) > 0 and self.n_matching_session_tokens < (len(self.embd_inp) * 3 / 4) + + # number of tokens to keep when resetting context + if ( + self.params.n_keep < 0 + or self.params.n_keep > len(self.embd_inp) + or self.params.instruct + ): + self.params.n_keep = len(self.embd_inp) + + self.inp_prefix = self._tokenize(self.params.instruct_inp_prefix) + self.inp_suffix = self._tokenize(self.params.instruct_inp_suffix, False) + + # in instruct mode, we inject a prefix and a suffix to each input by the user + self.antiecho = None + if self.params.instruct: + self.params.interactive_start = True + _ptn = self._tokenize(self.params.instruct_inp_prefix.strip(), False) + self.first_antiprompt.append(_ptn) + self.antiecho = util.IterSearch(_ptn) + + # enable interactive mode if reverse prompt or interactive start is specified + if len(self.params.antiprompt) != 0 or self.params.interactive_start: + self.params.interactive = True + + # determine newline token + self.llama_token_newline = self._tokenize("\n", False) + self.llama_token_eot = self._tokenize(" [end of text]\n", False) + + if self.params.verbose_prompt: + print( + f""" prompt: '{self.params.prompt}' -number of tokens in prompt = {len(self.embd_inp)}""", file=sys.stderr) - - for i in range(len(self.embd_inp)): - print(f"{self.embd_inp[i]} -> '{self.token_to_str(self.embd_inp[i])}'", file=sys.stderr) - - if (self.params.n_keep > 0): - print("static prompt based on n_keep: '") - for i in range(self.params.n_keep): - print(self.token_to_str(self.embd_inp[i]), file=sys.stderr) - print("'", file=sys.stderr) - print(file=sys.stderr) - - if (self.params.interactive): - print("interactive mode on.", file=sys.stderr) - - if (len(self.params.antiprompt) > 0): - for antiprompt in self.params.antiprompt: - print(f"Reverse prompt: '{antiprompt}'", file=sys.stderr) - - if len(self.params.input_prefix) > 0: - print(f"Input prefix: '{self.params.input_prefix}'", file=sys.stderr) - - print(f"""sampling: repeat_last_n = {self.params.repeat_last_n},\ +number of tokens in prompt = {len(self.embd_inp)}""", + file=sys.stderr, + ) + + for i in range(len(self.embd_inp)): + print( + f"{self.embd_inp[i]} -> '{self.token_to_str(self.embd_inp[i])}'", + file=sys.stderr, + ) + + if self.params.n_keep > 0: + print("static prompt based on n_keep: '") + for i in range(self.params.n_keep): + print(self.token_to_str(self.embd_inp[i]), file=sys.stderr) + print("'", file=sys.stderr) + print(file=sys.stderr) + + if self.params.interactive: + print("interactive mode on.", file=sys.stderr) + + if len(self.params.antiprompt) > 0: + for antiprompt in self.params.antiprompt: + print(f"Reverse prompt: '{antiprompt}'", file=sys.stderr) + + if len(self.params.input_prefix) > 0: + print(f"Input prefix: '{self.params.input_prefix}'", file=sys.stderr) + + print( + f"""sampling: repeat_last_n = {self.params.repeat_last_n},\ repeat_penalty = {self.params.repeat_penalty},\ presence_penalty = {self.params.presence_penalty},\ frequency_penalty = {self.params.frequency_penalty},\ @@ -224,77 +288,96 @@ def __init__(self, params: GptParams) -> None: n_predict = {self.params.n_predict},\ n_keep = {self.params.n_keep} -""", file=sys.stderr) +""", + file=sys.stderr, + ) - # determine antiprompt tokens - for i in self.params.antiprompt: - self.first_antiprompt.append(self._tokenize(i, False)) + # determine antiprompt tokens + for i in self.params.antiprompt: + self.first_antiprompt.append(self._tokenize(i, False)) - self.last_n_tokens = [0]*self.n_ctx #TODO: deque doesnt support slices + self.last_n_tokens = [0] * self.n_ctx # TODO: deque doesnt support slices - if (params.interactive): - print("""== Running in interactive mode. == + if params.interactive: + print( + """== Running in interactive mode. == - Press Ctrl+C to interject at any time. - Press Return to return control to LLaMa. - If you want to submit another line, end your input in '\\'. -""", file=sys.stderr) - self.set_color(util.CONSOLE_COLOR_PROMPT) - - # tokenize a prompt - def _tokenize(self, prompt, bos=True): - _arr = (llama_cpp.llama_token * ((len(prompt) + 1) * 4))() - _n = llama_cpp.llama_tokenize(self.ctx, prompt.encode("utf8", errors="ignore"), _arr, len(_arr), bos) - return _arr[:_n] - - def set_color(self, c): - if (self.params.use_color): - print(c, end="") - - def use_antiprompt(self): - return len(self.first_antiprompt) > 0 - - # generate tokens - def generate(self): - while self.remaining_tokens > 0 or self.params.interactive or self.params.n_predict == -1: - # predict - if len(self.embd) > 0: - # infinite text generation via context swapping - # if we run out of context: - # - take the n_keep first tokens from the original prompt (via n_past) - # - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch - if (self.n_past + len(self.embd) > self.n_ctx): - n_left = self.n_past - self.params.n_keep - self.n_past = self.params.n_keep - - # insert n_left/2 tokens at the start of embd from last_n_tokens - _insert = self.last_n_tokens[ - self.n_ctx - int(n_left/2) - len(self.embd):-len(self.embd) - ] - self.embd = _insert + self.embd - self.params.path_session = "" - - # try to reuse a matching prefix from the loaded session instead of re-eval (via n_past) - if self.n_session_consumed < len(self.session_tokens): - for i in range(len(self.embd)): - if self.embd[i] != self.session_tokens[self.n_session_consumed]: - self.session_tokens = self.session_tokens[:self.n_session_consumed] - break - - self.n_past += 1 - self.n_session_consumed += 1 - - if self.n_session_consumed >= len(self.session_tokens): - i += 1 - break - - if i > 0: - self.embd = self.embd[i:] - - # evaluate tokens in batches - # embd is typically prepared beforehand to fit within a batch, but not always - #TODO BUG: The batching code causes nonsensical generation - """for i in range(0, len(self.embd), self.params.n_batch): +""", + file=sys.stderr, + ) + self.set_color(util.CONSOLE_COLOR_PROMPT) + + # tokenize a prompt + def _tokenize(self, prompt, bos=True): + _arr = (llama_cpp.llama_token * ((len(prompt) + 1) * 4))() + _n = llama_cpp.llama_tokenize( + self.model, + prompt.encode("utf8", errors="ignore"), + len(prompt), + _arr, + len(_arr), + bos, + False, + ) + return _arr[:_n] + + def set_color(self, c): + if self.params.use_color: + print(c, end="") + + def use_antiprompt(self): + return len(self.first_antiprompt) > 0 + + # generate tokens + def generate(self): + while ( + self.remaining_tokens > 0 + or self.params.interactive + or self.params.n_predict == -1 + ): + # predict + if len(self.embd) > 0: + # infinite text generation via context swapping + # if we run out of context: + # - take the n_keep first tokens from the original prompt (via n_past) + # - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch + if self.n_past + len(self.embd) > self.n_ctx: + n_left = self.n_past - self.params.n_keep + self.n_past = self.params.n_keep + + # insert n_left/2 tokens at the start of embd from last_n_tokens + _insert = self.last_n_tokens[ + self.n_ctx - int(n_left / 2) - len(self.embd) : -len(self.embd) + ] + self.embd = _insert + self.embd + self.params.path_session = "" + + # try to reuse a matching prefix from the loaded session instead of re-eval (via n_past) + if self.n_session_consumed < len(self.session_tokens): + for i in range(len(self.embd)): + if self.embd[i] != self.session_tokens[self.n_session_consumed]: + self.session_tokens = self.session_tokens[ + : self.n_session_consumed + ] + break + + self.n_past += 1 + self.n_session_consumed += 1 + + if self.n_session_consumed >= len(self.session_tokens): + i += 1 + break + + if i > 0: + self.embd = self.embd[i:] + + # evaluate tokens in batches + # embd is typically prepared beforehand to fit within a batch, but not always + # TODO BUG: The batching code causes nonsensical generation + """for i in range(0, len(self.embd), self.params.n_batch): n_eval = self.params.n_batch _arr = (llama_cpp.llama_token * n_eval)(*self.embd[i:i + n_eval]) if llama_cpp.llama_eval(self.ctx, _arr, n_eval, self.n_past, self.params.n_threads) != 0: @@ -303,265 +386,358 @@ def generate(self): self.n_past += n_eval""" - if (llama_cpp.llama_eval( - self.ctx, (llama_cpp.llama_token * len(self.embd))(*self.embd), len(self.embd), self.n_past, self.params.n_threads - ) != 0): - raise Exception("Failed to llama_eval!") - - if len(self.embd) > 0 and len(self.params.path_session) > 0: - self.session_tokens.extend(self.embd) - self.n_session_consumed = len(self.session_tokens) - - self.n_past += len(self.embd) - self.embd = [] - if len(self.embd_inp) <= self.input_consumed: #&& !is_interacting - # out of user input, sample next token - top_k = llama_cpp.llama_n_vocab(self.ctx) if self.params.top_k <= 0 else self.params.top_k - repeat_last_n = self.n_ctx if self.params.repeat_last_n < 0 else self.params.repeat_last_n - - # optionally save the session on first sample (for faster prompt loading next time) - if len(self.params.path_session) > 0 and self.need_to_save_session: - self.need_to_save_session = False - llama_cpp.llama_save_session_file( - self.ctx, - self.params.path_session.encode("utf8"), - (llama_cpp.llama_token * len(self.session_tokens))(*self.session_tokens), - len(self.session_tokens) - ) - - id = 0 - - logits = llama_cpp.llama_get_logits(self.ctx) - n_vocab = llama_cpp.llama_n_vocab(self.ctx) - - # Apply params.logit_bias map - for key, value in self.params.logit_bias.items(): - logits[key] += value - - _arr = (llama_cpp.llama_token_data * n_vocab)(*[ - llama_cpp.llama_token_data(token_id, logits[token_id], 0.0) - for token_id in range(n_vocab) - ]) - candidates_p = llama_cpp.ctypes.pointer(llama_cpp.llama_token_data_array(_arr, len(_arr), False)) - - # Apply penalties - nl_logit = logits[llama_cpp.llama_token_nl(self.ctx)] - last_n_repeat = min(len(self.last_n_tokens), repeat_last_n, self.n_ctx) - - _arr = (llama_cpp.llama_token * last_n_repeat)(*self.last_n_tokens[len(self.last_n_tokens) - last_n_repeat:]) - llama_cpp.llama_sample_repetition_penalty(self.ctx, candidates_p, - _arr, - last_n_repeat, llama_cpp.c_float(self.params.repeat_penalty)) - llama_cpp.llama_sample_frequency_and_presence_penalties(self.ctx, candidates_p, - _arr, - last_n_repeat, llama_cpp.c_float(self.params.frequency_penalty), llama_cpp.c_float(self.params.presence_penalty)) - - if not self.params.penalize_nl: - logits[llama_cpp.llama_token_nl()] = nl_logit - - if self.params.temp <= 0: - # Greedy sampling - id = llama_cpp.llama_sample_token_greedy(self.ctx, candidates_p) - else: - if self.params.mirostat == 1: - mirostat_mu = 2.0 * self.params.mirostat_tau - mirostat_m = 100 - llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp)) - id = llama_cpp.llama_sample_token_mirostat(self.ctx, candidates_p, llama_cpp.c_float(self.params.mirostat_tau), llama_cpp.c_float(self.params.mirostat_eta), llama_cpp.c_int(mirostat_m), llama_cpp.c_float(mirostat_mu)) - elif self.params.mirostat == 2: - mirostat_mu = 2.0 * self.params.mirostat_tau - llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp)) - id = llama_cpp.llama_sample_token_mirostat_v2(self.ctx, candidates_p, llama_cpp.c_float(self.params.mirostat_tau), llama_cpp.c_float(self.params.mirostat_eta), llama_cpp.c_float(mirostat_mu)) - else: - # Temperature sampling - llama_cpp.llama_sample_top_k(self.ctx, candidates_p, top_k, min_keep=llama_cpp.c_size_t(1)) - llama_cpp.llama_sample_tail_free(self.ctx, candidates_p, llama_cpp.c_float(self.params.tfs_z), min_keep=llama_cpp.c_size_t(1)) - llama_cpp.llama_sample_typical(self.ctx, candidates_p, llama_cpp.c_float(self.params.typical_p), min_keep=llama_cpp.c_size_t(1)) - llama_cpp.llama_sample_top_p(self.ctx, candidates_p, llama_cpp.c_float(self.params.top_p), min_keep=llama_cpp.c_size_t(1)) - llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp)) - id = llama_cpp.llama_sample_token(self.ctx, candidates_p) - # print("`{}`".format(candidates_p.size)) - - self.last_n_tokens.pop(0) - self.last_n_tokens.append(id) - - # replace end of text token with newline token when in interactive mode - if (id == llama_cpp.llama_token_eos(self.ctx) and self.params.interactive and not self.params.instruct): - id = self.llama_token_newline[0] - self.embd.append(id) - if (self.use_antiprompt()): - # tokenize and inject first reverse prompt - self.embd_inp += self.first_antiprompt[0] - for id in self.first_antiprompt[0]: - self.embd.append(id) - else: - # add it to the context - self.embd.append(id) - - # echo this to console - self.output_echo = True - - # decrement remaining sampling budget - self.remaining_tokens -= 1 - else: - # output to console if input echo is on - self.output_echo = self.params.input_echo - - # some user input remains from prompt or interaction, forward it to processing - while len(self.embd_inp) > self.input_consumed: - self.embd.append(self.embd_inp[self.input_consumed]) - self.last_n_tokens.pop(0) - self.last_n_tokens.append(self.embd_inp[self.input_consumed]) - self.input_consumed += 1 - if len(self.embd) >= self.params.n_batch: - break - - # display tokens - if self.output_echo: - for id in self.embd: - if self.antiecho != None: - for r in self.antiecho(id): - yield r - else: - yield id - - # reset color to default if we there is no pending user input - if (self.params.input_echo and len(self.embd_inp) == self.input_consumed): - self.set_color(util.CONSOLE_COLOR_DEFAULT) - - if (self.params.interactive and len(self.embd_inp) <= self.input_consumed): - # if antiprompt is present, stop - if (self.use_antiprompt()): - if True in [ - i == self.last_n_tokens[-len(i):] - for i in self.first_antiprompt - ]: - break - - # if we are using instruction mode, and we have processed the initial prompt - if (self.params.interactive_start): - break - - # end of text token - if len(self.embd) > 0 and self.embd[-1] == llama_cpp.llama_token_eos(self.ctx): - if (not self.params.instruct): - for i in self.llama_token_eot: - yield i - break - - # respect n_predict even if antiprompt is present - if (self.params.interactive and self.remaining_tokens <= 0 and self.params.n_predict != -1): - # If we arent in instruction mode, fix the current generation by appending the antiprompt. - # Makes it so if chat ends prematurely you dont append the AI's text etc. - if not self.params.instruct: - self.embd_inp += self.first_antiprompt[0] - self.n_remain = self.params.n_predict - break - - self.params.interactive_start = False - - def __enter__(self): - return self - - def __exit__(self, type, value, tb): - self.exit() - - def exit(self): - llama_cpp.llama_free(self.ctx) - self.set_color(util.CONSOLE_COLOR_DEFAULT) - - def token_to_str(self, token_id: int) -> bytes: - size = 32 - buffer = (ctypes.c_char * size)() - n = llama_cpp.llama_token_to_piece_with_model( - self.model, llama_cpp.llama_token(token_id), buffer, size) - assert n <= size - return bytes(buffer[:n]) - - # return past text - def past(self): - for id in self.last_n_tokens[-self.n_past:]: - yield self.token_to_str(id).decode("utf8", errors="ignore") - - # write input - def input(self, prompt: str): - if (self.params.instruct and self.last_n_tokens[-len(self.inp_prefix):] != self.inp_prefix): - self.embd_inp += self.inp_prefix - self.embd_inp += self._tokenize(prompt) - if (self.params.instruct): - self.embd_inp += self.inp_suffix - - # write output - def output(self): - self.remaining_tokens = self.params.n_predict - for id in self.generate(): - cur_char = self.token_to_str(id) - - # Add remainder of missing bytes - if None in self.multibyte_fix: - self.multibyte_fix[self.multibyte_fix.index(None)] = cur_char - - # Return completed utf char - if len(self.multibyte_fix) > 0 and not None in self.multibyte_fix: - yield (b"".join(self.multibyte_fix)).decode("utf8") - self.multibyte_fix = [] - continue - - # Contains multi-byte UTF8 - for num, pattern in [(2, 192), (3, 224), (4, 240)]: - # Bitwise AND check - if pattern & int.from_bytes(cur_char, 'little') == pattern: - self.multibyte_fix = [cur_char] + ([None] * (num-1)) - - # Stop incomplete bytes from passing - if len(self.multibyte_fix) > 0: - continue - - yield cur_char.decode("utf8") - - # read user input - def read_input(self): - out = "" - while (t := input()).endswith("\\"): - out += t[:-1] + "\n" - return out + t + "\n" - - # interactive mode - def interact(self): - for i in self.output(): - print(i,end="",flush=True) - self.params.input_echo = False - - while self.params.interactive: - self.set_color(util.CONSOLE_COLOR_USER_INPUT) - if (self.params.instruct): - print('\n> ', end="") - self.input(self.read_input()) - else: - print(self.params.input_prefix, end="") - self.input(f"{self.params.input_prefix}{self.read_input()}{self.params.input_suffix}") - print(self.params.input_suffix,end="") - self.set_color(util.CONSOLE_COLOR_DEFAULT) - - try: - for i in self.output(): - print(i,end="",flush=True) - except KeyboardInterrupt: - self.set_color(util.CONSOLE_COLOR_DEFAULT) - if not self.params.instruct: - print(self.params.fix_prefix,end="") - self.input(self.params.fix_prefix) + if ( + llama_cpp.llama_eval( + self.ctx, + (llama_cpp.llama_token * len(self.embd))(*self.embd), + len(self.embd), + self.n_past, + ) + != 0 + ): + raise Exception("Failed to llama_eval!") + + if len(self.embd) > 0 and len(self.params.path_session) > 0: + self.session_tokens.extend(self.embd) + self.n_session_consumed = len(self.session_tokens) + + self.n_past += len(self.embd) + self.embd = [] + if len(self.embd_inp) <= self.input_consumed: # && !is_interacting + # out of user input, sample next token + top_k = ( + llama_cpp.llama_n_vocab(self.ctx) + if self.params.top_k <= 0 + else self.params.top_k + ) + repeat_last_n = ( + self.n_ctx + if self.params.repeat_last_n < 0 + else self.params.repeat_last_n + ) + + # optionally save the session on first sample (for faster prompt loading next time) + if len(self.params.path_session) > 0 and self.need_to_save_session: + self.need_to_save_session = False + llama_cpp.llama_save_session_file( + self.ctx, + self.params.path_session.encode("utf8"), + (llama_cpp.llama_token * len(self.session_tokens))( + *self.session_tokens + ), + len(self.session_tokens), + ) + + id = 0 + + logits = llama_cpp.llama_get_logits(self.ctx) + n_vocab = llama_cpp.llama_n_vocab(self.model) + + # Apply params.logit_bias map + for key, value in self.params.logit_bias.items(): + logits[key] += value + + _arr = (llama_cpp.llama_token_data * n_vocab)( + *[ + llama_cpp.llama_token_data(token_id, logits[token_id], 0.0) + for token_id in range(n_vocab) + ] + ) + candidates_p = llama_cpp.ctypes.pointer( + llama_cpp.llama_token_data_array(_arr, len(_arr), False) + ) + + # Apply penalties + nl_logit = logits[llama_cpp.llama_token_nl(self.ctx)] + last_n_repeat = min(len(self.last_n_tokens), repeat_last_n, self.n_ctx) + + _arr = (llama_cpp.llama_token * last_n_repeat)( + *self.last_n_tokens[len(self.last_n_tokens) - last_n_repeat :] + ) + llama_cpp.llama_sample_repetition_penalties( + ctx=self.ctx, + candidates=candidates_p, + last_tokens_data=_arr, + penalty_last_n=last_n_repeat, + penalty_repeat=llama_cpp.c_float(self.params.repeat_penalty), + penalty_freq=llama_cpp.c_float(self.params.frequency_penalty), + penalty_present=llama_cpp.c_float(self.params.presence_penalty), + ) + + # NOT PRESENT IN CURRENT VERSION ? + # llama_cpp.llama_sample_frequency_and_presence_penalti(self.ctx, candidates_p, + # _arr, + # last_n_repeat, llama_cpp.c_float(self.params.frequency_penalty), llama_cpp.c_float(self.params.presence_penalty)) + + if not self.params.penalize_nl: + logits[llama_cpp.llama_token_nl()] = nl_logit + + if self.params.temp <= 0: + # Greedy sampling + id = llama_cpp.llama_sample_token_greedy(self.ctx, candidates_p) + else: + if self.params.mirostat == 1: + mirostat_mu = 2.0 * self.params.mirostat_tau + mirostat_m = 100 + llama_cpp.llama_sample_temperature( + self.ctx, candidates_p, llama_cpp.c_float(self.params.temp) + ) + id = llama_cpp.llama_sample_token_mirostat( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.mirostat_tau), + llama_cpp.c_float(self.params.mirostat_eta), + llama_cpp.c_int(mirostat_m), + llama_cpp.c_float(mirostat_mu), + ) + elif self.params.mirostat == 2: + mirostat_mu = 2.0 * self.params.mirostat_tau + llama_cpp.llama_sample_temperature( + self.ctx, candidates_p, llama_cpp.c_float(self.params.temp) + ) + id = llama_cpp.llama_sample_token_mirostat_v2( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.mirostat_tau), + llama_cpp.c_float(self.params.mirostat_eta), + llama_cpp.c_float(mirostat_mu), + ) + else: + # Temperature sampling + llama_cpp.llama_sample_top_k( + self.ctx, + candidates_p, + top_k, + min_keep=llama_cpp.c_size_t(1), + ) + llama_cpp.llama_sample_tail_free( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.tfs_z), + min_keep=llama_cpp.c_size_t(1), + ) + llama_cpp.llama_sample_typical( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.typical_p), + min_keep=llama_cpp.c_size_t(1), + ) + llama_cpp.llama_sample_top_p( + self.ctx, + candidates_p, + llama_cpp.c_float(self.params.top_p), + min_keep=llama_cpp.c_size_t(1), + ) + llama_cpp.llama_sample_temperature( + self.ctx, candidates_p, llama_cpp.c_float(self.params.temp) + ) + id = llama_cpp.llama_sample_token(self.ctx, candidates_p) + # print("`{}`".format(candidates_p.size)) + + self.last_n_tokens.pop(0) + self.last_n_tokens.append(id) + + # replace end of text token with newline token when in interactive mode + if ( + id == llama_cpp.llama_token_eos(self.ctx) + and self.params.interactive + and not self.params.instruct + ): + id = self.llama_token_newline[0] + self.embd.append(id) + if self.use_antiprompt(): + # tokenize and inject first reverse prompt + self.embd_inp += self.first_antiprompt[0] + for id in self.first_antiprompt[0]: + self.embd.append(id) + else: + # add it to the context + self.embd.append(id) + + # echo this to console + self.output_echo = True + + # decrement remaining sampling budget + self.remaining_tokens -= 1 + else: + # output to console if input echo is on + self.output_echo = self.params.input_echo + + # some user input remains from prompt or interaction, forward it to processing + while len(self.embd_inp) > self.input_consumed: + self.embd.append(self.embd_inp[self.input_consumed]) + self.last_n_tokens.pop(0) + self.last_n_tokens.append(self.embd_inp[self.input_consumed]) + self.input_consumed += 1 + if len(self.embd) >= self.params.n_batch: + break + + # display tokens + if self.output_echo: + for id in self.embd: + if self.antiecho != None: + for r in self.antiecho(id): + yield r + else: + yield id + + # reset color to default if we there is no pending user input + if self.params.input_echo and len(self.embd_inp) == self.input_consumed: + self.set_color(util.CONSOLE_COLOR_DEFAULT) + + if self.params.interactive and len(self.embd_inp) <= self.input_consumed: + # if antiprompt is present, stop + if self.use_antiprompt(): + if True in [ + i == self.last_n_tokens[-len(i) :] + for i in self.first_antiprompt + ]: + break + + # if we are using instruction mode, and we have processed the initial prompt + if self.params.interactive_start: + break + + # end of text token + if len(self.embd) > 0 and self.embd[-1] == llama_cpp.llama_token_eos( + self.ctx + ): + if not self.params.instruct: + for i in self.llama_token_eot: + yield i + break + + # respect n_predict even if antiprompt is present + if ( + self.params.interactive + and self.remaining_tokens <= 0 + and self.params.n_predict != -1 + ): + # If we arent in instruction mode, fix the current generation by appending the antiprompt. + # Makes it so if chat ends prematurely you dont append the AI's text etc. + if not self.params.instruct: + self.embd_inp += self.first_antiprompt[0] + self.n_remain = self.params.n_predict + break + + self.params.interactive_start = False + + def __enter__(self): + return self + + def __exit__(self, type, value, tb): + self.exit() + + def exit(self): + llama_cpp.llama_free(self.ctx) + self.set_color(util.CONSOLE_COLOR_DEFAULT) + + def token_to_str(self, token_id: int) -> bytes: + size = 32 + buffer = (ctypes.c_char * size)() + n = llama_cpp.llama_token_to_piece( + self.model, llama_cpp.llama_token(token_id), buffer, size + ) + assert n <= size + return bytes(buffer[:n]) + + # return past text + def past(self): + for id in self.last_n_tokens[-self.n_past :]: + yield self.token_to_str(id).decode("utf8", errors="ignore") + + # write input + def input(self, prompt: str): + if ( + self.params.instruct + and self.last_n_tokens[-len(self.inp_prefix) :] != self.inp_prefix + ): + self.embd_inp += self.inp_prefix + self.embd_inp += self._tokenize(prompt) + if self.params.instruct: + self.embd_inp += self.inp_suffix + + # write output + def output(self): + self.remaining_tokens = self.params.n_predict + for id in self.generate(): + cur_char = self.token_to_str(id) + + # Add remainder of missing bytes + if None in self.multibyte_fix: + self.multibyte_fix[self.multibyte_fix.index(None)] = cur_char + + # Return completed utf char + if len(self.multibyte_fix) > 0 and not None in self.multibyte_fix: + yield (b"".join(self.multibyte_fix)).decode("utf8") + self.multibyte_fix = [] + continue + + # Contains multi-byte UTF8 + for num, pattern in [(2, 192), (3, 224), (4, 240)]: + # Bitwise AND check + if pattern & int.from_bytes(cur_char, "little") == pattern: + self.multibyte_fix = [cur_char] + ([None] * (num - 1)) + + # Stop incomplete bytes from passing + if len(self.multibyte_fix) > 0: + continue + + yield cur_char.decode("utf8") + + # read user input + def read_input(self): + out = "" + while (t := input()).endswith("\\"): + out += t[:-1] + "\n" + return out + t + "\n" + + # interactive mode + def interact(self): + for i in self.output(): + print(i, end="", flush=True) + self.params.input_echo = False + + # Using string instead of tokens to check for antiprompt, + # It is more reliable than tokens for interactive mode. + generated_str = "" + while self.params.interactive: + self.set_color(util.CONSOLE_COLOR_USER_INPUT) + if self.params.instruct: + print("\n> ", end="") + self.input(self.read_input()) + else: + print(self.params.input_prefix, end="") + self.input( + f"{self.params.input_prefix}{self.read_input()}{self.params.input_suffix}" + ) + print(self.params.input_suffix, end="") + self.set_color(util.CONSOLE_COLOR_DEFAULT) + + try: + for i in self.output(): + print(i, end="", flush=True) + generated_str += i + for ap in self.params.antiprompt: + if generated_str.endswith(ap): + raise KeyboardInterrupt + except KeyboardInterrupt: + self.set_color(util.CONSOLE_COLOR_DEFAULT) + if not self.params.instruct: + print(self.params.fix_prefix, end="") + self.input(self.params.fix_prefix) + if __name__ == "__main__": - from datetime import datetime + from datetime import datetime - USER_NAME="User" - AI_NAME="ChatLLaMa" + USER_NAME = "User" + AI_NAME = "ChatLLaMa" - time_now = datetime.now() - prompt = f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}. + time_now = datetime.now() + prompt = f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}. {AI_NAME} is helpful, kind, honest, friendly, good at writing and never fails to answer {USER_NAME}’s requests immediately and with details and precision. -There are no annotations like (30 seconds passed...) or (to himself), just what {USER_NAME} and {AI_NAME} say aloud to each other. +Transcript below contains only the recorded dialog between two, without any annotations like (30 seconds passed...) or (to himself), just what {USER_NAME} and {AI_NAME} say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. The transcript only includes text, it does not include markup like HTML and Markdown. @@ -575,8 +751,11 @@ def interact(self): {AI_NAME}: A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae. {USER_NAME}: Name a color. {AI_NAME}: Blue -{USER_NAME}:""" - params = gpt_params_parse() +{USER_NAME}: """ + + params = gpt_params_parse() + if params.prompt is None and params.file is None: + params.prompt = prompt - with LLaMAInteract(params) as m: - m.interact() + with LLaMAInteract(params) as m: + m.interact() diff --git a/examples/low_level_api/low_level_api_llama_cpp.py b/examples/low_level_api/low_level_api_llama_cpp.py index ef1b2c016..ba3545771 100644 --- a/examples/low_level_api/low_level_api_llama_cpp.py +++ b/examples/low_level_api/low_level_api_llama_cpp.py @@ -7,23 +7,20 @@ llama_cpp.llama_backend_init(numa=False) N_THREADS = multiprocessing.cpu_count() -MODEL_PATH = os.environ.get('MODEL', "../models/7B/ggml-model.bin") +MODEL_PATH = os.environ.get("MODEL", "../models/7B/ggml-model.bin") prompt = b"\n\n### Instruction:\nWhat is the capital of France?\n\n### Response:\n" lparams = llama_cpp.llama_model_default_params() cparams = llama_cpp.llama_context_default_params() -model = llama_cpp.llama_load_model_from_file(MODEL_PATH.encode('utf-8'), lparams) +model = llama_cpp.llama_load_model_from_file(MODEL_PATH.encode("utf-8"), lparams) ctx = llama_cpp.llama_new_context_with_model(model, cparams) # determine the required inference memory per token: tmp = [0, 1, 2, 3] llama_cpp.llama_eval( - ctx = ctx, - tokens=(llama_cpp.c_int * len(tmp))(*tmp), - n_tokens=len(tmp), - n_past=0 - )# Deprecated + ctx=ctx, tokens=(llama_cpp.c_int * len(tmp))(*tmp), n_tokens=len(tmp), n_past=0 +) # Deprecated n_past = 0 @@ -32,12 +29,12 @@ embd_inp = (llama_cpp.llama_token * (len(prompt) + 1))() n_of_tok = llama_cpp.llama_tokenize( model=model, - text=bytes(str(prompt),'utf-8'), - text_len=len(embd_inp), + text=bytes(str(prompt), "utf-8"), + text_len=len(embd_inp), tokens=embd_inp, n_max_tokens=len(embd_inp), add_bos=False, - special=False + special=False, ) embd_inp = embd_inp[:n_of_tok] @@ -63,11 +60,11 @@ while remaining_tokens > 0: if len(embd) > 0: llama_cpp.llama_eval( - ctx = ctx, + ctx=ctx, tokens=(llama_cpp.c_int * len(embd))(*embd), n_tokens=len(embd), - n_past=n_past - )# Deprecated + n_past=n_past, + ) # Deprecated n_past += len(embd) embd = [] @@ -75,20 +72,26 @@ logits = llama_cpp.llama_get_logits(ctx) n_vocab = llama_cpp.llama_n_vocab(model) - _arr = (llama_cpp.llama_token_data * n_vocab)(*[ - llama_cpp.llama_token_data(token_id, logits[token_id], 0.0) - for token_id in range(n_vocab) - ]) + _arr = (llama_cpp.llama_token_data * n_vocab)( + *[ + llama_cpp.llama_token_data(token_id, logits[token_id], 0.0) + for token_id in range(n_vocab) + ] + ) candidates_p = llama_cpp.ctypes.pointer( - llama_cpp.llama_token_data_array(_arr, len(_arr), False)) + llama_cpp.llama_token_data_array(_arr, len(_arr), False) + ) _arr = (llama_cpp.c_int * len(last_n_tokens_data))(*last_n_tokens_data) - llama_cpp.llama_sample_repetition_penalties(ctx, candidates_p, + llama_cpp.llama_sample_repetition_penalties( + ctx, + candidates_p, _arr, penalty_last_n=last_n_repeat, penalty_repeat=repeat_penalty, penalty_freq=frequency_penalty, - penalty_present=presence_penalty) + penalty_present=presence_penalty, + ) llama_cpp.llama_sample_top_k(ctx, candidates_p, k=40, min_keep=1) llama_cpp.llama_sample_top_p(ctx, candidates_p, p=0.8, min_keep=1) @@ -111,10 +114,11 @@ size = 32 buffer = (ctypes.c_char * size)() n = llama_cpp.llama_token_to_piece( - model, llama_cpp.llama_token(id), buffer, size) + model, llama_cpp.llama_token(id), buffer, size + ) assert n <= size print( - buffer[:n].decode('utf-8'), + buffer[:n].decode("utf-8"), end="", flush=True, ) diff --git a/examples/low_level_api/quantize.py b/examples/low_level_api/quantize.py index 8bd03f88a..057ac389e 100644 --- a/examples/low_level_api/quantize.py +++ b/examples/low_level_api/quantize.py @@ -4,14 +4,16 @@ def main(args): + fname_inp = args.fname_inp.encode("utf-8") + fname_out = args.fname_out.encode("utf-8") if not os.path.exists(fname_inp): raise RuntimeError(f"Input file does not exist ({fname_inp})") if os.path.exists(fname_out): raise RuntimeError(f"Output file already exists ({fname_out})") - fname_inp = args.fname_inp.encode("utf-8") - fname_out = args.fname_out.encode("utf-8") - itype = args.itype - return_code = llama_cpp.llama_model_quantize(fname_inp, fname_out, itype) + ftype = args.type + args = llama_cpp.llama_model_quantize_default_params() + args.ftype = ftype + return_code = llama_cpp.llama_model_quantize(fname_inp, fname_out, args) if return_code != 0: raise RuntimeError("Failed to quantize model") @@ -20,6 +22,10 @@ def main(args): parser = argparse.ArgumentParser() parser.add_argument("fname_inp", type=str, help="Path to input model") parser.add_argument("fname_out", type=str, help="Path to output model") - parser.add_argument("type", type=int, help="Type of quantization (2: q4_0, 3: q4_1)") + parser.add_argument( + "type", + type=int, + help="Type of quantization (2: q4_0, 3: q4_1), see llama_cpp.py for enum", + ) args = parser.parse_args() main(args) diff --git a/examples/low_level_api/util.py b/examples/low_level_api/util.py index 9d0ec2f70..ef8b1c1ee 100644 --- a/examples/low_level_api/util.py +++ b/examples/low_level_api/util.py @@ -1,4 +1,3 @@ - ANSI_COLOR_RESET = "\x1b[0m" ANSI_COLOR_YELLOW = "\x1b[33m" ANSI_BOLD = "\x1b[1m" @@ -8,88 +7,95 @@ CONSOLE_COLOR_PROMPT = ANSI_COLOR_YELLOW CONSOLE_COLOR_USER_INPUT = ANSI_BOLD + ANSI_COLOR_GREEN + # Iterative search # Actively searches and prevents a pattern from being returned class IterSearch: - def __init__(self, pattern): - self.pattern = list(pattern) - self.buffer = [] - - def __call__(self, char): - self.buffer += [char] + def __init__(self, pattern): + self.pattern = list(pattern) + self.buffer = [] - if (self.pattern[:len(self.buffer)] == self.buffer): - if (len(self.buffer) >= len(self.pattern)): - self.buffer.clear() - return [] + def __call__(self, char): + self.buffer += [char] - _tmp = self.buffer[:] - self.buffer.clear() - return _tmp + if self.pattern[: len(self.buffer)] == self.buffer: + if len(self.buffer) >= len(self.pattern): + self.buffer.clear() + return [] -class Circle: - def __init__(self, size, default=0): - self.list = [default] * size - self.maxsize = size - self.size = 0 - self.offset = 0 - - def append(self, elem): - if self.size < self.maxsize: - self.list[self.size] = elem - self.size += 1 - else: - self.list[self.offset] = elem - self.offset = (self.offset + 1) % self.maxsize - - def __getitem__(self, val): - if isinstance(val, int): - if 0 > val or val >= self.size: - raise IndexError('Index out of range') - return self.list[val] if self.size < self.maxsize else self.list[(self.offset + val) % self.maxsize] - elif isinstance(val, slice): - start, stop, step = val.start, val.stop, val.step - if step is None: - step = 1 - if start is None: - start = 0 - if stop is None: - stop = self.size - if start < 0: - start = self.size + start - if stop < 0: - stop = self.size + stop - - indices = range(start, stop, step) - return [self.list[(self.offset + i) % self.maxsize] for i in indices if i < self.size] - else: - raise TypeError('Invalid argument type') + _tmp = self.buffer[:] + self.buffer.clear() + return _tmp +class Circle: + def __init__(self, size, default=0): + self.list = [default] * size + self.maxsize = size + self.size = 0 + self.offset = 0 + + def append(self, elem): + if self.size < self.maxsize: + self.list[self.size] = elem + self.size += 1 + else: + self.list[self.offset] = elem + self.offset = (self.offset + 1) % self.maxsize + + def __getitem__(self, val): + if isinstance(val, int): + if 0 > val or val >= self.size: + raise IndexError("Index out of range") + return ( + self.list[val] + if self.size < self.maxsize + else self.list[(self.offset + val) % self.maxsize] + ) + elif isinstance(val, slice): + start, stop, step = val.start, val.stop, val.step + if step is None: + step = 1 + if start is None: + start = 0 + if stop is None: + stop = self.size + if start < 0: + start = self.size + start + if stop < 0: + stop = self.size + stop + + indices = range(start, stop, step) + return [ + self.list[(self.offset + i) % self.maxsize] + for i in indices + if i < self.size + ] + else: + raise TypeError("Invalid argument type") if __name__ == "__main__": - c = Circle(5) - - c.append(1) - print(c.list) - print(c[:]) - assert c[0] == 1 - assert c[:5] == [1] - - for i in range(2,5+1): - c.append(i) - print(c.list) - print(c[:]) - assert c[0] == 1 - assert c[:5] == [1,2,3,4,5] - - for i in range(5+1,9+1): - c.append(i) - print(c.list) - print(c[:]) - assert c[0] == 5 - assert c[:5] == [5,6,7,8,9] - #assert c[:-5] == [5,6,7,8,9] - assert c[:10] == [5,6,7,8,9] - + c = Circle(5) + + c.append(1) + print(c.list) + print(c[:]) + assert c[0] == 1 + assert c[:5] == [1] + + for i in range(2, 5 + 1): + c.append(i) + print(c.list) + print(c[:]) + assert c[0] == 1 + assert c[:5] == [1, 2, 3, 4, 5] + + for i in range(5 + 1, 9 + 1): + c.append(i) + print(c.list) + print(c[:]) + assert c[0] == 5 + assert c[:5] == [5, 6, 7, 8, 9] + # assert c[:-5] == [5,6,7,8,9] + assert c[:10] == [5, 6, 7, 8, 9] diff --git a/examples/notebooks/Batching.ipynb b/examples/notebooks/Batching.ipynb index 687316b32..be7fe9b52 100644 --- a/examples/notebooks/Batching.ipynb +++ b/examples/notebooks/Batching.ipynb @@ -6,6 +6,7 @@ "metadata": {}, "outputs": [], "source": [ + "import ctypes\n", "import llama_cpp" ] }, @@ -13,18 +14,7 @@ "cell_type": "code", "execution_count": 2, "metadata": {}, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "ggml_init_cublas: GGML_CUDA_FORCE_MMQ: no\n", - "ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes\n", - "ggml_init_cublas: found 1 CUDA devices:\n", - " Device 0: NVIDIA GeForce RTX 2060, compute capability 7.5\n" - ] - } - ], + "outputs": [], "source": [ "llama_cpp.llama_backend_init(numa=False)" ] @@ -38,361 +28,103 @@ "name": "stderr", "output_type": "stream", "text": [ - "llama_model_loader: loaded meta data with 16 key-value pairs and 291 tensors from ../../models/mistral-7b-v0.1-GGUF/ggml-model-Q4_K.gguf (version GGUF V2)\n", - "llama_model_loader: - tensor 0: token_embd.weight q4_K [ 4096, 32000, 1, 1 ]\n", - "llama_model_loader: - tensor 1: output_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 2: output.weight q6_K [ 4096, 32000, 1, 1 ]\n", - "llama_model_loader: - tensor 3: blk.0.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 4: blk.0.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 5: blk.0.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 6: blk.0.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 7: blk.0.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 8: blk.0.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 9: blk.0.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 10: blk.0.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 11: blk.0.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 12: blk.1.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 13: blk.1.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 14: blk.1.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 15: blk.1.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 16: blk.1.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 17: blk.1.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 18: blk.1.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 19: blk.1.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 20: blk.1.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 21: blk.2.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 22: blk.2.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 23: blk.2.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 24: blk.2.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 25: blk.2.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 26: blk.2.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 27: blk.2.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 28: blk.2.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 29: blk.2.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 30: blk.3.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 31: blk.3.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 32: blk.3.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 33: blk.3.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 34: blk.3.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 35: blk.3.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 36: blk.3.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 37: blk.3.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 38: blk.3.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 39: blk.4.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 40: blk.4.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 41: blk.4.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 42: blk.4.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 43: blk.4.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 44: blk.4.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 45: blk.4.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 46: blk.4.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 47: blk.4.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 48: blk.5.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 49: blk.5.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 50: blk.5.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 51: blk.5.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 52: blk.5.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 53: blk.5.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 54: blk.5.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 55: blk.5.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 56: blk.5.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 57: blk.6.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 58: blk.6.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 59: blk.6.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 60: blk.6.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 61: blk.6.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 62: blk.6.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 63: blk.6.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 64: blk.6.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 65: blk.6.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 66: blk.7.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 67: blk.7.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 68: blk.7.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 69: blk.7.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 70: blk.7.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 71: blk.7.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 72: blk.7.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 73: blk.7.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 74: blk.7.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 75: blk.8.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 76: blk.8.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 77: blk.8.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 78: blk.8.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 79: blk.8.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 80: blk.8.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 81: blk.8.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 82: blk.8.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 83: blk.8.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 84: blk.9.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 85: blk.9.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 86: blk.9.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 87: blk.9.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 88: blk.9.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 89: blk.9.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 90: blk.9.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 91: blk.9.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 92: blk.9.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 93: blk.10.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 94: blk.10.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 95: blk.10.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 96: blk.10.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 97: blk.10.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 98: blk.10.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 99: blk.10.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 100: blk.10.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 101: blk.10.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 102: blk.11.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 103: blk.11.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 104: blk.11.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 105: blk.11.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 106: blk.11.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 107: blk.11.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 108: blk.11.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 109: blk.11.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 110: blk.11.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 111: blk.12.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 112: blk.12.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 113: blk.12.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 114: blk.12.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 115: blk.12.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 116: blk.12.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 117: blk.12.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 118: blk.12.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 119: blk.12.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 120: blk.13.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 121: blk.13.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 122: blk.13.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 123: blk.13.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 124: blk.13.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 125: blk.13.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 126: blk.13.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 127: blk.13.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 128: blk.13.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 129: blk.14.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 130: blk.14.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 131: blk.14.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 132: blk.14.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 133: blk.14.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 134: blk.14.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 135: blk.14.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 136: blk.14.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 137: blk.14.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 138: blk.15.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 139: blk.15.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 140: blk.15.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 141: blk.15.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 142: blk.15.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 143: blk.15.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 144: blk.15.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 145: blk.15.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 146: blk.15.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 147: blk.16.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 148: blk.16.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 149: blk.16.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 150: blk.16.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 151: blk.16.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 152: blk.16.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 153: blk.16.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 154: blk.16.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 155: blk.16.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 156: blk.17.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 157: blk.17.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 158: blk.17.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 159: blk.17.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 160: blk.17.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 161: blk.17.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 162: blk.17.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 163: blk.17.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 164: blk.17.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 165: blk.18.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 166: blk.18.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 167: blk.18.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 168: blk.18.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 169: blk.18.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 170: blk.18.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 171: blk.18.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 172: blk.18.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 173: blk.18.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 174: blk.19.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 175: blk.19.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 176: blk.19.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 177: blk.19.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 178: blk.19.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 179: blk.19.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 180: blk.19.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 181: blk.19.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 182: blk.19.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 183: blk.20.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 184: blk.20.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 185: blk.20.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 186: blk.20.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 187: blk.20.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 188: blk.20.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 189: blk.20.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 190: blk.20.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 191: blk.20.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 192: blk.21.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 193: blk.21.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 194: blk.21.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 195: blk.21.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 196: blk.21.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 197: blk.21.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 198: blk.21.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 199: blk.21.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 200: blk.21.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 201: blk.22.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 202: blk.22.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 203: blk.22.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 204: blk.22.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 205: blk.22.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 206: blk.22.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 207: blk.22.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 208: blk.22.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 209: blk.22.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 210: blk.23.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 211: blk.23.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 212: blk.23.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 213: blk.23.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 214: blk.23.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 215: blk.23.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 216: blk.23.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 217: blk.23.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 218: blk.23.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 219: blk.24.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 220: blk.24.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 221: blk.24.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 222: blk.24.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 223: blk.24.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 224: blk.24.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 225: blk.24.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 226: blk.24.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 227: blk.24.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 228: blk.25.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 229: blk.25.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 230: blk.25.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 231: blk.25.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 232: blk.25.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 233: blk.25.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 234: blk.25.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 235: blk.25.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 236: blk.25.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 237: blk.26.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 238: blk.26.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 239: blk.26.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 240: blk.26.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 241: blk.26.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 242: blk.26.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 243: blk.26.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 244: blk.26.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 245: blk.26.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 246: blk.27.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 247: blk.27.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 248: blk.27.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 249: blk.27.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 250: blk.27.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 251: blk.27.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 252: blk.27.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 253: blk.27.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 254: blk.27.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 255: blk.28.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 256: blk.28.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 257: blk.28.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 258: blk.28.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 259: blk.28.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 260: blk.28.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 261: blk.28.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 262: blk.28.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 263: blk.28.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 264: blk.29.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 265: blk.29.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 266: blk.29.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 267: blk.29.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 268: blk.29.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 269: blk.29.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 270: blk.29.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 271: blk.29.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 272: blk.29.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 273: blk.30.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 274: blk.30.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 275: blk.30.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 276: blk.30.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 277: blk.30.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 278: blk.30.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 279: blk.30.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 280: blk.30.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 281: blk.30.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 282: blk.31.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 283: blk.31.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 284: blk.31.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", - "llama_model_loader: - tensor 285: blk.31.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 286: blk.31.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 287: blk.31.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", - "llama_model_loader: - tensor 288: blk.31.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", - "llama_model_loader: - tensor 289: blk.31.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - tensor 290: blk.31.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", - "llama_model_loader: - kv 0: general.architecture str \n", - "llama_model_loader: - kv 1: general.name str \n", - "llama_model_loader: - kv 2: llama.context_length u32 \n", - "llama_model_loader: - kv 3: llama.embedding_length u32 \n", - "llama_model_loader: - kv 4: llama.block_count u32 \n", - "llama_model_loader: - kv 5: llama.feed_forward_length u32 \n", - "llama_model_loader: - kv 6: llama.rope.dimension_count u32 \n", - "llama_model_loader: - kv 7: llama.attention.head_count u32 \n", - "llama_model_loader: - kv 8: llama.attention.head_count_kv u32 \n", - "llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 \n", - "llama_model_loader: - kv 10: general.file_type u32 \n", - "llama_model_loader: - kv 11: tokenizer.ggml.model str \n", - "llama_model_loader: - kv 12: tokenizer.ggml.tokens arr \n", - "llama_model_loader: - kv 13: tokenizer.ggml.scores arr \n", - "llama_model_loader: - kv 14: tokenizer.ggml.token_type arr \n", - "llama_model_loader: - kv 15: general.quantization_version u32 \n", + "llama_model_loader: loaded meta data with 20 key-value pairs and 291 tensors from /workspaces/llama-cpp-python/mistral-7b-v0.1.Q2_K.gguf (version GGUF V2)\n", + "llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", + "llama_model_loader: - kv 0: general.architecture str = llama\n", + "llama_model_loader: - kv 1: general.name str = mistralai_mistral-7b-v0.1\n", + "llama_model_loader: - kv 2: llama.context_length u32 = 32768\n", + "llama_model_loader: - kv 3: llama.embedding_length u32 = 4096\n", + "llama_model_loader: - kv 4: llama.block_count u32 = 32\n", + "llama_model_loader: - kv 5: llama.feed_forward_length u32 = 14336\n", + "llama_model_loader: - kv 6: llama.rope.dimension_count u32 = 128\n", + "llama_model_loader: - kv 7: llama.attention.head_count u32 = 32\n", + "llama_model_loader: - kv 8: llama.attention.head_count_kv u32 = 8\n", + "llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010\n", + "llama_model_loader: - kv 10: llama.rope.freq_base f32 = 10000.000000\n", + "llama_model_loader: - kv 11: general.file_type u32 = 10\n", + "llama_model_loader: - kv 12: tokenizer.ggml.model str = llama\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "llama_model_loader: - kv 13: tokenizer.ggml.tokens arr[str,32000] = [\"\", \"\", \"\", \"<0x00>\", \"<...\n", + "llama_model_loader: - kv 14: tokenizer.ggml.scores arr[f32,32000] = [0.000000, 0.000000, 0.000000, 0.0000...\n", + "llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,32000] = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...\n", + "llama_model_loader: - kv 16: tokenizer.ggml.bos_token_id u32 = 1\n", + "llama_model_loader: - kv 17: tokenizer.ggml.eos_token_id u32 = 2\n", + "llama_model_loader: - kv 18: tokenizer.ggml.unknown_token_id u32 = 0\n", + "llama_model_loader: - kv 19: general.quantization_version u32 = 2\n", "llama_model_loader: - type f32: 65 tensors\n", - "llama_model_loader: - type q4_K: 193 tensors\n", - "llama_model_loader: - type q6_K: 33 tensors\n", - "llm_load_vocab: special tokens definition check successful ( 259/32000 ).\n", + "llama_model_loader: - type q2_K: 65 tensors\n", + "llama_model_loader: - type q3_K: 160 tensors\n", + "llama_model_loader: - type q6_K: 1 tensors\n", + "llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect\n", + "llm_load_vocab: special tokens cache size = 3\n", + "llm_load_vocab: token to piece cache size = 0.1637 MB\n", "llm_load_print_meta: format = GGUF V2\n", "llm_load_print_meta: arch = llama\n", "llm_load_print_meta: vocab type = SPM\n", "llm_load_print_meta: n_vocab = 32000\n", "llm_load_print_meta: n_merges = 0\n", - "llm_load_print_meta: n_ctx_train = 4096\n", + "llm_load_print_meta: vocab_only = 0\n", + "llm_load_print_meta: n_ctx_train = 32768\n", "llm_load_print_meta: n_embd = 4096\n", + "llm_load_print_meta: n_layer = 32\n", "llm_load_print_meta: n_head = 32\n", "llm_load_print_meta: n_head_kv = 8\n", - "llm_load_print_meta: n_layer = 32\n", "llm_load_print_meta: n_rot = 128\n", + "llm_load_print_meta: n_swa = 0\n", + "llm_load_print_meta: n_embd_head_k = 128\n", + "llm_load_print_meta: n_embd_head_v = 128\n", "llm_load_print_meta: n_gqa = 4\n", + "llm_load_print_meta: n_embd_k_gqa = 1024\n", + "llm_load_print_meta: n_embd_v_gqa = 1024\n", "llm_load_print_meta: f_norm_eps = 0.0e+00\n", "llm_load_print_meta: f_norm_rms_eps = 1.0e-05\n", "llm_load_print_meta: f_clamp_kqv = 0.0e+00\n", "llm_load_print_meta: f_max_alibi_bias = 0.0e+00\n", + "llm_load_print_meta: f_logit_scale = 0.0e+00\n", "llm_load_print_meta: n_ff = 14336\n", + "llm_load_print_meta: n_expert = 0\n", + "llm_load_print_meta: n_expert_used = 0\n", + "llm_load_print_meta: causal attn = 1\n", + "llm_load_print_meta: pooling type = 0\n", + "llm_load_print_meta: rope type = 0\n", + "llm_load_print_meta: rope scaling = linear\n", "llm_load_print_meta: freq_base_train = 10000.0\n", "llm_load_print_meta: freq_scale_train = 1\n", + "llm_load_print_meta: n_ctx_orig_yarn = 32768\n", + "llm_load_print_meta: rope_finetuned = unknown\n", + "llm_load_print_meta: ssm_d_conv = 0\n", + "llm_load_print_meta: ssm_d_inner = 0\n", + "llm_load_print_meta: ssm_d_state = 0\n", + "llm_load_print_meta: ssm_dt_rank = 0\n", + "llm_load_print_meta: ssm_dt_b_c_rms = 0\n", "llm_load_print_meta: model type = 7B\n", - "llm_load_print_meta: model ftype = mostly Q4_K - Medium\n", + "llm_load_print_meta: model ftype = Q2_K - Medium\n", "llm_load_print_meta: model params = 7.24 B\n", - "llm_load_print_meta: model size = 4.07 GiB (4.83 BPW) \n", - "llm_load_print_meta: general.name = LLaMA v2\n", - "llm_load_print_meta: BOS token = 1 ''\n", - "llm_load_print_meta: EOS token = 2 ''\n", - "llm_load_print_meta: UNK token = 0 ''\n", - "llm_load_print_meta: LF token = 13 '<0x0A>'\n", - "llm_load_tensors: ggml ctx size = 0.10 MB\n", - "llm_load_tensors: using CUDA for GPU acceleration\n", - "llm_load_tensors: mem required = 70.41 MB\n", - "llm_load_tensors: offloading 32 repeating layers to GPU\n", - "llm_load_tensors: offloading non-repeating layers to GPU\n", - "llm_load_tensors: offloaded 35/35 layers to GPU\n", - "llm_load_tensors: VRAM used: 4095.05 MB\n", - ".................................................................................................\n" + "llm_load_print_meta: model size = 2.87 GiB (3.41 BPW) \n", + "llm_load_print_meta: general.name = mistralai_mistral-7b-v0.1\n", + "llm_load_print_meta: BOS token = 1 ''\n", + "llm_load_print_meta: EOS token = 2 ''\n", + "llm_load_print_meta: UNK token = 0 ''\n", + "llm_load_print_meta: LF token = 13 '<0x0A>'\n", + "llm_load_print_meta: EOG token = 2 ''\n", + "llm_load_print_meta: max token length = 48\n", + "llm_load_tensors: ggml ctx size = 0.14 MiB\n", + "llm_load_tensors: CPU buffer size = 2939.57 MiB\n", + "..................................................................................................\n" ] } ], "source": [ "params = llama_cpp.llama_model_default_params()\n", "params.n_gpu_layers = 35\n", - "model = llama_cpp.llama_load_model_from_file(b\"../../models/mistral-7b-v0.1-GGUF/ggml-model-Q4_K.gguf\", params=params) # Update this to whatever" + "model = llama_cpp.llama_load_model_from_file(\n", + " b\"/workspaces/llama-cpp-python/mistral-7b-v0.1.Q2_K.gguf\", params\n", + ") # Update this to whatever" ] }, { @@ -404,7 +136,7 @@ "name": "stdout", "output_type": "stream", "text": [ - "[1, 1014, 2936, 9060, 285, 1142]\n", + "[1, 415, 2936, 9060, 285, 1142]\n", "58\n" ] } @@ -416,7 +148,9 @@ "prompt = b\"The quick brown fox\"\n", "\n", "tokens = (llama_cpp.llama_token * n_ctx)()\n", - "tokens_len = llama_cpp.llama_tokenize(model, prompt, len(prompt), tokens, len(tokens), True, True)\n", + "tokens_len = llama_cpp.llama_tokenize(\n", + " model, prompt, len(prompt), tokens, len(tokens), True, True\n", + ")\n", "print(tokens[:tokens_len])\n", "\n", "n_kv_req = tokens_len + (n_len - tokens_len) * n_parallel\n", @@ -432,22 +166,22 @@ "name": "stderr", "output_type": "stream", "text": [ - "llama_new_context_with_model: n_ctx = 58\n", + "llama_new_context_with_model: n_ctx = 64\n", + "llama_new_context_with_model: n_batch = 32\n", + "llama_new_context_with_model: n_ubatch = 32\n", + "llama_new_context_with_model: flash_attn = 0\n", "llama_new_context_with_model: freq_base = 10000.0\n", "llama_new_context_with_model: freq_scale = 1\n", - "llama_kv_cache_init: offloading v cache to GPU\n", - "llama_kv_cache_init: offloading k cache to GPU\n", - "llama_kv_cache_init: VRAM kv self = 7.25 MB\n", - "llama_new_context_with_model: kv self size = 7.25 MB\n", - "llama_build_graph: non-view tensors processed: 740/740\n", - "llama_new_context_with_model: compute buffer total size = 10.63 MB\n", - "llama_new_context_with_model: VRAM scratch buffer: 4.51 MB\n", - "llama_new_context_with_model: total VRAM used: 4106.81 MB (model: 4095.05 MB, context: 11.76 MB)\n" + "llama_kv_cache_init: CPU KV buffer size = 8.00 MiB\n", + "llama_new_context_with_model: KV self size = 8.00 MiB, K (f16): 4.00 MiB, V (f16): 4.00 MiB\n", + "llama_new_context_with_model: CPU output buffer size = 0.12 MiB\n", + "llama_new_context_with_model: CPU compute buffer size = 5.01 MiB\n", + "llama_new_context_with_model: graph nodes = 1030\n", + "llama_new_context_with_model: graph splits = 1\n" ] } ], "source": [ - "\n", "ctx_params = llama_cpp.llama_context_default_params()\n", "ctx_params.seed = 1234\n", "ctx_params.n_ctx = n_kv_req\n", @@ -473,7 +207,7 @@ "metadata": {}, "outputs": [], "source": [ - "import ctypes\n", + "\n", "\n", "batch.n_tokens = tokens_len\n", "for i in range(tokens_len):\n", @@ -499,10 +233,35 @@ " llama_cpp.llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens)" ] }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] + }, { "cell_type": "code", "execution_count": 9, "metadata": {}, + "outputs": [], + "source": [ + "\n", + "# Initialize sampler chain with default parameters\n", + "sparams = llama_cpp.llama_sampler_chain_default_params()\n", + "sampler_chain = llama_cpp.llama_sampler_chain_init(sparams)\n", + "\n", + "# Add top_k, top_p, temperature, and final distribution-based sampler\n", + "llama_cpp.llama_sampler_chain_add(sampler_chain, llama_cpp.llama_sampler_init_top_k(40))\n", + "llama_cpp.llama_sampler_chain_add(sampler_chain, llama_cpp.llama_sampler_init_top_p(0.9, 1))\n", + "llama_cpp.llama_sampler_chain_add(sampler_chain, llama_cpp.llama_sampler_init_temp(0.4))\n", + "llama_cpp.llama_sampler_chain_add(sampler_chain, llama_cpp.llama_sampler_init_dist(1234)) # Final \"dist\" sampler" + ] + }, + { + "cell_type": "code", + "execution_count": 10, + "metadata": {}, "outputs": [ { "name": "stdout", @@ -511,61 +270,59 @@ "7\n", "[' j', ' jumped']\n", "8\n", - "[' jumps', ' jumped over']\n", + "[' j over', ' jumped over']\n", "9\n", - "[' jumps over', ' jumped over the']\n", + "[' j over the', ' jumped over the']\n", "10\n", - "[' jumps over the', ' jumped over the lazy']\n", + "[' j over the lazy', ' jumped over the lazy']\n", "11\n", - "[' jumps over the lazy', ' jumped over the lazy dog']\n", + "[' j over the lazy dog', ' jumped over the lazy dog']\n", "12\n", - "[' jumps over the lazy dog', ' jumped over the lazy dog.']\n", + "[' j over the lazy dog.', ' jumped over the lazy dog\\n']\n", "13\n", - "[' jumps over the lazy dog.', ' jumped over the lazy dog.\\n']\n", + "[' j over the lazy dog. También', ' jumped over the lazy dog\\nGroupLayout']\n", "14\n", - "[' jumps over the lazy dog.\\n', ' jumped over the lazy dog.\\n\\n']\n", + "[' j over the lazy dog. También:', ' jumped over the lazy dog\\nGroupLayouting']\n", "15\n", - "[' jumps over the lazy dog.\\n\\n', ' jumped over the lazy dog.\\n\\nThe']\n", + "[' j over the lazy dog. También: is', ' jumped over the lazy dog\\nGroupLayouting is']\n", "16\n", - "[' jumps over the lazy dog.\\n\\nI', ' jumped over the lazy dog.\\n\\nThe quick']\n", + "[' j over the lazy dog. También: is a', ' jumped over the lazy dog\\nGroupLayouting is a']\n", "17\n", - "[' jumps over the lazy dog.\\n\\nI’', ' jumped over the lazy dog.\\n\\nThe quick brown']\n", + "[' j over the lazy dog. También: is a technique', ' jumped over the lazy dog\\nGroupLayouting is a common']\n", "18\n", - "[' jumps over the lazy dog.\\n\\nI’m', ' jumped over the lazy dog.\\n\\nThe quick brown f']\n", + "[' j over the lazy dog. También: is a technique practice', ' jumped over the lazy dog\\nGroupLayouting is a common practice']\n", "19\n", - "[' jumps over the lazy dog.\\n\\nI’m not', ' jumped over the lazy dog.\\n\\nThe quick brown fox']\n", + "[' j over the lazy dog. También: is a technique practice in', ' jumped over the lazy dog\\nGroupLayouting is a common practice in']\n", "20\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped']\n", + "[' j over the lazy dog. También: is a technique practice in the', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the']\n", "21\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over']\n", + "[' j over the lazy dog. También: is a technique practice in the real', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media']\n", "22\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the']\n", + "[' j over the lazy dog. También: is a technique practice in the real-', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry']\n", "23\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy']\n", + "[' j over the lazy dog. También: is a technique practice in the real-.', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry.']\n", "24\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However']\n", "25\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We,', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However,']\n", "26\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there']\n", "27\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\n']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has']\n", "28\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been']\n", "29\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little']\n", "30\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in the', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick brown']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little research', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little emp']\n", "31\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in the English', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick brown f']\n", + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little researchirical', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little empirical']\n", "32\n", - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in the English language', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick brown fox']\n" + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little researchirical research', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little empirical research']\n" ] } ], "source": [ - "import ctypes\n", - "\n", "streams = [\"\"] * n_parallel\n", "i_batch = [batch.n_tokens - 1] * n_parallel\n", "\n", @@ -577,35 +334,18 @@ " for i in range(n_parallel):\n", " if i_batch[i] < 0:\n", " continue\n", - " \n", - " n_vocab = llama_cpp.llama_n_vocab(model)\n", - " logits = llama_cpp.llama_get_logits_ith(ctx, i_batch[i])\n", - "\n", - " candidates = (llama_cpp.llama_token_data * n_vocab)()\n", - "\n", - " for token_id in range(n_vocab):\n", - " candidates[token_id].id = token_id\n", - " candidates[token_id].logit = logits[token_id]\n", - " candidates[token_id].p = 0.0\n", "\n", - " candidates_p = llama_cpp.llama_token_data_array(candidates, len(candidates), False)\n", - "\n", - " top_k = 40\n", - " top_p = 0.9\n", - " temp = 0.4\n", - "\n", - " llama_cpp.llama_sample_top_k(ctx, ctypes.byref(candidates_p), top_k, 1)\n", - " llama_cpp.llama_sample_top_p(ctx, ctypes.byref(candidates_p), top_p, 1)\n", - " llama_cpp.llama_sample_temp (ctx, ctypes.byref(candidates_p), temp)\n", - " \n", - " new_token_id = llama_cpp.llama_sample_token(ctx, ctypes.byref(candidates_p))\n", + " # Sample the next token using the sampler chain\n", + " new_token_id = llama_cpp.llama_sampler_sample(sampler_chain, ctx, -1)\n", "\n", " if new_token_id == llama_cpp.llama_token_eos(ctx) or n_cur == n_len:\n", " i_batch[i] = -1\n", " continue\n", "\n", " buf = (ctypes.c_char * 32)()\n", - " outlen = llama_cpp.llama_token_to_piece(model, new_token_id, buf, len(buf))\n", + " \n", + " # Convert token ID to text\n", + " outlen = llama_cpp.llama_token_to_piece(model, new_token_id, buf, len(buf), 0, False)\n", " streams[i] += bytes(buf[:outlen]).decode(\"utf-8\")\n", "\n", " batch.token[batch.n_tokens] = new_token_id\n", @@ -617,7 +357,7 @@ " i_batch[i] = batch.n_tokens\n", " batch.n_tokens += 1\n", " n_decode += 1\n", - " \n", + "\n", " if batch.n_tokens == 0:\n", " break\n", "\n", @@ -627,19 +367,19 @@ " print(\"Error decoding\", flush=True)\n", " break\n", " print(n_cur)\n", - " print(streams)\n" + " print(streams)" ] }, { "cell_type": "code", - "execution_count": 10, + "execution_count": 11, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ - "[' jumps over the lazy dog.\\n\\nI’m not sure if that’s the most famous sentence in the English language', ' jumped over the lazy dog.\\n\\nThe quick brown fox jumped over the lazy dog.\\n\\nThe quick brown fox']\n" + "[' j over the lazy dog. También: is a technique practice in the real-. We, when is been little researchirical research', ' jumped over the lazy dog\\nGroupLayouting is a common practice in the media industry. However, there has been little empirical research']\n" ] } ], @@ -649,7 +389,7 @@ }, { "cell_type": "code", - "execution_count": 11, + "execution_count": 12, "metadata": {}, "outputs": [], "source": [ @@ -658,7 +398,7 @@ }, { "cell_type": "code", - "execution_count": 12, + "execution_count": 13, "metadata": {}, "outputs": [], "source": [ @@ -667,7 +407,7 @@ }, { "cell_type": "code", - "execution_count": 13, + "execution_count": 14, "metadata": {}, "outputs": [], "source": [ @@ -676,7 +416,7 @@ }, { "cell_type": "code", - "execution_count": 14, + "execution_count": 15, "metadata": {}, "outputs": [], "source": [ @@ -700,7 +440,7 @@ ], "metadata": { "kernelspec": { - "display_name": ".venv", + "display_name": "Python 3", "language": "python", "name": "python3" }, @@ -714,7 +454,7 @@ "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", - "version": "3.11.5+" + "version": "3.12.1" }, "orig_nbformat": 4 }, diff --git a/examples/notebooks/Clients.ipynb b/examples/notebooks/Clients.ipynb index caebbb67f..fab82673e 100644 --- a/examples/notebooks/Clients.ipynb +++ b/examples/notebooks/Clients.ipynb @@ -37,11 +37,11 @@ "source": [ "import openai\n", "\n", - "openai.api_key = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + "openai.api_key = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", "openai.api_base = \"http://100.64.159.73:8000/v1\"\n", "\n", "openai.Completion.create(\n", - " model=\"text-davinci-003\", # currently can be anything\n", + " model=\"text-davinci-003\", # currently can be anything\n", " prompt=\"The quick brown fox jumps\",\n", " max_tokens=5,\n", ")" @@ -66,7 +66,9 @@ "source": [ "import os\n", "\n", - "os.environ[\"OPENAI_API_KEY\"] = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + "os.environ[\"OPENAI_API_KEY\"] = (\n", + " \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + ")\n", "os.environ[\"OPENAI_API_BASE\"] = \"http://100.64.159.73:8000/v1\"\n", "\n", "from langchain.llms import OpenAI\n", diff --git a/examples/notebooks/Functions.ipynb b/examples/notebooks/Functions.ipynb index 7a7b899ec..1f4138165 100644 --- a/examples/notebooks/Functions.ipynb +++ b/examples/notebooks/Functions.ipynb @@ -9,7 +9,7 @@ "The OpenAI compatbile web server in `llama-cpp-python` supports function calling.\n", "\n", "Function calling allows API clients to specify a schema that gives the model a format it should respond in.\n", - "Function calling in `llama-cpp-python` works by combining models pretrained for function calling such as [`functionary`](https://huggingface.co/abetlen/functionary-7b-v1-GGUF) with constrained sampling to produce a response that is compatible with the schema.\n", + "Function calling in `llama-cpp-python` works by combining models pretrained for function calling such as [`functionary`](https://huggingface.co/meetkai) with constrained sampling to produce a response that is compatible with the schema.\n", "\n", "Note however that this improves but does not guarantee that the response will be compatible with the schema.\n", "\n", @@ -45,10 +45,11 @@ "\n", "\n", "client = openai.OpenAI(\n", - " api_key = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\", # can be anything\n", - " base_url = \"http://100.64.159.73:8000/v1\" # NOTE: Replace with IP address and port of your llama-cpp-python server\n", + " api_key=\"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\", # can be anything\n", + " base_url=\"http://100.64.159.73:8000/v1\", # NOTE: Replace with IP address and port of your llama-cpp-python server\n", ")\n", "\n", + "\n", "# Example dummy function hard coded to return the same weather\n", "# In production, this could be your backend API or an external API\n", "def get_current_weather(location, unit=\"fahrenheit\"):\n", @@ -56,15 +57,23 @@ " if \"tokyo\" in location.lower():\n", " return json.dumps({\"location\": \"Tokyo\", \"temperature\": \"10\", \"unit\": \"celsius\"})\n", " elif \"san francisco\" in location.lower():\n", - " return json.dumps({\"location\": \"San Francisco\", \"temperature\": \"72\", \"unit\": \"fahrenheit\"})\n", + " return json.dumps(\n", + " {\"location\": \"San Francisco\", \"temperature\": \"72\", \"unit\": \"fahrenheit\"}\n", + " )\n", " elif \"paris\" in location.lower():\n", " return json.dumps({\"location\": \"Paris\", \"temperature\": \"22\", \"unit\": \"celsius\"})\n", " else:\n", " return json.dumps({\"location\": location, \"temperature\": \"unknown\"})\n", "\n", + "\n", "def run_conversation():\n", " # Step 1: send the conversation and available functions to the model\n", - " messages = [{\"role\": \"user\", \"content\": \"What's the weather like in San Francisco, Tokyo, and Paris?\"}]\n", + " messages = [\n", + " {\n", + " \"role\": \"user\",\n", + " \"content\": \"What's the weather like in San Francisco, Tokyo, and Paris?\",\n", + " }\n", + " ]\n", " tools = [\n", " {\n", " \"type\": \"function\",\n", @@ -123,6 +132,8 @@ " messages=messages,\n", " ) # get a new response from the model where it can see the function response\n", " return second_response\n", + "\n", + "\n", "print(run_conversation())" ] }, @@ -169,16 +180,18 @@ "# Enables `response_model`\n", "client = instructor.patch(client=client)\n", "\n", + "\n", "class UserDetail(BaseModel):\n", " name: str\n", " age: int\n", "\n", + "\n", "user = client.chat.completions.create(\n", " model=\"gpt-3.5-turbo\",\n", " response_model=UserDetail,\n", " messages=[\n", " {\"role\": \"user\", \"content\": \"Extract Jason is 25 years old\"},\n", - " ]\n", + " ],\n", ")\n", "\n", "assert isinstance(user, UserDetail)\n", @@ -213,17 +226,22 @@ "source": [ "import enum\n", "\n", + "\n", "class Labels(str, enum.Enum):\n", " \"\"\"Enumeration for single-label text classification.\"\"\"\n", + "\n", " SPAM = \"spam\"\n", " NOT_SPAM = \"not_spam\"\n", "\n", + "\n", "class SinglePrediction(BaseModel):\n", " \"\"\"\n", " Class for a single class label prediction.\n", " \"\"\"\n", + "\n", " class_label: Labels\n", "\n", + "\n", "def classify(data: str) -> SinglePrediction:\n", " \"\"\"Perform single-label classification on the input text.\"\"\"\n", " return client.chat.completions.create(\n", @@ -237,6 +255,7 @@ " ],\n", " ) # type: ignore\n", "\n", + "\n", "prediction = classify(\"Hello there I'm a Nigerian prince and I want to give you money\")\n", "assert prediction.class_label == Labels.SPAM\n", "print(prediction)" @@ -265,19 +284,23 @@ "source": [ "from typing import List\n", "\n", + "\n", "# Define Enum class for multiple labels\n", "class MultiLabels(str, enum.Enum):\n", " TECH_ISSUE = \"tech_issue\"\n", " BILLING = \"billing\"\n", " GENERAL_QUERY = \"general_query\"\n", "\n", + "\n", "# Define the multi-class prediction model\n", "class MultiClassPrediction(BaseModel):\n", " \"\"\"\n", " Class for a multi-class label prediction.\n", " \"\"\"\n", + "\n", " class_labels: List[MultiLabels]\n", "\n", + "\n", "def multi_classify(data: str) -> MultiClassPrediction:\n", " \"\"\"Perform multi-label classification on the input text.\"\"\"\n", " return client.chat.completions.create(\n", @@ -291,6 +314,7 @@ " ],\n", " ) # type: ignore\n", "\n", + "\n", "# Test multi-label classification\n", "ticket = \"My account is locked and I can't access my billing info.\"\n", "prediction = multi_classify(ticket)\n", @@ -331,10 +355,12 @@ "question = \"What is the meaning of life?\"\n", "context = \"The according to the devil the meaning of live is to live a life of sin and debauchery.\"\n", "\n", + "\n", "class QuestionAnswer(BaseModel):\n", " question: str\n", " answer: str\n", "\n", + "\n", "qa: QuestionAnswer = client.chat.completions.create(\n", " model=\"gpt-3.5-turbo\",\n", " response_model=QuestionAnswer,\n", @@ -351,6 +377,7 @@ ")\n", "print(qa)\n", "\n", + "\n", "class QuestionAnswerNoEvil(BaseModel):\n", " question: str\n", " answer: Annotated[\n", @@ -360,6 +387,7 @@ " ),\n", " ]\n", "\n", + "\n", "try:\n", " qa: QuestionAnswerNoEvil = client.chat.completions.create(\n", " model=\"gpt-3.5-turbo\",\n", @@ -405,6 +433,7 @@ "\n", "from pydantic import Field, BaseModel, model_validator, FieldValidationInfo\n", "\n", + "\n", "class Fact(BaseModel):\n", " fact: str = Field(...)\n", " substring_quote: List[str] = Field(...)\n", @@ -424,6 +453,7 @@ " for match in re.finditer(re.escape(quote), context):\n", " yield match.span()\n", "\n", + "\n", "class QuestionAnswer(BaseModel):\n", " question: str = Field(...)\n", " answer: List[Fact] = Field(...)\n", @@ -440,13 +470,17 @@ " temperature=0.0,\n", " response_model=QuestionAnswer,\n", " messages=[\n", - " {\"role\": \"system\", \"content\": \"You are a world class algorithm to answer questions with correct and exact citations.\"},\n", + " {\n", + " \"role\": \"system\",\n", + " \"content\": \"You are a world class algorithm to answer questions with correct and exact citations.\",\n", + " },\n", " {\"role\": \"user\", \"content\": f\"{context}\"},\n", - " {\"role\": \"user\", \"content\": f\"Question: {question}\"}\n", + " {\"role\": \"user\", \"content\": f\"Question: {question}\"},\n", " ],\n", " validation_context={\"text_chunk\": context},\n", " )\n", "\n", + "\n", "question = \"What did the author do during college?\"\n", "context = \"\"\"\n", "My name is Jason Liu, and I grew up in Toronto Canada but I was born in China.\n", diff --git a/examples/notebooks/Guidance.ipynb b/examples/notebooks/Guidance.ipynb index 045856ea2..c52559853 100644 --- a/examples/notebooks/Guidance.ipynb +++ b/examples/notebooks/Guidance.ipynb @@ -28,7 +28,9 @@ "source": [ "import os\n", "\n", - "os.environ[\"OPENAI_API_KEY\"] = \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + "os.environ[\"OPENAI_API_KEY\"] = (\n", + " \"sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\" # can be anything\n", + ")\n", "os.environ[\"OPENAI_API_BASE\"] = \"http://100.64.159.73:8000/v1\"\n", "os.environ[\"OPENAI_API_HOST\"] = \"http://100.64.159.73:8000\"\n", "\n", @@ -38,21 +40,23 @@ "guidance.llm = guidance.llms.OpenAI(\"text-davinci-003\", caching=False)\n", "\n", "# define a guidance program that adapts a proverb\n", - "program = guidance(\"\"\"Tweak this proverb to apply to model instructions instead.\n", + "program = guidance(\n", + " \"\"\"Tweak this proverb to apply to model instructions instead.\n", "\n", "{{proverb}}\n", "- {{book}} {{chapter}}:{{verse}}\n", "\n", "UPDATED\n", "Where there is no guidance{{gen 'rewrite' stop=\"\\\\n-\"}}\n", - "- GPT {{gen 'chapter'}}:{{gen 'verse'}}\"\"\")\n", + "- GPT {{gen 'chapter'}}:{{gen 'verse'}}\"\"\"\n", + ")\n", "\n", "# execute the program on a specific proverb\n", "executed_program = program(\n", " proverb=\"Where there is no guidance, a people falls,\\nbut in an abundance of counselors there is safety.\",\n", " book=\"Proverbs\",\n", " chapter=11,\n", - " verse=14\n", + " verse=14,\n", ")" ] }, diff --git a/examples/notebooks/Multimodal.ipynb b/examples/notebooks/Multimodal.ipynb index f1b8e9d15..8448ac1f7 100644 --- a/examples/notebooks/Multimodal.ipynb +++ b/examples/notebooks/Multimodal.ipynb @@ -11,7 +11,7 @@ }, { "cell_type": "code", - "execution_count": 1, + "execution_count": 13, "metadata": {}, "outputs": [ { @@ -25,7 +25,7 @@ "source": [ "from openai import OpenAI\n", "\n", - "client = OpenAI(base_url=\"http://100.64.159.73:8000/v1\", api_key=\"sk-1234\")\n", + "client = OpenAI(base_url=\"http://localhost:8000/v1\", api_key=\"llama.cpp\")\n", "response = client.chat.completions.create(\n", " model=\"gpt-4-vision-preview\",\n", " messages=[\n", @@ -38,13 +38,20 @@ " \"url\": \"https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png\",\n", " },\n", " },\n", - " {\"type\": \"text\", \"text\": \"What does the image say. Format your response as a json object with a single 'text' key.\"},\n", + " {\n", + " \"type\": \"text\",\n", + " \"text\": \"What does the image say. Format your response as a json object with a single 'text' key.\",\n", + " },\n", " ],\n", " }\n", " ],\n", - " response_format={ \"type\": \"json_object\" }\n", + " response_format={\n", + " \"type\": \"json_object\",\n", + " \"schema\": {\"type\": \"object\", \"properties\": {\"text\": {\"type\": \"string\"}}},\n", + " },\n", ")\n", "import json\n", + "\n", "print(json.loads(response.choices[0].message.content))" ] }, diff --git a/examples/notebooks/OpenHermesFunctionCalling.ipynb b/examples/notebooks/OpenHermesFunctionCalling.ipynb new file mode 100644 index 000000000..13128be04 --- /dev/null +++ b/examples/notebooks/OpenHermesFunctionCalling.ipynb @@ -0,0 +1,933 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": 1, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n" + ] + } + ], + "source": [ + "import json\n", + "import inspect\n", + "from typing import get_type_hints\n", + "\n", + "\n", + "class Article:\n", + " pass\n", + "\n", + "\n", + "class Weather:\n", + " pass\n", + "\n", + "\n", + "class Directions:\n", + " pass\n", + "\n", + "\n", + "def calculate_mortgage_payment(\n", + " loan_amount: int, interest_rate: float, loan_term: int\n", + ") -> float:\n", + " \"\"\"Get the monthly mortgage payment given an interest rate percentage.\"\"\"\n", + "\n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "\n", + "def get_article_details(\n", + " title: str,\n", + " authors: list[str],\n", + " short_summary: str,\n", + " date_published: str,\n", + " tags: list[str],\n", + ") -> Article:\n", + " '''Get article details from unstructured article text.\n", + " date_published: formatted as \"MM/DD/YYYY\"'''\n", + "\n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "\n", + "def get_weather(zip_code: str) -> Weather:\n", + " \"\"\"Get the current weather given a zip code.\"\"\"\n", + "\n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "\n", + "def get_directions(start: str, destination: str) -> Directions:\n", + " \"\"\"Get directions from Google Directions API.\n", + " start: start address as a string including zipcode (if any)\n", + " destination: end address as a string including zipcode (if any)\"\"\"\n", + "\n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "\n", + "def get_type_name(t):\n", + " name = str(t)\n", + " if \"list\" in name or \"dict\" in name:\n", + " return name\n", + " else:\n", + " return t.__name__\n", + "\n", + "\n", + "def serialize_function_to_json(func):\n", + " signature = inspect.signature(func)\n", + " type_hints = get_type_hints(func)\n", + "\n", + " function_info = {\n", + " \"name\": func.__name__,\n", + " \"description\": func.__doc__,\n", + " \"parameters\": {\"type\": \"object\", \"properties\": {}},\n", + " \"returns\": type_hints.get(\"return\", \"void\").__name__,\n", + " }\n", + "\n", + " for name, _ in signature.parameters.items():\n", + " param_type = get_type_name(type_hints.get(name, type(None)))\n", + " function_info[\"parameters\"][\"properties\"][name] = {\"type\": param_type}\n", + "\n", + " return json.dumps(function_info, indent=2)\n", + "\n", + "\n", + "print(serialize_function_to_json(get_article_details))" + ] + }, + { + "cell_type": "code", + "execution_count": 2, + "metadata": {}, + "outputs": [], + "source": [ + "import xml.etree.ElementTree as ET\n", + "import re\n", + "\n", + "\n", + "def extract_function_calls(completion):\n", + " completion = completion.strip()\n", + " pattern = r\"((.*?))\"\n", + " match = re.search(pattern, completion, re.DOTALL)\n", + " if not match:\n", + " return None\n", + "\n", + " multiplefn = match.group(1)\n", + " root = ET.fromstring(multiplefn)\n", + " functions = root.findall(\"functioncall\")\n", + " return [json.loads(fn.text) for fn in functions]" + ] + }, + { + "cell_type": "code", + "execution_count": 12, + "metadata": {}, + "outputs": [], + "source": [ + "def generate_hermes_prompt(prompt, functions):\n", + " functions = \"\\n\\n\".join([serialize_function_to_json(fn) for fn in functions])\n", + " prompt = f\"\"\"<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{functions}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {{\"name\": \"function_name\", \"arguments\": {{\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}}}} \n", + " {{\"name\": \"function_name\", \"arguments\": {{\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}}}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "{prompt}<|im_end|>\n", + "<|im_start|>assistant\"\"\"\n", + " return prompt" + ] + }, + { + "cell_type": "code", + "execution_count": 13, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{\n", + " \"name\": \"get_weather\",\n", + " \"description\": \"Get the current weather given a zip code.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"zip_code\": {\n", + " \"type\": \"str\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Weather\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"calculate_mortgage_payment\",\n", + " \"description\": \"Get the monthly mortgage payment given an interest rate percentage.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"loan_amount\": {\n", + " \"type\": \"int\"\n", + " },\n", + " \"interest_rate\": {\n", + " \"type\": \"float\"\n", + " },\n", + " \"loan_term\": {\n", + " \"type\": \"int\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"float\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "What's the weather in 10001?<|im_end|>\n", + "<|im_start|>assistant\n", + "<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{\n", + " \"name\": \"get_weather\",\n", + " \"description\": \"Get the current weather given a zip code.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"zip_code\": {\n", + " \"type\": \"str\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Weather\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"calculate_mortgage_payment\",\n", + " \"description\": \"Get the monthly mortgage payment given an interest rate percentage.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"loan_amount\": {\n", + " \"type\": \"int\"\n", + " },\n", + " \"interest_rate\": {\n", + " \"type\": \"float\"\n", + " },\n", + " \"loan_term\": {\n", + " \"type\": \"int\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"float\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.<|im_end|>\n", + "<|im_start|>assistant\n", + "<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{\n", + " \"name\": \"get_weather\",\n", + " \"description\": \"Get the current weather given a zip code.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"zip_code\": {\n", + " \"type\": \"str\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Weather\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"calculate_mortgage_payment\",\n", + " \"description\": \"Get the monthly mortgage payment given an interest rate percentage.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"loan_amount\": {\n", + " \"type\": \"int\"\n", + " },\n", + " \"interest_rate\": {\n", + " \"type\": \"float\"\n", + " },\n", + " \"loan_term\": {\n", + " \"type\": \"int\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"float\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "What's the current exchange rate for USD to EUR?<|im_end|>\n", + "<|im_start|>assistant\n" + ] + } + ], + "source": [ + "prompts = [\n", + " \"What's the weather in 10001?\",\n", + " \"Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.\",\n", + " \"What's the current exchange rate for USD to EUR?\",\n", + "]\n", + "functions = [get_weather, calculate_mortgage_payment, get_article_details]\n", + "\n", + "for prompt in prompts:\n", + " print(generate_hermes_prompt(prompt, functions))" + ] + }, + { + "cell_type": "code", + "execution_count": 5, + "metadata": {}, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "ggml_init_cublas: GGML_CUDA_FORCE_MMQ: no\n", + "ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes\n", + "ggml_init_cublas: found 1 CUDA devices:\n", + " Device 0: NVIDIA GeForce RTX 2060, compute capability 7.5\n", + "llama_model_loader: loaded meta data with 20 key-value pairs and 291 tensors from ../../models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf (version GGUF V3 (latest))\n", + "llama_model_loader: - tensor 0: token_embd.weight q4_K [ 4096, 32002, 1, 1 ]\n", + "llama_model_loader: - tensor 1: blk.0.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 2: blk.0.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 3: blk.0.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 4: blk.0.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 5: blk.0.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 6: blk.0.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 7: blk.0.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 8: blk.0.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 9: blk.0.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 10: blk.1.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 11: blk.1.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 12: blk.1.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 13: blk.1.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 14: blk.1.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 15: blk.1.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 16: blk.1.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 17: blk.1.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 18: blk.1.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 19: blk.2.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 20: blk.2.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 21: blk.2.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 22: blk.2.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 23: blk.2.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 24: blk.2.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 25: blk.2.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 26: blk.2.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 27: blk.2.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 28: blk.3.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 29: blk.3.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 30: blk.3.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 31: blk.3.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 32: blk.3.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 33: blk.3.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 34: blk.3.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 35: blk.3.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 36: blk.3.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 37: blk.4.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 38: blk.4.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 39: blk.4.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 40: blk.4.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 41: blk.4.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 42: blk.4.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 43: blk.4.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 44: blk.4.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 45: blk.4.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 46: blk.5.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 47: blk.5.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 48: blk.5.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 49: blk.5.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 50: blk.5.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 51: blk.5.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 52: blk.5.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 53: blk.5.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 54: blk.5.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 55: blk.6.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 56: blk.6.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 57: blk.6.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 58: blk.6.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 59: blk.6.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 60: blk.6.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 61: blk.6.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 62: blk.6.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 63: blk.6.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 64: blk.7.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 65: blk.7.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 66: blk.7.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 67: blk.7.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 68: blk.7.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 69: blk.7.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 70: blk.7.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 71: blk.7.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 72: blk.7.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 73: blk.8.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 74: blk.8.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 75: blk.8.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 76: blk.8.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 77: blk.8.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 78: blk.8.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 79: blk.8.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 80: blk.8.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 81: blk.8.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 82: blk.9.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 83: blk.9.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 84: blk.9.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 85: blk.9.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 86: blk.9.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 87: blk.9.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 88: blk.9.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 89: blk.9.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 90: blk.9.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 91: blk.10.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 92: blk.10.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 93: blk.10.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 94: blk.10.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 95: blk.10.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 96: blk.10.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 97: blk.10.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 98: blk.10.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 99: blk.10.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 100: blk.11.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 101: blk.11.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 102: blk.11.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 103: blk.11.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 104: blk.11.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 105: blk.11.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 106: blk.11.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 107: blk.11.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 108: blk.11.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 109: blk.12.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 110: blk.12.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 111: blk.12.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 112: blk.12.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 113: blk.12.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 114: blk.12.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 115: blk.12.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 116: blk.12.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 117: blk.12.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 118: blk.13.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 119: blk.13.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 120: blk.13.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 121: blk.13.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 122: blk.13.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 123: blk.13.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 124: blk.13.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 125: blk.13.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 126: blk.13.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 127: blk.14.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 128: blk.14.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 129: blk.14.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 130: blk.14.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 131: blk.14.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 132: blk.14.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 133: blk.14.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 134: blk.14.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 135: blk.14.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 136: blk.15.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 137: blk.15.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 138: blk.15.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 139: blk.15.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 140: blk.15.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 141: blk.15.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 142: blk.15.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 143: blk.15.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 144: blk.15.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 145: blk.16.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 146: blk.16.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 147: blk.16.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 148: blk.16.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 149: blk.16.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 150: blk.16.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 151: blk.16.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 152: blk.16.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 153: blk.16.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 154: blk.17.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 155: blk.17.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 156: blk.17.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 157: blk.17.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 158: blk.17.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 159: blk.17.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 160: blk.17.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 161: blk.17.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 162: blk.17.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 163: blk.18.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 164: blk.18.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 165: blk.18.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 166: blk.18.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 167: blk.18.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 168: blk.18.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 169: blk.18.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 170: blk.18.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 171: blk.18.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 172: blk.19.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 173: blk.19.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 174: blk.19.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 175: blk.19.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 176: blk.19.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 177: blk.19.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 178: blk.19.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 179: blk.19.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 180: blk.19.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 181: blk.20.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 182: blk.20.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 183: blk.20.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 184: blk.20.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 185: blk.20.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 186: blk.20.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 187: blk.20.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 188: blk.20.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 189: blk.20.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 190: blk.21.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 191: blk.21.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 192: blk.21.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 193: blk.21.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 194: blk.21.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 195: blk.21.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 196: blk.21.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 197: blk.21.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 198: blk.21.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 199: blk.22.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 200: blk.22.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 201: blk.22.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 202: blk.22.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 203: blk.22.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 204: blk.22.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 205: blk.22.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 206: blk.22.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 207: blk.22.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 208: blk.23.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 209: blk.23.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 210: blk.23.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 211: blk.23.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 212: blk.23.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 213: blk.23.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 214: blk.23.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 215: blk.23.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 216: blk.23.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 217: blk.24.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 218: blk.24.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 219: blk.24.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 220: blk.24.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 221: blk.24.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 222: blk.24.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 223: blk.24.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 224: blk.24.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 225: blk.24.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 226: blk.25.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 227: blk.25.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 228: blk.25.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 229: blk.25.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 230: blk.25.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 231: blk.25.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 232: blk.25.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 233: blk.25.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 234: blk.25.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 235: blk.26.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 236: blk.26.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 237: blk.26.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 238: blk.26.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 239: blk.26.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 240: blk.26.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 241: blk.26.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 242: blk.26.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 243: blk.26.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 244: blk.27.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 245: blk.27.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 246: blk.27.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 247: blk.27.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 248: blk.27.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 249: blk.27.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 250: blk.27.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 251: blk.27.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 252: blk.27.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 253: blk.28.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 254: blk.28.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 255: blk.28.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 256: blk.28.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 257: blk.28.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 258: blk.28.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 259: blk.28.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 260: blk.28.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 261: blk.28.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 262: blk.29.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 263: blk.29.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 264: blk.29.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 265: blk.29.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 266: blk.29.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 267: blk.29.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 268: blk.29.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 269: blk.29.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 270: blk.29.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 271: blk.30.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 272: blk.30.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 273: blk.30.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 274: blk.30.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 275: blk.30.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 276: blk.30.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 277: blk.30.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 278: blk.30.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 279: blk.30.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 280: blk.31.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 281: blk.31.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 282: blk.31.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 283: blk.31.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 284: blk.31.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 285: blk.31.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 286: blk.31.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 287: blk.31.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 288: blk.31.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 289: output_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 290: output.weight q6_K [ 4096, 32002, 1, 1 ]\n", + "llama_model_loader: - kv 0: general.architecture str = llama\n", + "llama_model_loader: - kv 1: general.name str = teknium_openhermes-2.5-mistral-7b\n", + "llama_model_loader: - kv 2: llama.context_length u32 = 32768\n", + "llama_model_loader: - kv 3: llama.embedding_length u32 = 4096\n", + "llama_model_loader: - kv 4: llama.block_count u32 = 32\n", + "llama_model_loader: - kv 5: llama.feed_forward_length u32 = 14336\n", + "llama_model_loader: - kv 6: llama.rope.dimension_count u32 = 128\n", + "llama_model_loader: - kv 7: llama.attention.head_count u32 = 32\n", + "llama_model_loader: - kv 8: llama.attention.head_count_kv u32 = 8\n", + "llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010\n", + "llama_model_loader: - kv 10: llama.rope.freq_base f32 = 10000.000000\n", + "llama_model_loader: - kv 11: general.file_type u32 = 15\n", + "llama_model_loader: - kv 12: tokenizer.ggml.model str = llama\n", + "llama_model_loader: - kv 13: tokenizer.ggml.tokens arr[str,32002] = [\"\", \"\", \"\", \"<0x00>\", \"<...\n", + "llama_model_loader: - kv 14: tokenizer.ggml.scores arr[f32,32002] = [0.000000, 0.000000, 0.000000, 0.0000...\n", + "llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,32002] = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...\n", + "llama_model_loader: - kv 16: tokenizer.ggml.bos_token_id u32 = 1\n", + "llama_model_loader: - kv 17: tokenizer.ggml.eos_token_id u32 = 32000\n", + "llama_model_loader: - kv 18: tokenizer.ggml.padding_token_id u32 = 0\n", + "llama_model_loader: - kv 19: general.quantization_version u32 = 2\n", + "llama_model_loader: - type f32: 65 tensors\n", + "llama_model_loader: - type q4_K: 193 tensors\n", + "llama_model_loader: - type q6_K: 33 tensors\n", + "llm_load_vocab: special tokens definition check successful ( 261/32002 ).\n", + "llm_load_print_meta: format = GGUF V3 (latest)\n", + "llm_load_print_meta: arch = llama\n", + "llm_load_print_meta: vocab type = SPM\n", + "llm_load_print_meta: n_vocab = 32002\n", + "llm_load_print_meta: n_merges = 0\n", + "llm_load_print_meta: n_ctx_train = 32768\n", + "llm_load_print_meta: n_embd = 4096\n", + "llm_load_print_meta: n_head = 32\n", + "llm_load_print_meta: n_head_kv = 8\n", + "llm_load_print_meta: n_layer = 32\n", + "llm_load_print_meta: n_rot = 128\n", + "llm_load_print_meta: n_gqa = 4\n", + "llm_load_print_meta: f_norm_eps = 0.0e+00\n", + "llm_load_print_meta: f_norm_rms_eps = 1.0e-05\n", + "llm_load_print_meta: f_clamp_kqv = 0.0e+00\n", + "llm_load_print_meta: f_max_alibi_bias = 0.0e+00\n", + "llm_load_print_meta: n_ff = 14336\n", + "llm_load_print_meta: rope scaling = linear\n", + "llm_load_print_meta: freq_base_train = 10000.0\n", + "llm_load_print_meta: freq_scale_train = 1\n", + "llm_load_print_meta: n_yarn_orig_ctx = 32768\n", + "llm_load_print_meta: rope_finetuned = unknown\n", + "llm_load_print_meta: model type = 7B\n", + "llm_load_print_meta: model ftype = mostly Q4_K - Medium\n", + "llm_load_print_meta: model params = 7.24 B\n", + "llm_load_print_meta: model size = 4.07 GiB (4.83 BPW) \n", + "llm_load_print_meta: general.name = teknium_openhermes-2.5-mistral-7b\n", + "llm_load_print_meta: BOS token = 1 ''\n", + "llm_load_print_meta: EOS token = 32000 '<|im_end|>'\n", + "llm_load_print_meta: UNK token = 0 ''\n", + "llm_load_print_meta: PAD token = 0 ''\n", + "llm_load_print_meta: LF token = 13 '<0x0A>'\n", + "llm_load_tensors: ggml ctx size = 0.11 MiB\n", + "llm_load_tensors: using CUDA for GPU acceleration\n", + "llm_load_tensors: mem required = 70.42 MiB\n", + "llm_load_tensors: offloading 32 repeating layers to GPU\n", + "llm_load_tensors: offloading non-repeating layers to GPU\n", + "llm_load_tensors: offloaded 35/35 layers to GPU\n", + "llm_load_tensors: VRAM used: 4095.06 MiB\n", + "...............................................................................................\n", + "llama_new_context_with_model: n_ctx = 2048\n", + "llama_new_context_with_model: freq_base = 10000.0\n", + "llama_new_context_with_model: freq_scale = 1\n", + "llama_kv_cache_init: offloading v cache to GPU\n", + "llama_kv_cache_init: offloading k cache to GPU\n", + "llama_kv_cache_init: VRAM kv self = 256.00 MiB\n", + "llama_new_context_with_model: kv self size = 256.00 MiB\n", + "llama_build_graph: non-view tensors processed: 740/740\n", + "llama_new_context_with_model: compute buffer total size = 159.07 MiB\n", + "llama_new_context_with_model: VRAM scratch buffer: 156.00 MiB\n", + "llama_new_context_with_model: total VRAM used: 4507.07 MiB (model: 4095.06 MiB, context: 412.00 MiB)\n" + ] + } + ], + "source": [ + "import llama_cpp\n", + "\n", + "llama = llama_cpp.Llama(\n", + " model_path=\"../../models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf\",\n", + " n_gpu_layers=-1,\n", + " n_ctx=2048,\n", + " verbose=False,\n", + ")" + ] + }, + { + "cell_type": "code", + "execution_count": 22, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "[{'name': 'get_weather', 'arguments': {'zip_code': '10001'}}]\n", + "====================================================================================================\n", + "[{'name': 'calculate_mortgage_payment', 'arguments': {'loan_amount': 200000, 'interest_rate': 0.04, 'loan_term': 30}}]\n", + "====================================================================================================\n", + "Unfortunately, I do not have a built-in function to check currency exchange rates. However, you can use third-party APIs or websites like Google Finance or XE to get this information.\n", + "====================================================================================================\n" + ] + } + ], + "source": [ + "prompts = [\n", + " \"What's the weather in 10001?\",\n", + " \"Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.\",\n", + " \"What's the current exchange rate for USD to EUR?\",\n", + "]\n", + "functions = [get_weather, calculate_mortgage_payment, get_article_details]\n", + "\n", + "for prompt in prompts:\n", + " prompt = generate_hermes_prompt(prompt, functions)\n", + " completion = llama.create_completion(prompt, max_tokens=-1)[\"choices\"][0][\"text\"]\n", + " function_calls = extract_function_calls(completion)\n", + " if function_calls:\n", + " print(function_calls)\n", + " else:\n", + " print(completion.strip())\n", + " print(\"=\" * 100)" + ] + }, + { + "cell_type": "code", + "execution_count": 23, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "get_weather\n", + "{'zip_code': '05751'}\n", + "====================================================================================================\n", + "get_weather\n", + "{'zip_code': '05751'}\n", + "get_weather\n", + "{'zip_code': '07030'}\n", + "calculate_mortgage_payment\n", + "{'loan_amount': 250000, 'interest_rate': 4.18, 'loan_term': 30}\n", + "====================================================================================================\n", + "I don't have a function to get exchange rates, but I can provide some resources where you can find this information. You can check websites like Google Finance, XE.com, or Yahoo Finance for up-to-date currency exchange rates.\n", + "====================================================================================================\n" + ] + } + ], + "source": [ + "prompts = [\n", + " \"What's the weather in 05751?\",\n", + " \"I'm planning a trip to Killington, Vermont (05751) from Hoboken, NJ (07030). Can you get me weather for both locations and directions?\",\n", + " \"What's the current exchange rate for USD to EUR?\",\n", + "]\n", + "\n", + "for prompt in prompts:\n", + " completion = llama.create_completion(\n", + " generate_hermes_prompt(prompt, functions), max_tokens=-1\n", + " )[\"choices\"][0][\"text\"]\n", + " function_calls = extract_function_calls(completion)\n", + "\n", + " if function_calls:\n", + " for function in function_calls:\n", + " print(function[\"name\"])\n", + " print(function[\"arguments\"])\n", + " else:\n", + " print(completion.strip())\n", + "\n", + " print(\"=\" * 100)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] + } + ], + "metadata": { + "kernelspec": { + "display_name": ".venv", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.11.5+" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/examples/notebooks/PerformanceTuning.ipynb b/examples/notebooks/PerformanceTuning.ipynb index 76e26fbd1..ba74e4a41 100644 --- a/examples/notebooks/PerformanceTuning.ipynb +++ b/examples/notebooks/PerformanceTuning.ipynb @@ -13,6 +13,7 @@ "import llama_cpp\n", "\n", "import numpy as np\n", + "\n", "np.int = int\n", "\n", "from skopt.space import Integer, Categorical\n", @@ -25,7 +26,7 @@ " Categorical([True, False], name=\"f16_kv\"),\n", " Categorical([True, False], name=\"use_mlock\"),\n", " Integer(1, multiprocessing.cpu_count(), name=\"n_threads\"),\n", - " Integer(1, 2048, name=\"n_batch\")\n", + " Integer(1, 2048, name=\"n_batch\"),\n", "]\n", "\n", "# TODO: Make this a random prompt to avoid any cache related inconsistencies\n", @@ -41,18 +42,25 @@ "\n", "from skopt.utils import use_named_args\n", "\n", + "\n", "@use_named_args(space)\n", "def objective(**params):\n", " f16_kv = params[\"f16_kv\"]\n", " use_mlock = params[\"use_mlock\"]\n", " n_threads = params[\"n_threads\"]\n", " n_batch = params[\"n_batch\"]\n", - " llm = llama_cpp.Llama(model_path=MODEL_PATH, f16_kv=f16_kv, use_mlock=use_mlock, n_threads=n_threads, n_batch=n_batch)\n", + " llm = llama_cpp.Llama(\n", + " model_path=MODEL_PATH,\n", + " f16_kv=f16_kv,\n", + " use_mlock=use_mlock,\n", + " n_threads=n_threads,\n", + " n_batch=n_batch,\n", + " )\n", "\n", " t1 = time.time()\n", " output = llm(\n", " PROMPT,\n", - " max_tokens=1, # Only optimize prompt processing\n", + " max_tokens=1, # Only optimize prompt processing\n", " stop=[\"###\", \"\\n\"],\n", " echo=True,\n", " )\n", @@ -5240,10 +5248,7 @@ "source": [ "from skopt import gp_minimize\n", "\n", - "res = gp_minimize(\n", - " objective,\n", - " space\n", - ")" + "res = gp_minimize(objective, space)" ] }, { diff --git a/examples/ray/README.md b/examples/ray/README.md new file mode 100644 index 000000000..6e338ba17 --- /dev/null +++ b/examples/ray/README.md @@ -0,0 +1,19 @@ +This is an example of doing LLM inference with [Ray](https://docs.ray.io/en/latest/index.html) and [Ray Serve](https://docs.ray.io/en/latest/serve/index.html). + +First, install the requirements: + +```bash +$ pip install -r requirements.txt +``` + +Deploy a GGUF model to Ray Serve with the following command: + +```bash +$ serve run llm:llm_builder model_path='../models/mistral-7b-instruct-v0.2.Q4_K_M.gguf' +``` + +This will start an API endpoint at `http://localhost:8000/`. You can query the model like this: + +```bash +$ curl -k -d '{"prompt": "tell me a joke", "max_tokens": 128}' -X POST http://localhost:8000 +``` diff --git a/examples/ray/llm.py b/examples/ray/llm.py new file mode 100755 index 000000000..2325dd303 --- /dev/null +++ b/examples/ray/llm.py @@ -0,0 +1,21 @@ +from starlette.requests import Request +from typing import Dict +from ray import serve +from ray.serve import Application +from llama_cpp import Llama + + +@serve.deployment +class LlamaDeployment: + def __init__(self, model_path: str): + self._llm = Llama(model_path=model_path) + + async def __call__(self, http_request: Request) -> Dict: + input_json = await http_request.json() + prompt = input_json["prompt"] + max_tokens = input_json.get("max_tokens", 64) + return self._llm(prompt, max_tokens=max_tokens) + + +def llm_builder(args: Dict[str, str]) -> Application: + return LlamaDeployment.bind(args["model_path"]) diff --git a/examples/ray/requirements.txt b/examples/ray/requirements.txt new file mode 100644 index 000000000..a409fb882 --- /dev/null +++ b/examples/ray/requirements.txt @@ -0,0 +1,3 @@ +ray[serve] +--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu +llama-cpp-python diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index 00031ba49..2c9c527cd 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.2.24" \ No newline at end of file +__version__ = "0.3.9" diff --git a/llama_cpp/_ctypes_extensions.py b/llama_cpp/_ctypes_extensions.py new file mode 100644 index 000000000..e88ed387d --- /dev/null +++ b/llama_cpp/_ctypes_extensions.py @@ -0,0 +1,131 @@ +from __future__ import annotations + +import sys +import os +import ctypes +import functools +import pathlib + +from typing import ( + Any, + Callable, + List, + Union, + Optional, + TYPE_CHECKING, + TypeVar, + Generic, +) +from typing_extensions import TypeAlias + + +# Load the library +def load_shared_library(lib_base_name: str, base_path: pathlib.Path): + """Platform independent shared library loader""" + # Searching for the library in the current directory under the name "libllama" (default name + # for llamacpp) and "llama" (default name for this repo) + lib_paths: List[pathlib.Path] = [] + # Determine the file extension based on the platform + if sys.platform.startswith("linux") or sys.platform.startswith("freebsd"): + lib_paths += [ + base_path / f"lib{lib_base_name}.so", + ] + elif sys.platform == "darwin": + lib_paths += [ + base_path / f"lib{lib_base_name}.so", + base_path / f"lib{lib_base_name}.dylib", + ] + elif sys.platform == "win32": + lib_paths += [ + base_path / f"{lib_base_name}.dll", + base_path / f"lib{lib_base_name}.dll", + ] + else: + raise RuntimeError("Unsupported platform") + + cdll_args = dict() # type: ignore + + # Add the library directory to the DLL search path on Windows (if needed) + if sys.platform == "win32": + os.add_dll_directory(str(base_path)) + os.environ["PATH"] = str(base_path) + os.pathsep + os.environ["PATH"] + + if sys.platform == "win32" and sys.version_info >= (3, 8): + os.add_dll_directory(str(base_path)) + if "CUDA_PATH" in os.environ: + os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin")) + os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib")) + if "HIP_PATH" in os.environ: + os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "bin")) + os.add_dll_directory(os.path.join(os.environ["HIP_PATH"], "lib")) + cdll_args["winmode"] = ctypes.RTLD_GLOBAL + + # Try to load the shared library, handling potential errors + for lib_path in lib_paths: + if lib_path.exists(): + try: + return ctypes.CDLL(str(lib_path), **cdll_args) # type: ignore + except Exception as e: + raise RuntimeError(f"Failed to load shared library '{lib_path}': {e}") + + raise FileNotFoundError( + f"Shared library with base name '{lib_base_name}' not found" + ) + + +# ctypes sane type hint helpers +# +# - Generic Pointer and Array types +# - PointerOrRef type with a type hinted byref function +# +# NOTE: Only use these for static type checking not for runtime checks +# no good will come of that + +if TYPE_CHECKING: + CtypesCData = TypeVar("CtypesCData", bound=ctypes._CData) # type: ignore + + CtypesArray: TypeAlias = ctypes.Array[CtypesCData] # type: ignore + + CtypesPointer: TypeAlias = ctypes._Pointer[CtypesCData] # type: ignore + + CtypesVoidPointer: TypeAlias = ctypes.c_void_p + + class CtypesRef(Generic[CtypesCData]): + pass + + CtypesPointerOrRef: TypeAlias = Union[ + CtypesPointer[CtypesCData], CtypesRef[CtypesCData] + ] + + CtypesFuncPointer: TypeAlias = ctypes._FuncPointer # type: ignore + +F = TypeVar("F", bound=Callable[..., Any]) + + +def ctypes_function_for_shared_library(lib: ctypes.CDLL): + """Decorator for defining ctypes functions with type hints""" + + def ctypes_function( + name: str, argtypes: List[Any], restype: Any, enabled: bool = True + ): + def decorator(f: F) -> F: + if enabled: + func = getattr(lib, name) + func.argtypes = argtypes + func.restype = restype + functools.wraps(f)(func) + return func + else: + return f + + return decorator + + return ctypes_function + + +def _byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCData]: + """Type-annotated version of ctypes.byref""" + ... + + +byref = _byref if TYPE_CHECKING else ctypes.byref diff --git a/llama_cpp/_ggml.py b/llama_cpp/_ggml.py new file mode 100644 index 000000000..5bee8a93b --- /dev/null +++ b/llama_cpp/_ggml.py @@ -0,0 +1,12 @@ +"""Internal module use at your own risk + +This module provides a minimal interface for working with ggml tensors from llama-cpp-python +""" +import os +import pathlib + +import llama_cpp._ctypes_extensions as ctypes_ext + +libggml_base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" +libggml = ctypes_ext.load_shared_library("ggml", libggml_base_path) + diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py new file mode 100644 index 000000000..343581dce --- /dev/null +++ b/llama_cpp/_internals.py @@ -0,0 +1,879 @@ +from __future__ import annotations + +import os +import ctypes + +from typing import ( + Dict, + List, + Tuple, + Optional, + Sequence, +) +from dataclasses import dataclass, field +from contextlib import ExitStack + +import numpy as np +import numpy.typing as npt + +from .llama_types import * +from .llama_grammar import LlamaGrammar +from ._utils import suppress_stdout_stderr + +import llama_cpp.llama_cpp as llama_cpp + + +# Python wrappers over llama.h structs + + +class LlamaModel: + """Intermediate Python wrapper for a llama.cpp llama_model. + NOTE: For stability it's recommended you use the Llama class instead.""" + + def __init__( + self, + *, + path_model: str, + params: llama_cpp.llama_model_params, + verbose: bool = True, + ): + self.path_model = path_model + self.params = params + self.verbose = verbose + self._exit_stack = ExitStack() + + model = None + + if not os.path.exists(path_model): + raise ValueError(f"Model path does not exist: {path_model}") + + with suppress_stdout_stderr(disable=verbose): + model = llama_cpp.llama_load_model_from_file( + self.path_model.encode("utf-8"), self.params + ) + + if model is None: + raise ValueError(f"Failed to load model from file: {path_model}") + + vocab = llama_cpp.llama_model_get_vocab(model) + + if vocab is None: + raise ValueError(f"Failed to get vocab from model: {path_model}") + + self.model = model + self.vocab = vocab + + def free_model(): + if self.model is None: + return + llama_cpp.llama_free_model(self.model) + self.model = None + + self._exit_stack.callback(free_model) + + def close(self): + self._exit_stack.close() + + def __del__(self): + self.close() + + def vocab_type(self) -> int: + return llama_cpp.llama_vocab_type(self.model) + + def n_vocab(self) -> int: + return llama_cpp.llama_n_vocab(self.vocab) + + def n_ctx_train(self) -> int: + return llama_cpp.llama_n_ctx_train(self.model) + + def n_embd(self) -> int: + return llama_cpp.llama_n_embd(self.model) + + def rope_freq_scale_train(self) -> float: + return llama_cpp.llama_model_rope_freq_scale_train(self.model) + + def desc(self) -> str: + buf = ctypes.create_string_buffer(1024) + llama_cpp.llama_model_desc(self.model, buf, 1024) + return buf.value.decode("utf-8") + + def size(self) -> int: + return llama_cpp.llama_model_size(self.model) + + def n_params(self) -> int: + return llama_cpp.llama_model_n_params(self.model) + + def get_tensor(self, name: str) -> ctypes.c_void_p: + raise NotImplementedError("get_tensor is not implemented in llama.cpp") + + # Vocab + + def token_get_text(self, token: int) -> str: + return llama_cpp.llama_token_get_text(self.vocab, token).decode("utf-8") + + def token_get_score(self, token: int) -> float: + return llama_cpp.llama_token_get_score(self.vocab, token) + + def token_get_attr(self, token: int) -> int: + return llama_cpp.llama_token_get_attr(self.vocab, token) + + # Special tokens + + def token_bos(self) -> int: + return llama_cpp.llama_token_bos(self.vocab) + + def token_eos(self) -> int: + return llama_cpp.llama_token_eos(self.vocab) + + def token_cls(self) -> int: + return llama_cpp.llama_token_cls(self.vocab) + + def token_sep(self) -> int: + return llama_cpp.llama_token_sep(self.vocab) + + def token_nl(self) -> int: + return llama_cpp.llama_token_nl(self.vocab) + + def token_prefix(self) -> int: + raise NotImplementedError("token_prefix is not implemented in llama.cpp") + + def token_middle(self) -> int: + raise NotImplementedError("token_middle is not implemented in llama.cpp") + + def token_suffix(self) -> int: + raise NotImplementedError("token_suffix is not implemented in llama.cpp") + + def token_eot(self) -> int: + return llama_cpp.llama_token_eot(self.vocab) + + def add_bos_token(self) -> bool: + return llama_cpp.llama_add_bos_token(self.vocab) + + def add_eos_token(self) -> bool: + return llama_cpp.llama_add_eos_token(self.vocab) + + # Tokenization + + def tokenize(self, text: bytes, add_bos: bool, special: bool): + n_ctx = self.n_ctx_train() + tokens = (llama_cpp.llama_token * n_ctx)() + n_tokens = llama_cpp.llama_tokenize( + self.vocab, text, len(text), tokens, n_ctx, add_bos, special + ) + if n_tokens < 0: + n_tokens = abs(n_tokens) + tokens = (llama_cpp.llama_token * n_tokens)() + n_tokens = llama_cpp.llama_tokenize( + self.vocab, text, len(text), tokens, n_tokens, add_bos, special + ) + if n_tokens < 0: + raise RuntimeError( + f'Failed to tokenize: text="{text}" n_tokens={n_tokens}' + ) + return list(tokens[:n_tokens]) + + def token_to_piece(self, token: int, special: bool = False) -> bytes: + buf = ctypes.create_string_buffer(32) + llama_cpp.llama_token_to_piece(self.vocab, token, buf, 32, 0, special) + return bytes(buf) + + def detokenize(self, tokens: List[int], special: bool = False) -> bytes: + output = b"" + size = 32 + buffer = (ctypes.c_char * size)() + for token in tokens: + n = llama_cpp.llama_token_to_piece( + self.vocab, llama_cpp.llama_token(token), buffer, size, 0, special + ) + assert n <= size + output += bytes(buffer[:n]) + # NOTE: Llama1 models automatically added a space at the start of the prompt + # this line removes a leading space if the first token is a beginning of sentence token + return ( + output[1:] + if len(tokens) > 0 and tokens[0] == self.token_bos() and output[0:1] == b" " + else output + ) + + # Extra + def metadata(self) -> Dict[str, str]: + metadata: Dict[str, str] = {} + buffer_size = 1024 + buffer = ctypes.create_string_buffer(buffer_size) + # zero the buffer + buffer.value = b"\0" * buffer_size + # iterate over model keys + for i in range(llama_cpp.llama_model_meta_count(self.model)): + nbytes = llama_cpp.llama_model_meta_key_by_index( + self.model, i, buffer, buffer_size + ) + if nbytes > buffer_size: + buffer_size = nbytes + 1 + buffer = ctypes.create_string_buffer(buffer_size) + nbytes = llama_cpp.llama_model_meta_key_by_index( + self.model, i, buffer, buffer_size + ) + key = buffer.value.decode("utf-8") + nbytes = llama_cpp.llama_model_meta_val_str_by_index( + self.model, i, buffer, buffer_size + ) + if nbytes > buffer_size: + buffer_size = nbytes + 1 + buffer = ctypes.create_string_buffer(buffer_size) + nbytes = llama_cpp.llama_model_meta_val_str_by_index( + self.model, i, buffer, buffer_size + ) + value = buffer.value.decode("utf-8") + metadata[key] = value + return metadata + + @staticmethod + def default_params(): + """Get the default llama_model_params.""" + return llama_cpp.llama_model_default_params() + + +class LlamaContext: + """Intermediate Python wrapper for a llama.cpp llama_context. + NOTE: For stability it's recommended you use the Llama class instead.""" + + def __init__( + self, + *, + model: LlamaModel, + params: llama_cpp.llama_context_params, + verbose: bool = True, + ): + self.model = model + self.params = params + self.verbose = verbose + self._exit_stack = ExitStack() + + ctx = llama_cpp.llama_new_context_with_model(self.model.model, self.params) + + if ctx is None: + raise ValueError("Failed to create llama_context") + + self.ctx = ctx + + def free_ctx(): + if self.ctx is None: + return + llama_cpp.llama_free(self.ctx) + self.ctx = None + + self._exit_stack.callback(free_ctx) + + def close(self): + self._exit_stack.close() + + def __del__(self): + self.close() + + def n_ctx(self) -> int: + return llama_cpp.llama_n_ctx(self.ctx) + + def pooling_type(self) -> int: + return llama_cpp.llama_pooling_type(self.ctx) + + def kv_cache_clear(self): + llama_cpp.llama_kv_cache_clear(self.ctx) + + def kv_cache_seq_rm(self, seq_id: int, p0: int, p1: int): + llama_cpp.llama_kv_cache_seq_rm(self.ctx, seq_id, p0, p1) + + def kv_cache_seq_cp(self, seq_id_src: int, seq_id_dst: int, p0: int, p1: int): + llama_cpp.llama_kv_cache_seq_cp(self.ctx, seq_id_src, seq_id_dst, p0, p1) + + def kv_cache_seq_keep(self, seq_id: int): + llama_cpp.llama_kv_cache_seq_keep(self.ctx, seq_id) + + def kv_cache_seq_shift(self, seq_id: int, p0: int, p1: int, shift: int): + llama_cpp.llama_kv_cache_seq_add(self.ctx, seq_id, p0, p1, shift) + + def get_state_size(self) -> int: + return llama_cpp.llama_get_state_size(self.ctx) + + # TODO: copy_state_data + + # TODO: set_state_data + + # TODO: llama_load_session_file + + # TODO: llama_save_session_file + + def decode(self, batch: LlamaBatch): + return_code = llama_cpp.llama_decode( + self.ctx, + batch.batch, + ) + if return_code != 0: + raise RuntimeError(f"llama_decode returned {return_code}") + + def set_n_threads(self, n_threads: int, n_threads_batch: int): + llama_cpp.llama_set_n_threads(self.ctx, n_threads, n_threads_batch) + + def get_logits(self): + return llama_cpp.llama_get_logits(self.ctx) + + def get_logits_ith(self, i: int): + return llama_cpp.llama_get_logits_ith(self.ctx, i) + + def get_embeddings(self): + return llama_cpp.llama_get_embeddings(self.ctx) + + # Sampling functions + + def set_rng_seed(self, seed: int): + # TODO: Fix + # llama_cpp.llama_set_rng_seed(self.ctx, seed) + raise NotImplementedError("set_rng_seed is not implemented in llama.cpp") + + def sample_repetition_penalties( + self, + candidates: "_LlamaTokenDataArray", + last_tokens_data: "llama_cpp.Array[llama_cpp.llama_token]", + penalty_last_n: int, + penalty_repeat: float, + penalty_freq: float, + penalty_present: float, + ): + # llama_cpp.llama_sample_repetition_penalties( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # last_tokens_data, + # penalty_last_n, + # penalty_repeat, + # penalty_freq, + # penalty_present, + # ) + raise NotImplementedError("sample_repetition_penalties is not implemented in llama.cpp") + + def sample_softmax(self, candidates: "_LlamaTokenDataArray"): + # llama_cpp.llama_sample_softmax( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # ) + raise NotImplementedError("sample_softmax is not implemented in llama.cpp") + + def sample_top_k(self, candidates: "_LlamaTokenDataArray", k: int, min_keep: int): + # llama_cpp.llama_sample_top_k( + # self.ctx, llama_cpp.byref(candidates.candidates), k, min_keep + # ) + raise NotImplementedError("sample_top_k is not implemented in llama.cpp") + + def sample_top_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): + # llama_cpp.llama_sample_top_p( + # self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep + # ) + raise NotImplementedError("sample_top_p is not implemented in llama.cpp") + + def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): + # llama_cpp.llama_sample_min_p( + # self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep + # ) + raise NotImplementedError("sample_min_p is not implemented in llama.cpp") + + def sample_typical( + self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int + ): + # llama_cpp.llama_sample_typical( + # self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep + # ) + raise NotImplementedError("sample_typical is not implemented in llama.cpp") + + def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float): + # llama_cpp.llama_sample_temp( + # self.ctx, llama_cpp.byref(candidates.candidates), temp + # ) + raise NotImplementedError("sample_temp is not implemented in llama.cpp") + + def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGrammar): + # llama_cpp.llama_sample_grammar( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # grammar.grammar, + # ) + raise NotImplementedError("sample_grammar is not implemented in llama.cpp") + + def sample_token_mirostat( + self, + candidates: "_LlamaTokenDataArray", + tau: float, + eta: float, + m: int, + mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], + ) -> int: + raise NotImplementedError("sample_token_mirostat is not implemented in llama.cpp") + # return llama_cpp.llama_sample_token_mirostat( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # tau, + # eta, + # m, + # mu, + # ) + + def sample_token_mirostat_v2( + self, + candidates: "_LlamaTokenDataArray", + tau: float, + eta: float, + mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], + ) -> int: + raise NotImplementedError("sample_token_mirostat_v2 is not implemented in llama.cpp") + # return llama_cpp.llama_sample_token_mirostat_v2( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # tau, + # eta, + # mu, + # ) + + def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int: + raise NotImplementedError("sample_token_greedy is not implemented in llama.cpp") + # return llama_cpp.llama_sample_token_greedy( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # ) + + def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: + raise NotImplementedError("sample_token is not implemented in llama.cpp") + # return llama_cpp.llama_sample_token( + # self.ctx, + # llama_cpp.byref(candidates.candidates), + # ) + + # Grammar + def grammar_accept_token(self, grammar: LlamaGrammar, token: int): + raise NotImplementedError("grammar_accept_token is not implemented in llama.cpp") + # llama_cpp.llama_grammar_accept_token(grammar.grammar, self.ctx, token) + + def reset_timings(self): + llama_cpp.llama_perf_context_reset(self.ctx) + + def print_timings(self): + llama_cpp.llama_perf_context_print(self.ctx) + + # Utility functions + @staticmethod + def default_params(): + """Get the default llama_context_params.""" + return llama_cpp.llama_context_default_params() + + +class LlamaBatch: + def __init__( + self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True + ): + self._n_tokens = n_tokens + self.embd = embd + self.n_seq_max = n_seq_max + self.verbose = verbose + self._exit_stack = ExitStack() + + batch = llama_cpp.llama_batch_init(self._n_tokens, self.embd, self.n_seq_max) + + if batch is None: + raise ValueError("Failed to create llama_batch") + + self.batch = batch + + def free_batch(): + if self.batch is None: + return + llama_cpp.llama_batch_free(self.batch) + self.batch = None + + self._exit_stack.callback(free_batch) + + def close(self): + self._exit_stack.close() + + def __del__(self): + self.close() + + def n_tokens(self) -> int: + return self.batch.n_tokens + + def reset(self): + self.batch.n_tokens = 0 + + def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool): + n_tokens = len(batch) + self.batch.n_tokens = n_tokens + for i in range(n_tokens): + self.batch.token[i] = batch[i] + self.batch.pos[i] = n_past + i + self.batch.seq_id[i][0] = 0 + self.batch.n_seq_id[i] = 1 + self.batch.logits[i] = logits_all + self.batch.logits[n_tokens - 1] = True + + def add_sequence(self, batch: Sequence[int], seq_id: int, logits_all: bool): + n_tokens = len(batch) + n_tokens0 = self.batch.n_tokens + self.batch.n_tokens += n_tokens + for i in range(n_tokens): + j = n_tokens0 + i + self.batch.token[j] = batch[i] + self.batch.pos[j] = i + self.batch.seq_id[j][0] = seq_id + self.batch.n_seq_id[j] = 1 + self.batch.logits[j] = logits_all + self.batch.logits[n_tokens - 1] = True + + +class LlamaTokenDataArray: + def __init__(self, *, n_vocab: int): + self.n_vocab = n_vocab + self.candidates_data = np.recarray( + (self.n_vocab,), + dtype=np.dtype( + [("id", np.intc), ("logit", np.single), ("p", np.single)], align=True + ), + ) + self.candidates = llama_cpp.llama_token_data_array( + data=self.candidates_data.ctypes.data_as(llama_cpp.llama_token_data_p), + size=self.n_vocab, + sorted=False, + ) + self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc) # type: ignore + self.default_candidates_data_p = np.zeros(self.n_vocab, dtype=np.single) + + def copy_logits(self, logits: npt.NDArray[np.single]): + self.candidates_data.id[:] = self.default_candidates_data_id + self.candidates_data.logit[:] = logits + self.candidates_data.p[:] = self.default_candidates_data_p + self.candidates.sorted = False + self.candidates.size = self.n_vocab + + +# Embedding functions + + +def normalize_embedding(embedding): + norm = float(np.linalg.norm(embedding)) + if norm == 0.0: + return embedding + return [v / norm for v in embedding] + + +# Python wrappers over common/sampling structs + + +@dataclass +class LlamaSamplingParams: + n_prev: int = 64 + n_probs: int = 0 + top_k: int = 40 + top_p: float = 0.95 + min_p: float = 0.05 + tfs_z: float = 1.00 + typical_p: float = 1.00 + temp: float = 0.80 + penalty_last_n: int = 64 + penalty_repeat: float = 1.0 + penalty_freq: float = 0.00 + penalty_present: float = 0.00 + mirostat: int = 0 + mirostat_tau: float = 5.00 + mirostat_eta: float = 0.10 + penalize_nl: bool = True + + grammar: str = "" + + cfg_negative_prompt: str = "" + cfg_scale: float = 1.00 + + logit_bias: dict[int, float] = field(default_factory=dict) + + +@dataclass +class LlamaSamplingContext: + params: LlamaSamplingParams = field(default_factory=LlamaSamplingParams) + mirostat_mu: ctypes.c_float = field(default_factory=ctypes.c_float) + grammar: Optional[LlamaGrammar] = None + # NOTE: Missing parsed_grammar + prev: list[int] = field(default_factory=list) + cur: list[llama_cpp.llama_token_data] = field(default_factory=list) + + def reset(self): + self.prev = [] + self.cur = [] + if self.grammar is not None: + self.grammar.reset() + + def cp(self): + return LlamaSamplingContext( + params=self.params, + mirostat_mu=self.mirostat_mu, + grammar=self.grammar, + prev=self.prev.copy(), + cur=self.cur.copy(), + ) + + def last(self) -> Optional[int]: + if len(self.prev) > 0: + return self.prev[-1] + else: + return None + + def prev_str(self, ctx_main: LlamaContext, n: int) -> str: + return ctx_main.model.detokenize(self.prev[-n:]).decode("utf-8") + + def sample( + self, + ctx_main: LlamaContext, + idx: int = 0, + logits_array: Optional[npt.NDArray[np.single]] = None, + ): + n_vocab = ctx_main.model.n_vocab() + id: int = 0 + + if logits_array is None: + logits = ctx_main.get_logits_ith(idx) + logits_array = np.array( + ctypes.cast(logits, ctypes.POINTER(ctypes.c_float * n_vocab)).contents, + dtype=np.single, + ) + + # apply logit_bias + for token, logit_bias in self.params.logit_bias.items(): + logits_array[token] += logit_bias + + token_data_array = LlamaTokenDataArray( + n_vocab=n_vocab + ) # TODO: Only create this once + token_data_array.copy_logits(logits_array) + + # apply penalties + if len(self.prev) > 0: + nl_token = ctx_main.model.token_nl() + nl_logit = logits_array[nl_token] + last_tokens = self.prev[-self.params.penalty_last_n :] + last_tokens_size = min(len(last_tokens), self.params.penalty_last_n) + if last_tokens_size > 0: + last_tokens_p = (llama_cpp.llama_token * len(last_tokens))(*last_tokens) + ctx_main.sample_repetition_penalties( + token_data_array, + last_tokens_p, + last_tokens_size, + self.params.penalty_repeat, + self.params.penalty_freq, + self.params.penalty_present, + ) + if not self.params.penalize_nl: + token_data_array.candidates_data.logit[nl_token] = nl_logit + + if self.grammar is not None: + ctx_main.sample_grammar(token_data_array, self.grammar) + + if self.params.temp < 0: + ctx_main.sample_softmax(token_data_array) + id = token_data_array.candidates_data.id[0] + elif self.params.temp == 0: + id = ctx_main.sample_token_greedy(token_data_array) + else: + if self.params.mirostat == 1: + mirostat_m = 100 + ctx_main.sample_temp(token_data_array, self.params.temp) + id = ctx_main.sample_token_mirostat( + token_data_array, + self.params.mirostat_tau, + self.params.mirostat_eta, + mirostat_m, + ctypes.pointer(self.mirostat_mu), + ) + elif self.params.mirostat == 2: + ctx_main.sample_temp(token_data_array, self.params.temp) + id = ctx_main.sample_token_mirostat_v2( + token_data_array, + self.params.mirostat_tau, + self.params.mirostat_eta, + ctypes.pointer(self.mirostat_mu), + ) + else: + min_keep = max(1, self.params.n_probs) + ctx_main.sample_top_k( + token_data_array, self.params.top_k, min_keep=min_keep + ) + ctx_main.sample_typical( + token_data_array, self.params.typical_p, min_keep=min_keep + ) + ctx_main.sample_top_p( + token_data_array, self.params.top_p, min_keep=min_keep + ) + ctx_main.sample_min_p( + token_data_array, self.params.min_p, min_keep=min_keep + ) + ctx_main.sample_temp(token_data_array, self.params.temp) + id = ctx_main.sample_token(token_data_array) + return id + + def accept(self, ctx_main: LlamaContext, id: int, apply_grammar: bool): + if apply_grammar and self.grammar is not None: + ctx_main.grammar_accept_token(self.grammar, id) + self.prev.append(id) + + +from typing import List, Callable, Optional, Union +import ctypes +import llama_cpp + + +class CustomSampler: + def __init__( + self, apply_func: typing.Callable[[llama_cpp.llama_token_data_array], None] + ): + self.apply_func = apply_func + + def apply_wrapper( + sampler: llama_cpp.llama_sampler_p, + cur_p: llama_cpp.llama_token_data_array_p, + ): + self.apply_func(cur_p) + + def free_wrapper(sampler: llama_cpp.llama_sampler_p): + pass + + sampler_i = llama_cpp.llama_sampler_i() + sampler_i.apply = llama_cpp.llama_sampler_i_apply(apply_wrapper) + self._apply_wrapper_ref = apply_wrapper + + sampler_i.name = llama_cpp.llama_sampler_i_name(0) + sampler_i.accept = llama_cpp.llama_sampler_i_accept(0) + sampler_i.reset = llama_cpp.llama_sampler_i_reset(0) + sampler_i.clone = llama_cpp.llama_sampler_i_clone(0) + sampler_i.free = llama_cpp.llama_sampler_i_free(0) + + self.sampler = llama_cpp.llama_sampler() + self.sampler.iface = ctypes.pointer(sampler_i) + self.sampler.ctx = None + + def get_sampler(self) -> llama_cpp.llama_sampler_p: + return ctypes.pointer(self.sampler) + + +class LlamaSampler: + def __init__(self): + params = llama_cpp.llama_sampler_chain_params() + self.sampler = llama_cpp.llama_sampler_chain_init(params) + self.samplers: List[llama_cpp.llama_sampler_p] = [] + self.custom_samplers: List[Tuple[int, CustomSampler]] = [] + + def add_greedy(self): + sampler = llama_cpp.llama_sampler_init_greedy() + self._add_sampler(sampler) + + def add_dist(self, seed: int): + sampler = llama_cpp.llama_sampler_init_dist(seed) + self._add_sampler(sampler) + + def add_softmax(self): + sampler = llama_cpp.llama_sampler_init_softmax() + self._add_sampler(sampler) + + def add_top_k(self, k: int): + sampler = llama_cpp.llama_sampler_init_top_k(k) + self._add_sampler(sampler) + + def add_top_p(self, p: float, min_keep: int): + sampler = llama_cpp.llama_sampler_init_top_p(p, min_keep) + self._add_sampler(sampler) + + def add_min_p(self, p: float, min_keep: int): + sampler = llama_cpp.llama_sampler_init_min_p(p, min_keep) + self._add_sampler(sampler) + + def add_typical(self, p: float, min_keep: int): + sampler = llama_cpp.llama_sampler_init_typical(p, min_keep) + self._add_sampler(sampler) + + def add_temp(self, temp: float): + sampler = llama_cpp.llama_sampler_init_temp(temp) + self._add_sampler(sampler) + + def add_temp_ext(self, t: float, delta: float, exponent: float): + sampler = llama_cpp.llama_sampler_init_temp_ext(t, delta, exponent) + self._add_sampler(sampler) + + def add_mirostat(self, n_vocab: int, seed: int, tau: float, eta: float, m: int): + sampler = llama_cpp.llama_sampler_init_mirostat(n_vocab, seed, tau, eta, m) + self._add_sampler(sampler) + + def add_mirostat_v2(self, seed: int, tau: float, eta: float): + sampler = llama_cpp.llama_sampler_init_mirostat_v2(seed, tau, eta) + self._add_sampler(sampler) + + def add_grammar(self, model: LlamaModel, grammar: LlamaGrammar): + sampler = llama_cpp.llama_sampler_init_grammar( + model.vocab, grammar._grammar.encode("utf-8"), grammar._root.encode("utf-8") + ) + self._add_sampler(sampler) + + def add_penalties( + self, + n_vocab: int, + special_eos_id: int, + linefeed_id: int, + penalty_last_n: int, + penalty_repeat: float, + penalty_freq: float, + penalty_present: float, + penalize_nl: bool, + ignore_eos: bool, + ): + sampler = llama_cpp.llama_sampler_init_penalties( + penalty_last_n, + penalty_repeat, + penalty_freq, + penalty_present, + ) + self._add_sampler(sampler) + + def init_logit_bias( + self, n_vocab: int, n_logit_bias, logit_bias: llama_cpp.llama_logit_bias_p + ): + sampler = llama_cpp.llama_sampler_init_logit_bias( + n_vocab, n_logit_bias, logit_bias + ) + self._add_sampler(sampler) + + def add_custom( + self, apply_func: Callable[[llama_cpp.llama_token_data_array], None] + ): + custom_sampler = CustomSampler(apply_func) + sampler = custom_sampler.get_sampler() + self._add_sampler(sampler) + # NOTE: Must remove custom samplers before free or llama.cpp will try to free them + self.custom_samplers.append( + (llama_cpp.llama_sampler_chain_n(self.sampler) - 1, custom_sampler) + ) + + def _add_sampler(self, sampler: llama_cpp.llama_sampler_p): + assert self.sampler is not None + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) + self.samplers.append(sampler) + + def get_seed(self) -> int: + assert self.sampler is not None + return llama_cpp.llama_sampler_get_seed(self.sampler) + + def sample(self, ctx: LlamaContext, idx: int) -> int: + assert self.sampler is not None + assert ctx.ctx is not None + return llama_cpp.llama_sampler_sample(self.sampler, ctx.ctx, idx) + + def close(self): + if self.sampler: + # NOTE: Must remove custom samplers before free or llama.cpp will try to free them + for i, _ in reversed(self.custom_samplers): + llama_cpp.llama_sampler_chain_remove(self.sampler, i) + llama_cpp.llama_sampler_free(self.sampler) + self.sampler = None + self.samplers.clear() + self.custom_samplers.clear() + + def __del__(self): + self.close() diff --git a/llama_cpp/_logger.py b/llama_cpp/_logger.py new file mode 100644 index 000000000..787b3f108 --- /dev/null +++ b/llama_cpp/_logger.py @@ -0,0 +1,47 @@ +import sys +import ctypes +import logging + +import llama_cpp + +# enum ggml_log_level { +# GGML_LOG_LEVEL_NONE = 0, +# GGML_LOG_LEVEL_INFO = 1, +# GGML_LOG_LEVEL_WARN = 2, +# GGML_LOG_LEVEL_ERROR = 3, +# GGML_LOG_LEVEL_DEBUG = 4, +# GGML_LOG_LEVEL_CONT = 5, // continue previous log +# }; +GGML_LOG_LEVEL_TO_LOGGING_LEVEL = { + 0: logging.CRITICAL, + 1: logging.INFO, + 2: logging.WARNING, + 3: logging.ERROR, + 4: logging.DEBUG, + 5: logging.DEBUG, +} + +logger = logging.getLogger("llama-cpp-python") + +_last_log_level = GGML_LOG_LEVEL_TO_LOGGING_LEVEL[0] + +# typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data); +@llama_cpp.llama_log_callback +def llama_log_callback( + level: int, + text: bytes, + user_data: ctypes.c_void_p, +): + # TODO: Correctly implement continue previous log + global _last_log_level + log_level = GGML_LOG_LEVEL_TO_LOGGING_LEVEL[level] if level != 5 else _last_log_level + if logger.level <= GGML_LOG_LEVEL_TO_LOGGING_LEVEL[level]: + print(text.decode("utf-8"), end="", flush=True, file=sys.stderr) + _last_log_level = log_level + + +llama_cpp.llama_log_set(llama_log_callback, ctypes.c_void_p(0)) + + +def set_verbose(verbose: bool): + logger.setLevel(logging.DEBUG if verbose else logging.ERROR) diff --git a/llama_cpp/_utils.py b/llama_cpp/_utils.py index 171f357f5..29628193b 100644 --- a/llama_cpp/_utils.py +++ b/llama_cpp/_utils.py @@ -1,11 +1,19 @@ import os import sys +from typing import Any, Dict + +# Avoid "LookupError: unknown encoding: ascii" when open() called in a destructor +outnull_file = open(os.devnull, "w") +errnull_file = open(os.devnull, "w") + +STDOUT_FILENO = 1 +STDERR_FILENO = 2 + class suppress_stdout_stderr(object): # NOTE: these must be "saved" here to avoid exceptions when using # this context manager inside of a __del__ method - open = open sys = sys os = os @@ -17,15 +25,8 @@ def __enter__(self): if self.disable: return self - # Check if sys.stdout and sys.stderr have fileno method - if not hasattr(self.sys.stdout, 'fileno') or not hasattr(self.sys.stderr, 'fileno'): - return self # Return the instance without making changes - - self.outnull_file = self.open(self.os.devnull, "w") - self.errnull_file = self.open(self.os.devnull, "w") - - self.old_stdout_fileno_undup = self.sys.stdout.fileno() - self.old_stderr_fileno_undup = self.sys.stderr.fileno() + self.old_stdout_fileno_undup = STDOUT_FILENO + self.old_stderr_fileno_undup = STDERR_FILENO self.old_stdout_fileno = self.os.dup(self.old_stdout_fileno_undup) self.old_stderr_fileno = self.os.dup(self.old_stderr_fileno_undup) @@ -33,11 +34,11 @@ def __enter__(self): self.old_stdout = self.sys.stdout self.old_stderr = self.sys.stderr - self.os.dup2(self.outnull_file.fileno(), self.old_stdout_fileno_undup) - self.os.dup2(self.errnull_file.fileno(), self.old_stderr_fileno_undup) + self.os.dup2(outnull_file.fileno(), self.old_stdout_fileno_undup) + self.os.dup2(errnull_file.fileno(), self.old_stderr_fileno_undup) - self.sys.stdout = self.outnull_file - self.sys.stderr = self.errnull_file + self.sys.stdout = outnull_file + self.sys.stderr = errnull_file return self def __exit__(self, *_): @@ -45,15 +46,33 @@ def __exit__(self, *_): return # Check if sys.stdout and sys.stderr have fileno method - if hasattr(self.sys.stdout, 'fileno') and hasattr(self.sys.stderr, 'fileno'): - self.sys.stdout = self.old_stdout - self.sys.stderr = self.old_stderr + self.sys.stdout = self.old_stdout + self.sys.stderr = self.old_stderr + + self.os.dup2(self.old_stdout_fileno, self.old_stdout_fileno_undup) + self.os.dup2(self.old_stderr_fileno, self.old_stderr_fileno_undup) + + self.os.close(self.old_stdout_fileno) + self.os.close(self.old_stderr_fileno) + + +class MetaSingleton(type): + """ + Metaclass for implementing the Singleton pattern. + """ + + _instances: Dict[type, Any] = {} + + def __call__(cls, *args: Any, **kwargs: Any) -> Any: + if cls not in cls._instances: + cls._instances[cls] = super(MetaSingleton, cls).__call__(*args, **kwargs) + return cls._instances[cls] - self.os.dup2(self.old_stdout_fileno, self.old_stdout_fileno_undup) - self.os.dup2(self.old_stderr_fileno, self.old_stderr_fileno_undup) - self.os.close(self.old_stdout_fileno) - self.os.close(self.old_stderr_fileno) +class Singleton(object, metaclass=MetaSingleton): + """ + Base class for implementing the Singleton pattern. + """ - self.outnull_file.close() - self.errnull_file.close() + def __init__(self): + super(Singleton, self).__init__() diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index c2c045549..7e9a6af23 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -1,724 +1,57 @@ +from __future__ import annotations + import os import sys import uuid import time +import json +import ctypes +import typing +import random +import fnmatch +import warnings +import contextlib import multiprocessing -from abc import ABC, abstractmethod + from typing import ( + Any, List, + Literal, Optional, Union, Generator, Sequence, Iterator, Deque, - Tuple, Callable, + Dict, ) -from collections import deque, OrderedDict +from collections import deque +from pathlib import Path -import diskcache -import ctypes from .llama_types import * from .llama_grammar import LlamaGrammar +from .llama_cache import ( + BaseLlamaCache, + LlamaCache, # type: ignore + LlamaDiskCache, # type: ignore + LlamaRAMCache, # type: ignore +) +from .llama_tokenizer import BaseLlamaTokenizer, LlamaTokenizer import llama_cpp.llama_cpp as llama_cpp import llama_cpp.llama_chat_format as llama_chat_format +from llama_cpp.llama_speculative import LlamaDraftModel + import numpy as np import numpy.typing as npt +import llama_cpp._internals as internals +from ._logger import set_verbose from ._utils import suppress_stdout_stderr -class BaseLlamaCache(ABC): - """Base cache class for a llama.cpp model.""" - - def __init__(self, capacity_bytes: int = (2 << 30)): - self.capacity_bytes = capacity_bytes - - @property - @abstractmethod - def cache_size(self) -> int: - raise NotImplementedError - - def _find_longest_prefix_key( - self, - key: Tuple[int, ...], - ) -> Optional[Tuple[int, ...]]: - pass - - @abstractmethod - def __getitem__(self, key: Sequence[int]) -> "LlamaState": - raise NotImplementedError - - @abstractmethod - def __contains__(self, key: Sequence[int]) -> bool: - raise NotImplementedError - - @abstractmethod - def __setitem__(self, key: Sequence[int], value: "LlamaState") -> None: - raise NotImplementedError - - -class LlamaRAMCache(BaseLlamaCache): - """Cache for a llama.cpp model using RAM.""" - - def __init__(self, capacity_bytes: int = (2 << 30)): - super().__init__(capacity_bytes) - self.capacity_bytes = capacity_bytes - self.cache_state: OrderedDict[Tuple[int, ...], "LlamaState"] = OrderedDict() - - @property - def cache_size(self): - return sum([state.llama_state_size for state in self.cache_state.values()]) - - def _find_longest_prefix_key( - self, - key: Tuple[int, ...], - ) -> Optional[Tuple[int, ...]]: - min_len = 0 - min_key = None - keys = ( - (k, Llama.longest_token_prefix(k, key)) for k in self.cache_state.keys() - ) - for k, prefix_len in keys: - if prefix_len > min_len: - min_len = prefix_len - min_key = k - return min_key - - def __getitem__(self, key: Sequence[int]) -> "LlamaState": - key = tuple(key) - _key = self._find_longest_prefix_key(key) - if _key is None: - raise KeyError("Key not found") - value = self.cache_state[_key] - self.cache_state.move_to_end(_key) - return value - - def __contains__(self, key: Sequence[int]) -> bool: - return self._find_longest_prefix_key(tuple(key)) is not None - - def __setitem__(self, key: Sequence[int], value: "LlamaState"): - key = tuple(key) - if key in self.cache_state: - del self.cache_state[key] - self.cache_state[key] = value - while self.cache_size > self.capacity_bytes and len(self.cache_state) > 0: - self.cache_state.popitem(last=False) - - -# Alias for backwards compatibility -LlamaCache = LlamaRAMCache - - -class LlamaDiskCache(BaseLlamaCache): - """Cache for a llama.cpp model using disk.""" - - def __init__( - self, cache_dir: str = ".cache/llama_cache", capacity_bytes: int = (2 << 30) - ): - super().__init__(capacity_bytes) - self.cache = diskcache.Cache(cache_dir) - - @property - def cache_size(self): - return int(self.cache.volume()) # type: ignore - - def _find_longest_prefix_key( - self, - key: Tuple[int, ...], - ) -> Optional[Tuple[int, ...]]: - min_len = 0 - min_key: Optional[Tuple[int, ...]] = None - for k in self.cache.iterkeys(): # type: ignore - prefix_len = Llama.longest_token_prefix(k, key) - if prefix_len > min_len: - min_len = prefix_len - min_key = k # type: ignore - return min_key - - def __getitem__(self, key: Sequence[int]) -> "LlamaState": - key = tuple(key) - _key = self._find_longest_prefix_key(key) - if _key is None: - raise KeyError("Key not found") - value: "LlamaState" = self.cache.pop(_key) # type: ignore - # NOTE: This puts an integer as key in cache, which breaks, - # Llama.longest_token_prefix(k, key) above since k is not a tuple of ints/tokens - # self.cache.push(_key, side="front") # type: ignore - return value - - def __contains__(self, key: Sequence[int]) -> bool: - return self._find_longest_prefix_key(tuple(key)) is not None - - def __setitem__(self, key: Sequence[int], value: "LlamaState"): - print("LlamaDiskCache.__setitem__: called", file=sys.stderr) - key = tuple(key) - if key in self.cache: - print("LlamaDiskCache.__setitem__: delete", file=sys.stderr) - del self.cache[key] - self.cache[key] = value - print("LlamaDiskCache.__setitem__: set", file=sys.stderr) - while self.cache_size > self.capacity_bytes and len(self.cache) > 0: - key_to_remove = next(iter(self.cache)) - del self.cache[key_to_remove] - print("LlamaDiskCache.__setitem__: trim", file=sys.stderr) - - -class LlamaState: - def __init__( - self, - input_ids: npt.NDArray[np.intc], - scores: npt.NDArray[np.single], - n_tokens: int, - llama_state: bytes, - llama_state_size: int, - ): - self.input_ids = input_ids - self.scores = scores - self.n_tokens = n_tokens - self.llama_state = llama_state - self.llama_state_size = llama_state_size - - -LogitsProcessor = Callable[ - [npt.NDArray[np.intc], npt.NDArray[np.single]], npt.NDArray[np.single] -] - - -class LogitsProcessorList(List[LogitsProcessor]): - def __call__( - self, input_ids: npt.NDArray[np.intc], scores: npt.NDArray[np.single] - ) -> npt.NDArray[np.single]: - for processor in self: - scores = processor(input_ids, scores) - return scores - - -StoppingCriteria = Callable[[npt.NDArray[np.intc], npt.NDArray[np.single]], bool] - - -class StoppingCriteriaList(List[StoppingCriteria]): - def __call__( - self, input_ids: npt.NDArray[np.intc], logits: npt.NDArray[np.single] - ) -> bool: - return any([stopping_criteria(input_ids, logits) for stopping_criteria in self]) - - -class _LlamaModel: - """Intermediate Python wrapper for a llama.cpp llama_model. - - NOTE: For stability it's recommended you use the Llama class instead.""" - - _llama_free_model = None - # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - suppress_stdout_stderr = suppress_stdout_stderr - - def __init__( - self, - *, - path_model: str, - params: llama_cpp.llama_model_params, - verbose: bool = True, - ): - self.path_model = path_model - self.params = params - self.verbose = verbose - - self._llama_free_model = llama_cpp._lib.llama_free_model # type: ignore - - if not os.path.exists(path_model): - raise ValueError(f"Model path does not exist: {path_model}") - - with suppress_stdout_stderr(disable=self.verbose): - self.model = llama_cpp.llama_load_model_from_file( - self.path_model.encode("utf-8"), self.params - ) - - def __del__(self): - with self.suppress_stdout_stderr(disable=self.verbose): - if self.model is not None and self._llama_free_model is not None: - self._llama_free_model(self.model) - self.model = None - - def vocab_type(self) -> int: - assert self.model is not None - return llama_cpp.llama_vocab_type(self.model) - - def n_vocab(self) -> int: - assert self.model is not None - return llama_cpp.llama_n_vocab(self.model) - - def n_ctx_train(self) -> int: - assert self.model is not None - return llama_cpp.llama_n_ctx_train(self.model) - - def n_embd(self) -> int: - assert self.model is not None - return llama_cpp.llama_n_embd(self.model) - - def rope_freq_scale_train(self) -> float: - assert self.model is not None - return llama_cpp.llama_rope_freq_scale_train(self.model) - - def desc(self) -> str: - assert self.model is not None - buf = ctypes.create_string_buffer(1024) - llama_cpp.llama_model_desc(self.model, buf, 1024) # type: ignore - return buf.value.decode("utf-8") - - def size(self) -> int: - assert self.model is not None - return llama_cpp.llama_model_size(self.model) - - def n_params(self) -> int: - assert self.model is not None - return llama_cpp.llama_model_n_params(self.model) - - def get_tensor(self, name: str) -> ctypes.c_void_p: - assert self.model is not None - return llama_cpp.llama_get_model_tensor(self.model, name.encode("utf-8")) - - def apply_lora_from_file( - self, - lora_path: str, - scale: float, - path_base_model: Optional[str], - n_threads: int, - ): - assert self.model is not None - return llama_cpp.llama_model_apply_lora_from_file( - self.model, - lora_path.encode("utf-8"), - scale, - path_base_model.encode("utf-8") - if path_base_model is not None - else llama_cpp.c_char_p(0), - n_threads, - ) - - # Vocab - - def token_get_text(self, token: int) -> str: - # TODO: Fix - assert self.model is not None - return llama_cpp.llama_token_get_text(self.model, token).decode("utf-8") - - def token_get_score(self, token: int) -> float: - assert self.model is not None - return llama_cpp.llama_token_get_score(self.model, token) - - def token_get_type(self, token: int) -> int: - assert self.model is not None - return llama_cpp.llama_token_get_type(self.model, token) - - # Special tokens - - def token_bos(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_bos(self.model) - - def token_eos(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_eos(self.model) - - def token_nl(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_nl(self.model) - - def token_prefix(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_prefix(self.model) - - def token_middle(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_middle(self.model) - - def token_suffix(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_suffix(self.model) - - def token_eot(self) -> int: - assert self.model is not None - return llama_cpp.llama_token_eot(self.model) - - # Tokenization - - def tokenize(self, text: bytes, add_bos: bool, special: bool): - assert self.model is not None - n_ctx = self.n_ctx_train() - tokens = (llama_cpp.llama_token * n_ctx)() - n_tokens = llama_cpp.llama_tokenize( - self.model, text, len(text), tokens, n_ctx, add_bos, special - ) - if n_tokens < 0: - n_tokens = abs(n_tokens) - tokens = (llama_cpp.llama_token * n_tokens)() - n_tokens = llama_cpp.llama_tokenize( - self.model, text, len(text), tokens, n_tokens, add_bos, special - ) - if n_tokens < 0: - raise RuntimeError( - f'Failed to tokenize: text="{text}" n_tokens={n_tokens}' - ) - return list(tokens[:n_tokens]) - - def token_to_piece(self, token: int) -> bytes: - assert self.model is not None - buf = ctypes.create_string_buffer(32) - llama_cpp.llama_token_to_piece(self.model, token, buf, 32) # type: ignore - return bytes(buf) - - def detokenize(self, tokens: List[int]) -> bytes: - assert self.model is not None - output = b"" - size = 32 - buffer = (ctypes.c_char * size)() - for token in tokens: - n = llama_cpp.llama_token_to_piece( - self.model, llama_cpp.llama_token(token), buffer, size - ) - assert n <= size - output += bytes(buffer[:n]) - # NOTE: Llama1 models automatically added a space at the start of the prompt - # this line removes a leading space if the first token is a beginning of sentence token - return ( - output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() else output - ) - - @staticmethod - def default_params(): - """Get the default llama_model_params.""" - return llama_cpp.llama_model_default_params() - - -class _LlamaContext: - """Intermediate Python wrapper for a llama.cpp llama_context. - - NOTE: For stability it's recommended you use the Llama class instead.""" - - _llama_free = None - # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - suppress_stdout_stderr = suppress_stdout_stderr - - def __init__( - self, - *, - model: _LlamaModel, - params: llama_cpp.llama_context_params, - verbose: bool = True, - ): - self.model = model - self.params = params - self.verbose = verbose - - self._llama_free = llama_cpp._lib.llama_free # type: ignore - - with suppress_stdout_stderr(disable=self.verbose): - self.ctx = llama_cpp.llama_new_context_with_model( - self.model.model, self.params - ) - - def __del__(self): - with self.suppress_stdout_stderr(disable=self.verbose): - if self.ctx is not None and self._llama_free is not None: - self._llama_free(self.ctx) - self.ctx = None - - def n_ctx(self) -> int: - assert self.ctx is not None - return llama_cpp.llama_n_ctx(self.ctx) - - def kv_cache_clear(self): - assert self.ctx is not None - llama_cpp.llama_kv_cache_clear(self.ctx) - - def kv_cache_seq_rm(self, seq_id: int, p0: int, p1: int): - assert self.ctx is not None - llama_cpp.llama_kv_cache_seq_rm(self.ctx, seq_id, p0, p1) - - def kv_cache_seq_cp(self, seq_id_src: int, seq_id_dst: int, p0: int, p1: int): - assert self.ctx is not None - llama_cpp.llama_kv_cache_seq_cp(self.ctx, seq_id_src, seq_id_dst, p0, p1) - - def kv_cache_seq_keep(self, seq_id: int): - assert self.ctx is not None - llama_cpp.llama_kv_cache_seq_keep(self.ctx, seq_id) - - def kv_cache_seq_shift(self, seq_id: int, p0: int, p1: int, shift: int): - assert self.ctx is not None - llama_cpp.llama_kv_cache_seq_shift(self.ctx, seq_id, p0, p1, shift) - - def get_state_size(self) -> int: - assert self.ctx is not None - return llama_cpp.llama_get_state_size(self.ctx) - - # TODO: copy_state_data - - # TODO: set_state_data - - # TODO: llama_load_session_file - - # TODO: llama_save_session_file - - def decode(self, batch: "_LlamaBatch"): - assert self.ctx is not None - assert batch.batch is not None - return_code = llama_cpp.llama_decode( - ctx=self.ctx, - batch=batch.batch, - ) - if return_code != 0: - raise RuntimeError(f"llama_decode returned {return_code}") - - def set_n_threads(self, n_threads: int, n_threads_batch: int): - assert self.ctx is not None - llama_cpp.llama_set_n_threads(self.ctx, n_threads, n_threads_batch) - - def get_logits(self): - assert self.ctx is not None - return llama_cpp.llama_get_logits(self.ctx) - - def get_logits_ith(self, i: int): - assert self.ctx is not None - return llama_cpp.llama_get_logits_ith(self.ctx, i) - - def get_embeddings(self): - assert self.ctx is not None - return llama_cpp.llama_get_embeddings(self.ctx) - - # Sampling functions - - def set_rng_seed(self, seed: int): - assert self.ctx is not None - llama_cpp.llama_set_rng_seed(self.ctx, seed) - - def sample_repetition_penalties( - self, - candidates: "_LlamaTokenDataArray", - last_tokens_data: "llama_cpp.Array[llama_cpp.llama_token]", - penalty_last_n: int, - penalty_repeat: float, - penalty_freq: float, - penalty_present: float, - ): - assert self.ctx is not None - llama_cpp.llama_sample_repetition_penalties( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - last_tokens_data, - penalty_last_n, - penalty_repeat, - penalty_freq, - penalty_present, - ) - - def sample_classifier_free_guidance( - self, - candidates: "_LlamaTokenDataArray", - guidance_ctx: "_LlamaContext", - scale: float, - ): - assert self.ctx is not None - assert guidance_ctx.ctx is not None - llama_cpp.llama_sample_classifier_free_guidance( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - guidance_ctx.ctx, - scale, - ) - - def sample_softmax(self, candidates: "_LlamaTokenDataArray"): - assert self.ctx is not None - llama_cpp.llama_sample_softmax( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - ) - - def sample_top_k(self, candidates: "_LlamaTokenDataArray", k: int, min_keep: int): - assert self.ctx is not None - llama_cpp.llama_sample_top_k( - self.ctx, ctypes.byref(candidates.candidates), k, min_keep # type: ignore - ) - - def sample_top_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - assert self.ctx is not None - llama_cpp.llama_sample_top_p( - self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore - ) - - def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - assert self.ctx is not None - llama_cpp.llama_sample_min_p( - self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore - ) - - def sample_tail_free( - self, candidates: "_LlamaTokenDataArray", z: float, min_keep: int - ): - assert self.ctx is not None - llama_cpp.llama_sample_tail_free( - self.ctx, ctypes.byref(candidates.candidates), z, min_keep # type: ignore - ) - - def sample_typical( - self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int - ): - assert self.ctx is not None - llama_cpp.llama_sample_typical( - self.ctx, ctypes.byref(candidates.candidates), p, min_keep # type: ignore - ) - - def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float): - assert self.ctx is not None - llama_cpp.llama_sample_temp( - self.ctx, ctypes.byref(candidates.candidates), temp # type: ignore - ) - - def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGrammar): - assert self.ctx is not None - assert grammar.grammar is not None - llama_cpp.llama_sample_grammar( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - grammar.grammar, - ) - - def sample_token_mirostat( - self, - candidates: "_LlamaTokenDataArray", - tau: float, - eta: float, - m: int, - mu: float, - ) -> int: - assert self.ctx is not None - return llama_cpp.llama_sample_token_mirostat( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - tau, - eta, - m, - ctypes.pointer(ctypes.c_float(mu)), - ) - - def sample_token_mirostat_v2( - self, candidates: "_LlamaTokenDataArray", tau: float, eta: float, mu: float - ) -> int: - assert self.ctx is not None - return llama_cpp.llama_sample_token_mirostat_v2( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - tau, - eta, - ctypes.pointer(ctypes.c_float(mu)), - ) - - def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int: - assert self.ctx is not None - return llama_cpp.llama_sample_token_greedy( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - ) - - def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: - assert self.ctx is not None - return llama_cpp.llama_sample_token( - self.ctx, - ctypes.byref(candidates.candidates), # type: ignore - ) - - # Grammar - def grammar_accept_token(self, grammar: LlamaGrammar, token: int): - assert self.ctx is not None - assert grammar.grammar is not None - llama_cpp.llama_grammar_accept_token(self.ctx, grammar.grammar, token) - - def reset_timings(self): - assert self.ctx is not None - llama_cpp.llama_reset_timings(self.ctx) - - def print_timings(self): - assert self.ctx is not None - llama_cpp.llama_print_timings(self.ctx) - - # Utility functions - @staticmethod - def default_params(): - """Get the default llama_context_params.""" - return llama_cpp.llama_context_default_params() - - -class _LlamaBatch: - _llama_batch_free = None - # NOTE: this must be "saved" here to avoid exceptions when calling __del__ - suppress_stdout_stderr = suppress_stdout_stderr - - def __init__( - self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True - ): - self.n_tokens = n_tokens - self.embd = embd - self.n_seq_max = n_seq_max - self.verbose = verbose - - self._llama_batch_free = llama_cpp._lib.llama_batch_free # type: ignore - - with suppress_stdout_stderr(disable=self.verbose): - self.batch = llama_cpp.llama_batch_init( - self.n_tokens, self.embd, self.n_seq_max - ) - - def __del__(self): - with self.suppress_stdout_stderr(disable=self.verbose): - if self.batch is not None and self._llama_batch_free is not None: - self._llama_batch_free(self.batch) - self.batch = None - - def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool): - assert self.batch is not None - n_tokens = len(batch) - self.batch.n_tokens = n_tokens - for i in range(n_tokens): - self.batch.token[i] = batch[i] - self.batch.pos[i] = n_past + i - self.batch.seq_id[i][0] = 0 - self.batch.n_seq_id[i] = 1 - self.batch.logits[i] = logits_all - self.batch.logits[n_tokens - 1] = True - - -class _LlamaTokenDataArray: - def __init__(self, *, n_vocab: int): - self.n_vocab = n_vocab - self.candidates_data = np.array( - [], - dtype=np.dtype( - [("id", np.intc), ("logit", np.single), ("p", np.single)], align=True - ), - ) - self.candidates_data.resize(3, self.n_vocab, refcheck=False) - self.candidates = llama_cpp.llama_token_data_array( - data=self.candidates_data.ctypes.data_as(llama_cpp.llama_token_data_p), - size=self.n_vocab, - sorted=False, - ) - self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc) - self.default_candidates_data_p = np.zeros(self.n_vocab, dtype=np.single) - - def copy_logits(self, logits: npt.NDArray[np.single]): - self.candidates_data["id"][:] = self.default_candidates_data_id - self.candidates_data["logit"][:] = logits - self.candidates_data["p"][:] = self.default_candidates_data_p - self.candidates.data = self.candidates_data.ctypes.data_as( - llama_cpp.llama_token_data_p - ) - self.candidates.sorted = llama_cpp.c_bool(False) - self.candidates.size = llama_cpp.c_size_t(self.n_vocab) - - class Llama: """High-level Python wrapper for a llama.cpp model.""" @@ -730,18 +63,25 @@ def __init__( *, # Model Params n_gpu_layers: int = 0, + split_mode: int = llama_cpp.LLAMA_SPLIT_MODE_LAYER, main_gpu: int = 0, tensor_split: Optional[List[float]] = None, + rpc_servers: Optional[str] = None, vocab_only: bool = False, use_mmap: bool = True, use_mlock: bool = False, + kv_overrides: Optional[Dict[str, Union[bool, int, float, str]]] = None, # Context Params seed: int = llama_cpp.LLAMA_DEFAULT_SEED, n_ctx: int = 512, n_batch: int = 512, + n_ubatch: int = 512, n_threads: Optional[int] = None, n_threads_batch: Optional[int] = None, - rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED, + rope_scaling_type: Optional[ + int + ] = llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, + pooling_type: int = llama_cpp.LLAMA_POOLING_TYPE_UNSPECIFIED, rope_freq_base: float = 0.0, rope_freq_scale: float = 0.0, yarn_ext_factor: float = -1.0, @@ -749,22 +89,31 @@ def __init__( yarn_beta_fast: float = 32.0, yarn_beta_slow: float = 1.0, yarn_orig_ctx: int = 0, - mul_mat_q: bool = True, logits_all: bool = False, embedding: bool = False, - offload_kqv: bool = False, + offload_kqv: bool = True, + flash_attn: bool = False, # Sampling Params + no_perf: bool = False, last_n_tokens_size: int = 64, # LoRA Params lora_base: Optional[str] = None, lora_scale: float = 1.0, lora_path: Optional[str] = None, # Backend Params - numa: bool = False, + numa: Union[bool, int] = False, # Chat Format Params - chat_format: str = "llama-2", + chat_format: Optional[str] = None, chat_handler: Optional[llama_chat_format.LlamaChatCompletionHandler] = None, + # Speculative Decoding + draft_model: Optional[LlamaDraftModel] = None, + # Tokenizer Override + tokenizer: Optional[BaseLlamaTokenizer] = None, + # KV cache quantization + type_k: Optional[int] = None, + type_v: Optional[int] = None, # Misc + spm_infill: bool = False, verbose: bool = True, # Extra Params **kwargs, # type: ignore @@ -798,17 +147,22 @@ def __init__( Args: model_path: Path to the model. n_gpu_layers: Number of layers to offload to GPU (-ngl). If -1, all layers are offloaded. - main_gpu: The GPU that is used for scratch and small tensors. + split_mode: How to split the model across GPUs. See llama_cpp.LLAMA_SPLIT_* for options. + main_gpu: main_gpu interpretation depends on split_mode: LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model. LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results. LLAMA_SPLIT_MODE_LAYER: ignored tensor_split: How split tensors should be distributed across GPUs. If None, the model is not split. + rpc_servers: Comma separated list of RPC servers to use for offloading vocab_only: Only load the vocabulary no weights. use_mmap: Use mmap if possible. use_mlock: Force the system to keep the model in RAM. + kv_overrides: Key-value overrides for the model. seed: RNG seed, -1 for random n_ctx: Text context, 0 = from model n_batch: Prompt processing maximum batch size + n_ubatch: Physical batch size n_threads: Number of threads to use for generation n_threads_batch: Number of threads to use for batch processing rope_scaling_type: RoPE scaling type, from `enum llama_rope_scaling_type`. ref: https://github.com/ggerganov/llama.cpp/pull/2054 + pooling_type: Pooling type, from `enum llama_pooling_type`. rope_freq_base: RoPE base frequency, 0 = from model rope_freq_scale: RoPE frequency scaling factor, 0 = from model yarn_ext_factor: YaRN extrapolation mix factor, negative = from model @@ -819,13 +173,20 @@ def __init__( logits_all: Return logits for all tokens, not just the last token. Must be True for completion to return logprobs. embedding: Embedding mode only. offload_kqv: Offload K, Q, V to GPU. + flash_attn: Use flash attention. + no_perf: Measure performance timings. last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque. lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model. lora_path: Path to a LoRA file to apply to the model. - numa: Enable NUMA support. (NOTE: The initial value of this parameter is used for the remainder of the program as this value is set in llama_backend_init) + numa: numa policy chat_format: String specifying the chat format to use when calling create_chat_completion. chat_handler: Optional chat handler to use when calling create_chat_completion. + draft_model: Optional draft model to use for speculative decoding. + tokenizer: Optional tokenizer to override the default tokenizer from llama.cpp. verbose: Print verbose output to stderr. + type_k: KV cache data type for K (default: f16) + type_v: KV cache data type for V (default: f16) + spm_infill: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. Raises: ValueError: If the model path does not exist. @@ -834,13 +195,28 @@ def __init__( A Llama instance. """ self.verbose = verbose + self._stack = contextlib.ExitStack() + + set_verbose(verbose) - self.numa = numa if not Llama.__backend_initialized: - with suppress_stdout_stderr(disable=self.verbose): - llama_cpp.llama_backend_init(self.numa) + with suppress_stdout_stderr(disable=verbose): + llama_cpp.llama_backend_init() Llama.__backend_initialized = True + if isinstance(numa, bool): + self.numa = ( + llama_cpp.GGML_NUMA_STRATEGY_DISTRIBUTE + if numa + else llama_cpp.GGML_NUMA_STRATEGY_DISABLED + ) + else: + self.numa = numa + + if self.numa != llama_cpp.GGML_NUMA_STRATEGY_DISABLED: + with suppress_stdout_stderr(disable=verbose): + llama_cpp.llama_numa_init(self.numa) + self.model_path = model_path # Model Params @@ -848,9 +224,15 @@ def __init__( self.model_params.n_gpu_layers = ( 0x7FFFFFFF if n_gpu_layers == -1 else n_gpu_layers ) # 0x7FFFFFFF is INT32 max, will be auto set to all layers + self.model_params.split_mode = split_mode self.model_params.main_gpu = main_gpu + if rpc_servers is not None: + self.model_params.rpc_servers = rpc_servers.encode("utf-8") + self._rpc_servers = rpc_servers + else: + self._rpc_servers = None self.tensor_split = tensor_split - self._p_tensor_split = None + self._c_tensor_split = None if self.tensor_split is not None: if len(self.tensor_split) > llama_cpp.LLAMA_MAX_DEVICES: raise ValueError( @@ -866,23 +248,80 @@ def __init__( self.model_params.use_mmap = use_mmap if lora_path is None else False self.model_params.use_mlock = use_mlock + # kv_overrides is the original python dict + self.kv_overrides = kv_overrides + if kv_overrides is not None: + # _kv_overrides_array is a ctypes.Array of llama_model_kv_override Structs + kvo_array_len = len(kv_overrides) + 1 # for sentinel element + self._kv_overrides_array = ( + llama_cpp.llama_model_kv_override * kvo_array_len + )() + + for i, (k, v) in enumerate(kv_overrides.items()): + self._kv_overrides_array[i].key = k.encode("utf-8") + if isinstance(v, bool): + self._kv_overrides_array[ + i + ].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_BOOL + self._kv_overrides_array[i].value.val_bool = v + elif isinstance(v, int): + self._kv_overrides_array[ + i + ].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_INT + self._kv_overrides_array[i].value.val_i64 = v + elif isinstance(v, float): + self._kv_overrides_array[ + i + ].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_FLOAT + self._kv_overrides_array[i].value.val_f64 = v + elif isinstance(v, str): # type: ignore + v_bytes = v.encode("utf-8") + if len(v_bytes) > 128: # TODO: Make this a constant + raise ValueError(f"Value for {k} is too long: {v}") + v_bytes = v_bytes.ljust(128, b"\0") + self._kv_overrides_array[ + i + ].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR + # copy min(v_bytes, 128) to str_value + address = typing.cast( + int, + ctypes.addressof(self._kv_overrides_array[i].value) + + llama_cpp.llama_model_kv_override_value.val_str.offset, + ) + buffer_start = ctypes.cast(address, ctypes.POINTER(ctypes.c_char)) + ctypes.memmove( + buffer_start, + v_bytes, + 128, + ) + else: + raise ValueError(f"Unknown value type for {k}: {v}") + + self._kv_overrides_array[ + -1 + ].key = b"\0" # ensure sentinel element is zeroed + self.model_params.kv_overrides = self._kv_overrides_array + self.n_batch = min(n_ctx, n_batch) # ??? self.n_threads = n_threads or max(multiprocessing.cpu_count() // 2, 1) - self.n_threads_batch = n_threads_batch or max( - multiprocessing.cpu_count() // 2, 1 - ) + self.n_threads_batch = n_threads_batch or multiprocessing.cpu_count() + + # Used by the sampler + self._seed = seed or llama_cpp.LLAMA_DEFAULT_SEED + # Context Params self.context_params = llama_cpp.llama_context_default_params() - self.context_params.seed = seed self.context_params.n_ctx = n_ctx self.context_params.n_batch = self.n_batch + self.context_params.n_ubatch = min(self.n_batch, n_ubatch) self.context_params.n_threads = self.n_threads self.context_params.n_threads_batch = self.n_threads_batch self.context_params.rope_scaling_type = ( rope_scaling_type if rope_scaling_type is not None - else llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED + else llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED ) + self.context_params.pooling_type = pooling_type self.context_params.rope_freq_base = ( rope_freq_base if rope_freq_base != 0.0 else 0 ) @@ -902,12 +341,19 @@ def __init__( yarn_beta_slow if yarn_beta_slow != 0.0 else 0 ) self.context_params.yarn_orig_ctx = yarn_orig_ctx if yarn_orig_ctx != 0 else 0 - self.context_params.mul_mat_q = mul_mat_q - self.context_params.logits_all = logits_all - self.context_params.embedding = embedding + self.context_params.logits_all = ( + logits_all if draft_model is None else True + ) # Must be set to True for speculative decoding + self.context_params.embeddings = embedding # TODO: Rename to embeddings self.context_params.offload_kqv = offload_kqv - + self.context_params.flash_attn = flash_attn + # KV cache quantization + if type_k is not None: + self.context_params.type_k = type_k + if type_v is not None: + self.context_params.type_v = type_v # Sampling Params + self.context_params.no_perf = no_perf self.last_n_tokens_size = last_n_tokens_size self.cache: Optional[BaseLlamaCache] = None @@ -916,41 +362,78 @@ def __init__( self.lora_scale = lora_scale self.lora_path = lora_path + self.spm_infill = spm_infill + if not os.path.exists(model_path): raise ValueError(f"Model path does not exist: {model_path}") - self._model = _LlamaModel( - path_model=self.model_path, params=self.model_params, verbose=self.verbose + self._model = self._stack.enter_context( + contextlib.closing( + internals.LlamaModel( + path_model=self.model_path, + params=self.model_params, + verbose=self.verbose, + ) + ) ) + + # Override tokenizer + self.tokenizer_ = tokenizer or LlamaTokenizer(self) + # Set the default value for the context and correct the batch if n_ctx == 0: n_ctx = self._model.n_ctx_train() self.n_batch = min(n_ctx, n_batch) self.context_params.n_ctx = self._model.n_ctx_train() self.context_params.n_batch = self.n_batch - - self._ctx = _LlamaContext( - model=self._model, - params=self.context_params, - verbose=self.verbose, + self.context_params.n_ubatch = min(self.n_batch, n_ubatch) + + self._ctx = self._stack.enter_context( + contextlib.closing( + internals.LlamaContext( + model=self._model, + params=self.context_params, + verbose=self.verbose, + ) + ) ) - self._batch = _LlamaBatch( - n_tokens=self.n_batch, - embd=0, - n_seq_max=self.context_params.n_ctx, - verbose=self.verbose, + self._batch = self._stack.enter_context( + contextlib.closing( + internals.LlamaBatch( + n_tokens=self.n_batch, + embd=0, + n_seq_max=self.context_params.n_ctx, + verbose=self.verbose, + ) + ) ) + self._lora_adapter: Optional[llama_cpp.llama_adapter_lora_p] = None + if self.lora_path: - if self._model.apply_lora_from_file( - self.lora_path, - self.lora_scale, - self.lora_base, - self.n_threads, + self._lora_adapter = llama_cpp.llama_adapter_lora_init( + self._model.model, + self.lora_path.encode("utf-8"), + ) + if self._lora_adapter is None: + raise RuntimeError( + f"Failed to initialize LoRA adapter from lora path: {self.lora_path}" + ) + + def free_lora_adapter(): + if self._lora_adapter is None: + return + llama_cpp.llama_adapter_lora_free(self._lora_adapter) + self._lora_adapter = None + + self._stack.callback(free_lora_adapter) + + if llama_cpp.llama_set_adapter_lora( + self._ctx.ctx, self._lora_adapter, self.lora_scale ): raise RuntimeError( - f"Failed to apply LoRA from lora path: {self.lora_path} to base path: {self.lora_base}" + f"Failed to set LoRA adapter from lora path: {self.lora_path}" ) if self.verbose: @@ -958,6 +441,11 @@ def __init__( self.chat_format = chat_format self.chat_handler = chat_handler + self._chat_handlers: Dict[ + str, llama_chat_format.LlamaChatCompletionHandler + ] = {} + + self.draft_model = draft_model self._n_vocab = self.n_vocab() self._n_ctx = self.n_ctx() @@ -965,22 +453,103 @@ def __init__( self._token_nl = self.token_nl() self._token_eos = self.token_eos() - self._candidates = _LlamaTokenDataArray(n_vocab=self._n_vocab) + self._candidates = internals.LlamaTokenDataArray(n_vocab=self._n_vocab) self.n_tokens = 0 self.input_ids: npt.NDArray[np.intc] = np.ndarray((n_ctx,), dtype=np.intc) self.scores: npt.NDArray[np.single] = np.ndarray( - (n_ctx, self._n_vocab), dtype=np.single + (n_ctx if logits_all == True else n_batch, self._n_vocab), dtype=np.single + ) + + self._mirostat_mu = ctypes.c_float( + 2.0 * 5.0 + ) # TODO: Move this to sampling context + + try: + self.metadata = self._model.metadata() + except Exception as e: + self.metadata = {} + if self.verbose: + print(f"Failed to load metadata: {e}", file=sys.stderr) + + if self.verbose: + print(f"Model metadata: {self.metadata}", file=sys.stderr) + + eos_token_id = self.token_eos() + bos_token_id = self.token_bos() + + eos_token = ( + self._model.token_get_text(eos_token_id) if eos_token_id != -1 else "" + ) + bos_token = ( + self._model.token_get_text(bos_token_id) if bos_token_id != -1 else "" + ) + + # Unfortunately the llama.cpp API does not return metadata arrays, so we can't get template names from tokenizer.chat_templates + template_choices = dict( + (name[10:], template) + for name, template in self.metadata.items() + if name.startswith("tokenizer.chat_template.") ) + if "tokenizer.chat_template" in self.metadata: + template_choices["chat_template.default"] = self.metadata[ + "tokenizer.chat_template" + ] + + if self.verbose and template_choices: + print( + f"Available chat formats from metadata: {', '.join(template_choices.keys())}", + file=sys.stderr, + ) + + for name, template in template_choices.items(): + self._chat_handlers[name] = llama_chat_format.Jinja2ChatFormatter( + template=template, + eos_token=eos_token, + bos_token=bos_token, + stop_token_ids=[eos_token_id], + ).to_chat_handler() + + if ( + self.chat_format is None + and self.chat_handler is None + and "chat_template.default" in template_choices + ): + chat_format = llama_chat_format.guess_chat_format_from_gguf_metadata( + self.metadata + ) + + if chat_format is not None: + self.chat_format = chat_format + if self.verbose: + print(f"Guessed chat format: {chat_format}", file=sys.stderr) + else: + if self.verbose: + print( + f"Using gguf chat template: {template_choices['chat_template.default']}", + file=sys.stderr, + ) + print(f"Using chat eos_token: {eos_token}", file=sys.stderr) + print(f"Using chat bos_token: {bos_token}", file=sys.stderr) + + self.chat_format = "chat_template.default" + + if self.chat_format is None and self.chat_handler is None: + self.chat_format = "llama-2" + if self.verbose: + print( + f"Using fallback chat format: {self.chat_format}", file=sys.stderr + ) + + self._sampler = None + @property def ctx(self) -> llama_cpp.llama_context_p: - assert self._ctx.ctx is not None return self._ctx.ctx @property def model(self) -> llama_cpp.llama_model_p: - assert self._model.model is not None return self._model.model @property @@ -1009,6 +578,8 @@ def tokenize( Args: text: The utf-8 encoded string to tokenize. + add_bos: Whether to add a beginning of sequence token. + special: Whether to tokenize special tokens. Raises: RuntimeError: If the tokenization failed. @@ -1016,18 +587,27 @@ def tokenize( Returns: A list of tokens. """ - return self._model.tokenize(text, add_bos, special) + return self.tokenizer_.tokenize(text, add_bos, special) - def detokenize(self, tokens: List[int]) -> bytes: + def detokenize( + self, + tokens: List[int], + prev_tokens: Optional[List[int]] = None, + special: bool = False, + ) -> bytes: """Detokenize a list of tokens. Args: tokens: The list of tokens to detokenize. + prev_tokens: The list of previous tokens. Offset mapping will be performed if provided. + special: Whether to detokenize special tokens. Returns: The detokenized string. """ - return self._model.detokenize(tokens) + return self.tokenizer_.detokenize( + tokens, prev_tokens=prev_tokens, special=special + ) def set_cache(self, cache: Optional[BaseLlamaCache]): """Set the cache. @@ -1043,8 +623,7 @@ def set_seed(self, seed: int): Args: seed: The random seed. """ - assert self._ctx.ctx is not None - llama_cpp.llama_set_rng_seed(self._ctx.ctx, seed) + self._seed = seed def reset(self): """Reset the model state.""" @@ -1056,8 +635,6 @@ def eval(self, tokens: Sequence[int]): Args: tokens: The list of tokens to evaluate. """ - assert self._ctx.ctx is not None - assert self._batch.batch is not None self._ctx.kv_cache_seq_rm(-1, self.n_tokens, -1) for i in range(0, len(tokens), self.n_batch): batch = tokens[i : min(len(tokens), i + self.n_batch)] @@ -1070,17 +647,114 @@ def eval(self, tokens: Sequence[int]): # Save tokens self.input_ids[n_past : n_past + n_tokens] = batch # Save logits - rows = n_tokens - cols = self._n_vocab - offset = ( - 0 if self.context_params.logits_all else n_tokens - 1 - ) # NOTE: Only save the last token logits if logits_all is False - self.scores[n_past + offset : n_past + n_tokens, :].reshape(-1)[ - : - ] = self._ctx.get_logits()[offset * cols : rows * cols] + if self.context_params.logits_all: + rows = n_tokens + cols = self._n_vocab + logits = np.ctypeslib.as_array( + self._ctx.get_logits(), shape=(rows * cols,) + ) + self.scores[n_past : n_past + n_tokens, :].reshape(-1)[::] = logits + else: + # rows = 1 + # cols = self._n_vocab + # logits = np.ctypeslib.as_array( + # self._ctx.get_logits(), shape=(rows * cols,) + # ) + # self.scores[n_past + n_tokens - 1, :].reshape(-1)[::] = logits + # NOTE: Now that sampling is done inside the sampler, logits are only needed for logprobs which requires logits_all + pass # Update n_tokens self.n_tokens += n_tokens + def _init_sampler( + self, + top_k: int = 40, + top_p: float = 0.95, + min_p: float = 0.05, + typical_p: float = 1.0, + temp: float = 0.80, + repeat_penalty: float = 1.0, + frequency_penalty: float = 0.0, + presence_penalty: float = 0.0, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_eta: float = 0.1, + mirostat_tau: float = 5.0, + penalize_nl: bool = True, + logits_processor: Optional[LogitsProcessorList] = None, + grammar: Optional[LlamaGrammar] = None, + ): + sampler = internals.LlamaSampler() + + if logits_processor is not None: + # Create and add a custom sampler + def apply_func(token_data_array: llama_cpp.llama_token_data_array_p): + size = token_data_array.contents.size + data_soa = token_data_array.contents.data + data_soa_address = ctypes.addressof(data_soa.contents) + # NOTE: This is probably broken + recarray = np.recarray( + shape=(size,), + dtype=np.dtype( + [("id", np.intc), ("logit", np.single), ("p", np.single)], + align=True, + ), + buf=(llama_cpp.llama_token_data * size).from_address( + data_soa_address + ), + ) + for logit_processor in logits_processor: + recarray.logit[:] = logit_processor(self._input_ids, recarray.logit) + + sampler.add_custom(apply_func) + + sampler.add_penalties( + n_vocab=self._n_vocab, + special_eos_id=self._token_eos, + linefeed_id=self._token_nl, + penalty_last_n=self.last_n_tokens_size, + penalty_repeat=repeat_penalty, + penalty_freq=frequency_penalty, + penalty_present=presence_penalty, + penalize_nl=penalize_nl, + ignore_eos=False, + ) + + if grammar is not None: + sampler.add_grammar(self._model, grammar) + + if temp < 0.0: + sampler.add_softmax() + sampler.add_dist(self._seed) + elif temp == 0.0: + sampler.add_greedy() + else: + if mirostat_mode == 1: + mirostat_m = 100 + sampler.add_mirostat( + self._n_vocab, + self._seed, + mirostat_tau, + mirostat_eta, + mirostat_m, + ) + elif mirostat_mode == 2: + sampler.add_mirostat_v2( + self._seed, + mirostat_tau, + mirostat_eta, + ) + else: + n_probs = 0 + min_keep = max(1, n_probs) + sampler.add_top_k(top_k) + sampler.add_typical(typical_p, min_keep) + sampler.add_top_p(top_p, min_keep) + sampler.add_min_p(min_p, min_keep) + sampler.add_temp(temp) + sampler.add_dist(self._seed) + return sampler + def sample( self, top_k: int = 40, @@ -1088,7 +762,7 @@ def sample( min_p: float = 0.05, typical_p: float = 1.0, temp: float = 0.80, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, frequency_penalty: float = 0.0, presence_penalty: float = 0.0, tfs_z: float = 1.0, @@ -1098,6 +772,7 @@ def sample( penalize_nl: bool = True, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, + idx: Optional[int] = None, ): """Sample a token from the model. @@ -1110,80 +785,37 @@ def sample( Returns: The sampled token. """ - assert self._ctx is not None assert self.n_tokens > 0 - last_n_tokens_data = [llama_cpp.llama_token(0)] * max( - 0, self.last_n_tokens_size - self.n_tokens - ) + self._input_ids[-self.last_n_tokens_size :].tolist() - last_n_tokens_size = len(last_n_tokens_data) - n_vocab = self._n_vocab - n_ctx = self._n_ctx - top_k = n_vocab if top_k <= 0 else top_k - last_n_tokens_size = n_ctx if last_n_tokens_size < 0 else last_n_tokens_size - last_n_tokens_data_c = (llama_cpp.llama_token * last_n_tokens_size)( - *last_n_tokens_data - ) - logits: npt.NDArray[np.single] = self._scores[-1, :] - if logits_processor is not None: - logits[:] = logits_processor(self._input_ids, logits) - - nl_logit = logits[self._token_nl] - self._candidates.copy_logits(logits) - self._ctx.sample_repetition_penalties( - candidates=self._candidates, - last_tokens_data=last_n_tokens_data_c, - penalty_last_n=last_n_tokens_size, - penalty_repeat=repeat_penalty, - penalty_freq=frequency_penalty, - penalty_present=presence_penalty, - ) - if not penalize_nl: - self._candidates.candidates.data[self._token_nl].logit = llama_cpp.c_float( - nl_logit - ) + tmp_sampler = False - if grammar is not None: - self._ctx.sample_grammar( - candidates=self._candidates, + if self._sampler is None: + tmp_sampler = True + self._sampler = self._init_sampler( + top_k=top_k, + top_p=top_p, + min_p=min_p, + typical_p=typical_p, + temp=temp, + repeat_penalty=repeat_penalty, + frequency_penalty=frequency_penalty, + presence_penalty=presence_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + penalize_nl=penalize_nl, + logits_processor=logits_processor, grammar=grammar, ) - if temp < 0.0: - self._ctx.sample_softmax(candidates=self._candidates) - id = self._candidates.candidates.data[0].id - elif temp == 0.0: - id = self._ctx.sample_token_greedy(candidates=self._candidates) - elif mirostat_mode == 1: - self._ctx.sample_temp(candidates=self._candidates, temp=temp) - id = self._ctx.sample_token_mirostat( - candidates=self._candidates, - tau=mirostat_tau, - eta=mirostat_eta, - mu=2.0 * mirostat_tau, - m=100, - ) - elif mirostat_mode == 2: - self._ctx.sample_temp(candidates=self._candidates, temp=temp) - id = self._ctx.sample_token_mirostat_v2( - candidates=self._candidates, - tau=mirostat_tau, - eta=mirostat_eta, - mu=2.0 * mirostat_tau, - ) - else: - self._ctx.sample_top_k(candidates=self._candidates, k=top_k, min_keep=1) - self._ctx.sample_tail_free(candidates=self._candidates, z=tfs_z, min_keep=1) - self._ctx.sample_typical( - candidates=self._candidates, p=typical_p, min_keep=1 - ) - self._ctx.sample_top_p(candidates=self._candidates, p=top_p, min_keep=1) - self._ctx.sample_min_p(candidates=self._candidates, p=min_p, min_keep=1) - self._ctx.sample_temp(candidates=self._candidates, temp=temp) - id = self._ctx.sample_token(candidates=self._candidates) - if grammar is not None: - self._ctx.grammar_accept_token(grammar=grammar, token=id) - return id + ridx = idx - self.n_tokens if idx is not None else -1 + + assert self.ctx is not None + token = self._sampler.sample(self._ctx, ridx) + if tmp_sampler: + self._sampler = None + return token def generate( self, @@ -1193,7 +825,7 @@ def generate( min_p: float = 0.05, typical_p: float = 1.0, temp: float = 0.80, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, reset: bool = True, frequency_penalty: float = 0.0, presence_penalty: float = 0.0, @@ -1201,6 +833,7 @@ def generate( mirostat_mode: int = 0, mirostat_tau: float = 5.0, mirostat_eta: float = 0.1, + penalize_nl: bool = True, logits_processor: Optional[LogitsProcessorList] = None, stopping_criteria: Optional[StoppingCriteriaList] = None, grammar: Optional[LlamaGrammar] = None, @@ -1210,7 +843,7 @@ def generate( Examples: >>> llama = Llama("models/ggml-7b.bin") >>> tokens = llama.tokenize(b"Hello, world!") - >>> for token in llama.generate(tokens, top_k=40, top_p=0.95, temp=1.0, repeat_penalty=1.1): + >>> for token in llama.generate(tokens, top_k=40, top_p=0.95, temp=1.0, repeat_penalty=1.0): ... print(llama.detokenize([token])) Args: @@ -1224,6 +857,27 @@ def generate( Yields: The generated tokens. """ + # Reset mirostat sampling + self._mirostat_mu = ctypes.c_float(2.0 * mirostat_tau) + self._sampler = self._init_sampler( + top_k=top_k, + top_p=top_p, + min_p=min_p, + typical_p=typical_p, + temp=temp, + repeat_penalty=repeat_penalty, + frequency_penalty=frequency_penalty, + presence_penalty=presence_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + penalize_nl=penalize_nl, + logits_processor=logits_processor, + grammar=grammar, + ) + + # Check for kv cache prefix match if reset and self.n_tokens > 0: longest_prefix = 0 for a, b in zip(self._input_ids, tokens[:-1]): @@ -1232,44 +886,76 @@ def generate( else: break if longest_prefix > 0: - if self.verbose: - print("Llama.generate: prefix-match hit", file=sys.stderr) reset = False tokens = tokens[longest_prefix:] self.n_tokens = longest_prefix + if self.verbose: + print( + f"Llama.generate: {longest_prefix} prefix-match hit, " + f"remaining {len(tokens)} prompt tokens to eval", + file=sys.stderr, + ) + # Reset the model state if reset: self.reset() - if grammar is not None: - grammar.reset() + # # Reset the grammar + # if grammar is not None: + # grammar.reset() + sample_idx = self.n_tokens + len(tokens) - 1 + tokens = list(tokens) + + # Eval and sample while True: self.eval(tokens) - token = self.sample( - top_k=top_k, - top_p=top_p, - min_p=min_p, - typical_p=typical_p, - temp=temp, - repeat_penalty=repeat_penalty, - frequency_penalty=frequency_penalty, - presence_penalty=presence_penalty, - tfs_z=tfs_z, - mirostat_mode=mirostat_mode, - mirostat_tau=mirostat_tau, - mirostat_eta=mirostat_eta, - logits_processor=logits_processor, - grammar=grammar, - ) - if stopping_criteria is not None and stopping_criteria( - self._input_ids, self._scores[-1, :] - ): - return - tokens_or_none = yield token - tokens = [token] - if tokens_or_none is not None: - tokens.extend(tokens_or_none) + while sample_idx < self.n_tokens: + token = self.sample( + top_k=top_k, + top_p=top_p, + min_p=min_p, + typical_p=typical_p, + temp=temp, + repeat_penalty=repeat_penalty, + frequency_penalty=frequency_penalty, + presence_penalty=presence_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + logits_processor=logits_processor, + grammar=grammar, + penalize_nl=penalize_nl, + idx=sample_idx, + ) + + sample_idx += 1 + if stopping_criteria is not None and stopping_criteria( + self._input_ids[: sample_idx], self._scores[sample_idx - self.n_tokens, :] + ): + return + tokens_or_none = yield token + tokens.clear() + tokens.append(token) + if tokens_or_none is not None: + tokens.extend(tokens_or_none) + + if sample_idx < self.n_tokens and token != self._input_ids[sample_idx]: + self.n_tokens = sample_idx + self._ctx.kv_cache_seq_rm(-1, self.n_tokens, -1) + break + + if self.draft_model is not None: + self.input_ids[self.n_tokens : self.n_tokens + len(tokens)] = tokens + draft_tokens = self.draft_model( + self.input_ids[: self.n_tokens + len(tokens)] + ) + tokens.extend( + draft_tokens.astype(int)[ + : self._n_ctx - self.n_tokens - len(tokens) + ] + ) def create_embedding( self, input: Union[str, List[str]], model: Optional[str] = None @@ -1282,44 +968,24 @@ def create_embedding( Returns: An embedding object. """ - assert self._ctx.ctx is not None - assert self._model.model is not None model_name: str = model if model is not None else self.model_path - if self.context_params.embedding == False: - raise RuntimeError( - "Llama model must be created with embedding=True to call this method" - ) - - if self.verbose: - llama_cpp.llama_reset_timings(self._ctx.ctx) - - if isinstance(input, str): - inputs = [input] - else: - inputs = input + input = input if isinstance(input, list) else [input] - data: List[Embedding] = [] - total_tokens = 0 - for index, input in enumerate(inputs): - tokens = self.tokenize(input.encode("utf-8"), special=True) - self.reset() - self.eval(tokens) - n_tokens = len(tokens) - total_tokens += n_tokens - embedding = llama_cpp.llama_get_embeddings(self._ctx.ctx)[ - : llama_cpp.llama_n_embd(self._model.model) - ] + # get numeric embeddings + embeds: Union[List[List[float]], List[List[List[float]]]] + total_tokens: int + embeds, total_tokens = self.embed(input, return_count=True) # type: ignore - data.append( - { - "object": "embedding", - "embedding": embedding, - "index": index, - } - ) - if self.verbose: - llama_cpp.llama_print_timings(self._ctx.ctx) + # convert to CreateEmbeddingResponse + data: List[Embedding] = [ + { + "object": "embedding", + "embedding": emb, + "index": idx, + } + for idx, emb in enumerate(embeds) + ] return { "object": "list", @@ -1331,7 +997,13 @@ def create_embedding( }, } - def embed(self, input: str) -> List[float]: + def embed( + self, + input: Union[str, List[str]], + normalize: bool = False, + truncate: bool = True, + return_count: bool = False, + ): """Embed a string. Args: @@ -1340,7 +1012,111 @@ def embed(self, input: str) -> List[float]: Returns: A list of embeddings """ - return list(map(float, self.create_embedding(input)["data"][0]["embedding"])) + n_embd = self.n_embd() + n_batch = self.n_batch + + # get pooling information + pooling_type = self.pooling_type() + logits_all = pooling_type == llama_cpp.LLAMA_POOLING_TYPE_NONE + + if self.context_params.embeddings is False: + raise RuntimeError( + "Llama model must be created with embedding=True to call this method" + ) + + if self.verbose: + llama_cpp.llama_perf_context_reset(self._ctx.ctx) + + if isinstance(input, str): + inputs = [input] + else: + inputs = input + + # reset batch + self._batch.reset() + + # decode and fetch embeddings + data: Union[List[List[float]], List[List[List[float]]]] = [] + + def decode_batch(seq_sizes: List[int]): + llama_cpp.llama_kv_cache_clear(self._ctx.ctx) + self._ctx.decode(self._batch) + self._batch.reset() + + # store embeddings + if pooling_type == llama_cpp.LLAMA_POOLING_TYPE_NONE: + pos: int = 0 + for i, size in enumerate(seq_sizes): + ptr = llama_cpp.llama_get_embeddings(self._ctx.ctx) + embedding: List[List[float]] = [ + ptr[pos + j * n_embd : pos + (j + 1) * n_embd] + for j in range(size) + ] + if normalize: + embedding = [ + internals.normalize_embedding(e) for e in embedding + ] + data.append(embedding) + pos += size + else: + for i in range(len(seq_sizes)): + ptr = llama_cpp.llama_get_embeddings_seq(self._ctx.ctx, i) + embedding: List[float] = ptr[:n_embd] + if normalize: + embedding = internals.normalize_embedding(embedding) + data.append(embedding) + + # init state + total_tokens = 0 + s_batch = [] + t_batch = 0 + p_batch = 0 + + # accumulate batches and encode + for text in inputs: + tokens = self.tokenize(text.encode("utf-8")) + if truncate: + tokens = tokens[:n_batch] + + n_tokens = len(tokens) + total_tokens += n_tokens + + # check for overrun + if n_tokens > n_batch: + raise ValueError( + f"Requested tokens ({n_tokens}) exceed batch size of {n_batch}" + ) + + # time to eval batch + if t_batch + n_tokens > n_batch: + decode_batch(s_batch) + s_batch = [] + t_batch = 0 + p_batch = 0 + + # add to batch + self._batch.add_sequence(tokens, p_batch, logits_all) + + # update batch stats + s_batch.append(n_tokens) + t_batch += n_tokens + p_batch += 1 + + # hanlde last batch + decode_batch(s_batch) + + if self.verbose: + llama_cpp.llama_perf_context_print(self._ctx.ctx) + + output = data[0] if isinstance(input, str) else data + + llama_cpp.llama_kv_cache_clear(self._ctx.ctx) + self.reset() + + if return_count: + return output, total_tokens + else: + return output def _create_completion( self, @@ -1356,7 +1132,7 @@ def _create_completion( stop: Optional[Union[str, List[str]]] = [], frequency_penalty: float = 0.0, presence_penalty: float = 0.0, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, top_k: int = 40, stream: bool = False, seed: Optional[int] = None, @@ -1368,28 +1144,91 @@ def _create_completion( stopping_criteria: Optional[StoppingCriteriaList] = None, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, + logit_bias: Optional[Dict[int, float]] = None, ) -> Union[ Iterator[CreateCompletionResponse], Iterator[CreateCompletionStreamResponse] ]: - assert self._ctx is not None assert suffix is None or suffix.__class__ is str completion_id: str = f"cmpl-{str(uuid.uuid4())}" created: int = int(time.time()) + bos_token_id: int = self.token_bos() + cls_token_id: int = self._model.token_cls() + sep_token_id: int = self._model.token_sep() + prefix_token_id: int = 0 # self._model.token_prefix() # TODO: Fix + middle_token_id: int = 0 # self._model.token_middle() # TODO: Fix + suffix_token_id: int = 0 # self._model.token_suffix() # TODO: Fix + add_space_prefix: bool = ( + self.metadata.get("tokenizer.ggml.add_space_prefix", "true") == "true" + ) + bos_tokens: List[int] = [cls_token_id if cls_token_id != -1 else bos_token_id] + eos_tokens: List[int] = [ + sep_token_id if sep_token_id != -1 else self.token_eos() + ] + + if ( + (isinstance(prompt, list) and suffix is None) + or not self._model.add_bos_token() + or bos_tokens[:1] == [-1] + ): + bos_tokens = [] + + if (isinstance(prompt, list) and suffix is None) or ( + not self._model.add_eos_token() and sep_token_id == -1 + ): + eos_tokens = [] + + suffix_space_prefix: int = 0 + # Tokenizer hack to remove leading space + if add_space_prefix and suffix_token_id >= 0 and suffix: + suffix = "☺" + suffix + suffix_space_prefix = 2 + # If prompt is empty, initialize completion with BOS token to avoid # detokenization including a space at the beginning of the completion - completion_tokens: List[int] = [] if len(prompt) > 0 else [self.token_bos()] + completion_tokens: List[int] = [] if len(prompt) > 0 else [bos_token_id] # Add blank space to start of prompt to match OG llama tokenizer - prompt_tokens: List[int] = ( + prefix_tokens: List[int] = ( + [prefix_token_id] if prefix_token_id >= 0 and suffix is not None else [] + ) + ( ( - self.tokenize(prompt.encode("utf-8"), special=True) + self.tokenize( + prompt.encode("utf-8"), + add_bos=False, + special=(prefix_token_id < 0 or suffix is None), + ) if prompt != "" - else [self.token_bos()] + else [] ) if isinstance(prompt, str) else prompt ) + suffix_tokens: List[int] = ( + ( + [suffix_token_id] + + ( + self.tokenize(suffix.encode("utf-8"), add_bos=False, special=False)[ + suffix_space_prefix: + ] + if suffix + else [] + ) + ) + if suffix_token_id >= 0 and suffix is not None + else [] + ) + middle_tokens: List[int] = ( + [middle_token_id] if middle_token_id >= 0 and suffix is not None else [] + ) + prompt_tokens: List[int] = ( + bos_tokens + + ( + (suffix_tokens + prefix_tokens + middle_tokens) + if self.spm_infill + else (prefix_tokens + suffix_tokens + middle_tokens) + ) + + eos_tokens + ) text: bytes = b"" returned_tokens: int = 0 stop = ( @@ -1397,6 +1236,12 @@ def _create_completion( ) model_name: str = model if model is not None else self.model_path + if prompt_tokens[:2] == [self.token_bos()] * 2: + warnings.warn( + f'Detected duplicate leading "{self._model.token_get_text(self.token_bos())}" in prompt, this will likely reduce response quality, consider removing it...', + RuntimeWarning, + ) + # NOTE: This likely doesn't work correctly for the first token in the prompt # because of the extra space added to the start of the prompt_tokens if logit_bias is not None: @@ -1466,7 +1311,9 @@ def logit_bias_processor( print("Llama._create_completion: cache miss", file=sys.stderr) if seed is not None: - self._ctx.set_rng_seed(seed) + self.set_seed(seed) + else: + self.set_seed(random.Random(self._seed).randint(0, 2 ** 32)) finish_reason = "length" multibyte_fix = 0 @@ -1488,14 +1335,14 @@ def logit_bias_processor( logits_processor=logits_processor, grammar=grammar, ): - if token == self._token_eos: - text = self.detokenize(completion_tokens) + if llama_cpp.llama_token_is_eog(self._model.vocab, token): + text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "stop" break completion_tokens.append(token) - all_text = self.detokenize(completion_tokens) + all_text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) # Contains multi-byte UTF8 for k, char in enumerate(all_text[-3:]): @@ -1519,7 +1366,10 @@ def logit_bias_processor( if stream: remaining_tokens = completion_tokens[returned_tokens:] - remaining_text = self.detokenize(remaining_tokens) + remaining_text = self.detokenize( + remaining_tokens, + prev_tokens=prompt_tokens + completion_tokens[:returned_tokens], + ) remaining_length = len(remaining_text) # We want to avoid yielding any characters from @@ -1539,19 +1389,31 @@ def logit_bias_processor( # not sure how to handle this branch when dealing # with CJK output, so keep it unchanged for token in remaining_tokens: - if token == self.token_bos(): + if token == bos_token_id: continue - token_end_position += len(self.detokenize([token])) + token_end_position += len( + self.detokenize( + [token], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ) + ) # Check if stop sequence is in the token if token_end_position > ( remaining_length - first_stop_position ): break - token_str = self.detokenize([token]).decode( - "utf-8", errors="ignore" - ) + token_str = self.detokenize( + [token], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ).decode("utf-8", errors="ignore") text_offset = len(prompt) + len( - self.detokenize(completion_tokens[:returned_tokens]) + self.detokenize( + completion_tokens[:returned_tokens], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ).decode("utf-8", errors="ignore") ) token_offset = len(prompt_tokens) + returned_tokens logits = self._scores[token_offset - 1, :] @@ -1571,9 +1433,11 @@ def logit_bias_processor( top_logprob.update({token_str: current_logprobs[int(token)]}) logprobs_or_none = { "tokens": [ - self.detokenize([token]).decode( - "utf-8", errors="ignore" - ) + self.detokenize( + [token], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ).decode("utf-8", errors="ignore") ], "text_offset": [text_offset], "token_logprobs": [current_logprobs[int(token)]], @@ -1587,9 +1451,11 @@ def logit_bias_processor( "model": model_name, "choices": [ { - "text": self.detokenize([token]).decode( - "utf-8", errors="ignore" - ), + "text": self.detokenize( + [token], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ).decode("utf-8", errors="ignore"), "index": 0, "logprobs": logprobs_or_none, "finish_reason": None, @@ -1601,7 +1467,11 @@ def logit_bias_processor( decode_success = False for i in range(1, len(remaining_tokens) + 1): try: - bs = self.detokenize(remaining_tokens[:i]) + bs = self.detokenize( + remaining_tokens[:i], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ) ts = bs.decode("utf-8") decode_success = True break @@ -1636,14 +1506,14 @@ def logit_bias_processor( } if len(completion_tokens) >= max_tokens: - text = self.detokenize(completion_tokens) + text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "length" break if stopping_criteria is not None and stopping_criteria( self._input_ids, self._scores[-1, :] ): - text = self.detokenize(completion_tokens) + text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "stop" if self.verbose: @@ -1651,26 +1521,38 @@ def logit_bias_processor( if stream: remaining_tokens = completion_tokens[returned_tokens:] - all_text = self.detokenize(remaining_tokens) - any_stop = [s for s in stop_sequences if s in all_text] + remaining_text = self.detokenize( + remaining_tokens, + prev_tokens=prompt_tokens + completion_tokens[:returned_tokens], + ) + any_stop = [s for s in stop_sequences if s in remaining_text] if len(any_stop) > 0: - end = min(all_text.index(stop) for stop in any_stop) + end = min(remaining_text.index(stop) for stop in any_stop) else: - end = len(all_text) + end = len(remaining_text) token_end_position = 0 for token in remaining_tokens: - token_end_position += len(self.detokenize([token])) + token_end_position += len( + self.detokenize( + [token], + prev_tokens=prompt_tokens + completion_tokens[:returned_tokens], + ) + ) logprobs_or_none: Optional[CompletionLogprobs] = None if logprobs is not None: - if token == self.token_bos(): + if token == bos_token_id: continue token_str = self.detokenize([token]).decode( "utf-8", errors="ignore" ) text_offset = len(prompt) + len( - self.detokenize(completion_tokens[:returned_tokens]) + self.detokenize( + completion_tokens[:returned_tokens], + prev_tokens=prompt_tokens + + completion_tokens[:returned_tokens], + ) ) token_offset = len(prompt_tokens) + returned_tokens - 1 logits = self._scores[token_offset, :] @@ -1752,7 +1634,8 @@ def logit_bias_processor( if self.verbose: print("Llama._create_completion: cache save", file=sys.stderr) self.cache[prompt_tokens + completion_tokens] = self.save_state() - print("Llama._create_completion: cache saved", file=sys.stderr) + if self.verbose: + print("Llama._create_completion: cache saved", file=sys.stderr) return if self.cache: @@ -1765,7 +1648,7 @@ def logit_bias_processor( if echo: text_str = prompt + text_str - if suffix is not None: + if suffix_token_id < 0 and suffix is not None: text_str = text_str + suffix logprobs_or_none: Optional[CompletionLogprobs] = None @@ -1778,24 +1661,35 @@ def logit_bias_processor( top_logprobs: List[Optional[Dict[str, float]]] = [] if echo: - # Remove leading BOS token - all_tokens = prompt_tokens[1:] + completion_tokens + # Remove leading BOS token if exists + all_tokens = ( + prompt_tokens[1 if prompt_tokens[0] == self.token_bos() else 0 :] + + completion_tokens + ) else: all_tokens = completion_tokens all_token_strs = [ - self.detokenize([token]).decode("utf-8", errors="ignore") - for token in all_tokens + self.detokenize([token], prev_tokens=all_tokens[:i]).decode( + "utf-8", errors="ignore" + ) + for i, token in enumerate(all_tokens) ] all_logprobs = Llama.logits_to_logprobs(self._scores)[token_offset:] # TODO: may be able to change this loop to use np.take_along_dim - for token, token_str, logprobs_token in zip( - all_tokens, all_token_strs, all_logprobs + for idx, (token, token_str, logprobs_token) in enumerate( + zip(all_tokens, all_token_strs, all_logprobs) ): - if token == self.token_bos(): + if token == bos_token_id: continue - text_offsets.append(text_offset) - text_offset += len(token_str) + text_offsets.append( + text_offset + + len( + self.detokenize(all_tokens[:idx]).decode( + "utf-8", errors="ignore" + ) + ) + ) tokens.append(token_str) sorted_logprobs = list( sorted( @@ -1804,7 +1698,9 @@ def logit_bias_processor( ) token_logprobs.append(logprobs_token[int(token)]) top_logprob: Optional[Dict[str, float]] = { - self.detokenize([i]).decode("utf-8", errors="ignore"): logprob + self.detokenize([i], prev_tokens=all_tokens[:idx]).decode( + "utf-8", errors="ignore" + ): logprob for logprob, i in sorted_logprobs[:logprobs] } top_logprob.update({token_str: logprobs_token[int(token)]}) @@ -1856,7 +1752,7 @@ def create_completion( stop: Optional[Union[str, List[str]]] = [], frequency_penalty: float = 0.0, presence_penalty: float = 0.0, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, top_k: int = 40, stream: bool = False, seed: Optional[int] = None, @@ -1868,7 +1764,7 @@ def create_completion( stopping_criteria: Optional[StoppingCriteriaList] = None, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, + logit_bias: Optional[Dict[int, float]] = None, ) -> Union[CreateCompletionResponse, Iterator[CreateCompletionStreamResponse]]: """Generate text from a prompt. @@ -1909,7 +1805,7 @@ def create_completion( completion_or_chunks = self._create_completion( prompt=prompt, suffix=suffix, - max_tokens=max_tokens, + max_tokens=-1 if max_tokens is None else max_tokens, temperature=temperature, top_p=top_p, min_p=min_p, @@ -1943,7 +1839,7 @@ def __call__( self, prompt: str, suffix: Optional[str] = None, - max_tokens: int = 128, + max_tokens: Optional[int] = 16, temperature: float = 0.8, top_p: float = 0.95, min_p: float = 0.05, @@ -1953,7 +1849,7 @@ def __call__( stop: Optional[Union[str, List[str]]] = [], frequency_penalty: float = 0.0, presence_penalty: float = 0.0, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, top_k: int = 40, stream: bool = False, seed: Optional[int] = None, @@ -1965,7 +1861,7 @@ def __call__( stopping_criteria: Optional[StoppingCriteriaList] = None, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, + logit_bias: Optional[Dict[int, float]] = None, ) -> Union[CreateCompletionResponse, Iterator[CreateCompletionStreamResponse]]: """Generate text from a prompt. @@ -2050,7 +1946,7 @@ def create_chat_completion( max_tokens: Optional[int] = None, presence_penalty: float = 0.0, frequency_penalty: float = 0.0, - repeat_penalty: float = 1.1, + repeat_penalty: float = 1.0, tfs_z: float = 1.0, mirostat_mode: int = 0, mirostat_tau: float = 5.0, @@ -2058,7 +1954,9 @@ def create_chat_completion( model: Optional[str] = None, logits_processor: Optional[LogitsProcessorList] = None, grammar: Optional[LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, + logit_bias: Optional[Dict[int, float]] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, ) -> Union[ CreateChatCompletionResponse, Iterator[CreateChatCompletionStreamResponse] ]: @@ -2095,8 +1993,10 @@ def create_chat_completion( Returns: Generated chat completion or a stream of chat completion chunks. """ - handler = self.chat_handler or llama_chat_format.get_chat_completion_handler( - self.chat_format + handler = ( + self.chat_handler + or self._chat_handlers.get(self.chat_format) + or llama_chat_format.get_chat_completion_handler(self.chat_format) ) return handler( llama=self, @@ -2110,6 +2010,8 @@ def create_chat_completion( top_k=top_k, min_p=min_p, typical_p=typical_p, + logprobs=logprobs, + top_logprobs=top_logprobs, stream=stream, stop=stop, seed=seed, @@ -2128,23 +2030,60 @@ def create_chat_completion( logit_bias=logit_bias, ) + def create_chat_completion_openai_v1( + self, + *args: Any, + **kwargs: Any, + ): + """Generate a chat completion with return type based on the the OpenAI v1 API. + + OpenAI python package is required to use this method. + + You can install it with `pip install openai`. + + Args: + *args: Positional arguments to pass to create_chat_completion. + **kwargs: Keyword arguments to pass to create_chat_completion. + + Returns: + Generated chat completion or a stream of chat completion chunks. + """ + try: + from openai.types.chat import ChatCompletion, ChatCompletionChunk + + stream = kwargs.get("stream", False) # type: ignore + assert isinstance(stream, bool) + if stream: + return (ChatCompletionChunk(**chunk) for chunk in self.create_chat_completion(*args, **kwargs)) # type: ignore + else: + return ChatCompletion(**self.create_chat_completion(*args, **kwargs)) # type: ignore + except ImportError: + raise ImportError( + "To use create_chat_completion_openai_v1, you must install the openai package." + "You can install it with `pip install openai`." + ) + def __getstate__(self): return dict( model_path=self.model_path, # Model Params n_gpu_layers=self.model_params.n_gpu_layers, + split_mode=self.model_params.split_mode, main_gpu=self.model_params.main_gpu, tensor_split=self.tensor_split, vocab_only=self.model_params.vocab_only, use_mmap=self.model_params.use_mmap, use_mlock=self.model_params.use_mlock, + kv_overrides=self.kv_overrides, # Context Params - seed=self.context_params.seed, + seed=self._seed, n_ctx=self.context_params.n_ctx, n_batch=self.n_batch, + n_ubatch=self.context_params.n_ubatch, n_threads=self.context_params.n_threads, n_threads_batch=self.context_params.n_threads_batch, rope_scaling_type=self.context_params.rope_scaling_type, + pooling_type=self.context_params.pooling_type, rope_freq_base=self.context_params.rope_freq_base, rope_freq_scale=self.context_params.rope_freq_scale, yarn_ext_factor=self.context_params.yarn_ext_factor, @@ -2152,10 +2091,12 @@ def __getstate__(self): yarn_beta_fast=self.context_params.yarn_beta_fast, yarn_beta_slow=self.context_params.yarn_beta_slow, yarn_orig_ctx=self.context_params.yarn_orig_ctx, - mul_mat_q=self.context_params.mul_mat_q, logits_all=self.context_params.logits_all, - embedding=self.context_params.embedding, + embedding=self.context_params.embeddings, + offload_kqv=self.context_params.offload_kqv, + flash_attn=self.context_params.flash_attn, # Sampling Params + no_perf=self.context_params.no_perf, last_n_tokens_size=self.last_n_tokens_size, # LoRA Params lora_base=self.lora_base, @@ -2166,59 +2107,26 @@ def __getstate__(self): # Chat Format Params chat_format=self.chat_format, chat_handler=self.chat_handler, + # Speculative Decidng + draft_model=self.draft_model, + # KV cache quantization + type_k=self.context_params.type_k, + type_v=self.context_params.type_v, # Misc + spm_infill=self.spm_infill, verbose=self.verbose, ) def __setstate__(self, state): - self.__init__( - model_path=state["model_path"], - # Model Params - n_gpu_layers=state["n_gpu_layers"], - main_gpu=state["main_gpu"], - tensor_split=state["tensor_split"], - vocab_only=state["vocab_only"], - use_mmap=state["use_mmap"], - use_mlock=state["use_mlock"], - # Context Params - seed=state["seed"], - n_ctx=state["n_ctx"], - n_batch=state["n_batch"], - n_threads=state["n_threads"], - n_threads_batch=state["n_threads_batch"], - rope_freq_base=state["rope_freq_base"], - rope_freq_scale=state["rope_freq_scale"], - rope_scaling_type=state["rope_scaling_type"], - yarn_ext_factor=state["yarn_ext_factor"], - yarn_attn_factor=state["yarn_attn_factor"], - yarn_beta_fast=state["yarn_beta_fast"], - yarn_beta_slow=state["yarn_beta_slow"], - yarn_orig_ctx=state["yarn_orig_ctx"], - mul_mat_q=state["mul_mat_q"], - logits_all=state["logits_all"], - embedding=state["embedding"], - # Sampling Params - last_n_tokens_size=state["last_n_tokens_size"], - # LoRA Params - lora_base=state["lora_base"], - lora_path=state["lora_path"], - # Backend Params - numa=state["numa"], - # Chat Format Params - chat_format=state["chat_format"], - chat_handler=state["chat_handler"], - # Misc - verbose=state["verbose"], - ) + self.__init__(**state) def save_state(self) -> LlamaState: - assert self._ctx.ctx is not None if self.verbose: print("Llama.save_state: saving llama state", file=sys.stderr) state_size = llama_cpp.llama_get_state_size(self._ctx.ctx) if self.verbose: print(f"Llama.save_state: got state size: {state_size}", file=sys.stderr) - llama_state = (llama_cpp.c_uint8 * int(state_size))() + llama_state = (ctypes.c_uint8 * int(state_size))() if self.verbose: print("Llama.save_state: allocated state", file=sys.stderr) n_bytes = llama_cpp.llama_copy_state_data(self._ctx.ctx, llama_state) @@ -2226,7 +2134,7 @@ def save_state(self) -> LlamaState: print(f"Llama.save_state: copied llama state: {n_bytes}", file=sys.stderr) if int(n_bytes) > int(state_size): raise RuntimeError("Failed to copy llama state data") - llama_state_compact = (llama_cpp.c_uint8 * int(n_bytes))() + llama_state_compact = (ctypes.c_uint8 * int(n_bytes))() llama_cpp.ctypes.memmove(llama_state_compact, llama_state, int(n_bytes)) if self.verbose: print( @@ -2234,20 +2142,24 @@ def save_state(self) -> LlamaState: file=sys.stderr, ) return LlamaState( - scores=self.scores.copy(), + scores=self._scores.copy(), input_ids=self.input_ids.copy(), n_tokens=self.n_tokens, llama_state=bytes(llama_state_compact), llama_state_size=n_bytes, + seed=self._seed, ) def load_state(self, state: LlamaState) -> None: - assert self._ctx.ctx is not None - self.scores = state.scores.copy() + # Only filling in up to `n_tokens` and then zero-ing out the rest + self.scores[: state.n_tokens, :] = state.scores.copy() + rest = self.scores[state.n_tokens :, :] + rest[rest > 0] = 0.0 self.input_ids = state.input_ids.copy() self.n_tokens = state.n_tokens + self._seed = state.seed state_size = state.llama_state_size - LLamaStateArrayType = llama_cpp.c_uint8 * state_size + LLamaStateArrayType = ctypes.c_uint8 * state_size llama_state = LLamaStateArrayType.from_buffer_copy(state.llama_state) if llama_cpp.llama_set_state_data(self._ctx.ctx, llama_state) != state_size: @@ -2265,8 +2177,8 @@ def n_vocab(self) -> int: """Return the vocabulary size.""" return self._model.n_vocab() - def tokenizer(self) -> "LlamaTokenizer": - """Return the tokenizer for this model.""" + def tokenizer(self) -> LlamaTokenizer: + """Return the llama tokenizer for this model.""" return LlamaTokenizer(self) def token_eos(self) -> int: @@ -2281,6 +2193,17 @@ def token_nl(self) -> int: """Return the newline token.""" return self._model.token_nl() + def pooling_type(self) -> str: + """Return the pooling type.""" + return self._ctx.pooling_type() + + def close(self) -> None: + """Explicitly free the model from memory.""" + self._stack.close() + + def __del__(self) -> None: + self.close() + @staticmethod def logits_to_logprobs( logits: Union[npt.NDArray[np.single], List], axis: int = -1 @@ -2309,19 +2232,187 @@ def longest_token_prefix(a: Sequence[int], b: Sequence[int]): break return longest_prefix + @classmethod + def from_pretrained( + cls, + repo_id: str, + filename: Optional[str], + additional_files: Optional[List] = None, + local_dir: Optional[Union[str, os.PathLike[str]]] = None, + local_dir_use_symlinks: Union[bool, Literal["auto"]] = "auto", + cache_dir: Optional[Union[str, os.PathLike[str]]] = None, + **kwargs: Any, + ) -> "Llama": + """Create a Llama model from a pretrained model name or path. + This method requires the huggingface-hub package. + You can install it with `pip install huggingface-hub`. + + Args: + repo_id: The model repo id. + filename: A filename or glob pattern to match the model file in the repo. + additional_files: A list of filenames or glob patterns to match additional model files in the repo. + local_dir: The local directory to save the model to. + local_dir_use_symlinks: Whether to use symlinks when downloading the model. + **kwargs: Additional keyword arguments to pass to the Llama constructor. + + Returns: + A Llama model.""" + try: + from huggingface_hub import hf_hub_download, HfFileSystem + from huggingface_hub.utils import validate_repo_id + except ImportError: + raise ImportError( + "Llama.from_pretrained requires the huggingface-hub package. " + "You can install it with `pip install huggingface-hub`." + ) + + validate_repo_id(repo_id) + + hffs = HfFileSystem() + + files = [ + file["name"] if isinstance(file, dict) else file + for file in hffs.ls(repo_id, recursive=True) + ] + + # split each file into repo_id, subfolder, filename + file_list: List[str] = [] + for file in files: + rel_path = Path(file).relative_to(repo_id) + file_list.append(str(rel_path)) + + # find the only/first shard file: + matching_files = [file for file in file_list if fnmatch.fnmatch(file, filename)] # type: ignore + + if len(matching_files) == 0: + raise ValueError( + f"No file found in {repo_id} that match {filename}\n\n" + f"Available Files:\n{json.dumps(file_list)}" + ) + + if len(matching_files) > 1: + raise ValueError( + f"Multiple files found in {repo_id} matching {filename}\n\n" + f"Available Files:\n{json.dumps(files)}" + ) + + (matching_file,) = matching_files -class LlamaTokenizer: - def __init__(self, llama: Llama): - self.llama = llama + subfolder = str(Path(matching_file).parent) + filename = Path(matching_file).name - def encode(self, text: str, add_bos: bool = True) -> List[int]: - return self.llama.tokenize( - text.encode("utf-8", errors="ignore"), add_bos=add_bos, special=True + # download the file + hf_hub_download( + repo_id=repo_id, + filename=filename, + subfolder=subfolder, + local_dir=local_dir, + local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cache_dir, ) - def decode(self, tokens: List[int]) -> str: - return self.llama.detokenize(tokens).decode("utf-8", errors="ignore") + if additional_files: + for additonal_file_name in additional_files: + # find the additional shard file: + matching_additional_files = [file for file in file_list if fnmatch.fnmatch(file, additonal_file_name)] - @classmethod - def from_ggml_file(cls, path: str) -> "LlamaTokenizer": - return cls(Llama(model_path=path, vocab_only=True)) + if len(matching_additional_files) == 0: + raise ValueError( + f"No file found in {repo_id} that match {additonal_file_name}\n\n" + f"Available Files:\n{json.dumps(file_list)}" + ) + + if len(matching_additional_files) > 1: + raise ValueError( + f"Multiple files found in {repo_id} matching {additonal_file_name}\n\n" + f"Available Files:\n{json.dumps(files)}" + ) + + (matching_additional_file,) = matching_additional_files + + # download the additional file + hf_hub_download( + repo_id=repo_id, + filename=matching_additional_file, + subfolder=subfolder, + local_dir=local_dir, + local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cache_dir, + ) + + if local_dir is None: + model_path = hf_hub_download( + repo_id=repo_id, + filename=filename, + subfolder=subfolder, + local_dir=local_dir, + local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cache_dir, + local_files_only=True, + ) + else: + model_path = os.path.join(local_dir, filename) + + # loading the first file of a sharded GGUF loads all remaining shard files in the subfolder + return cls( + model_path=model_path, + **kwargs, + ) + + +class LlamaState: + def __init__( + self, + input_ids: npt.NDArray[np.intc], + scores: npt.NDArray[np.single], + n_tokens: int, + llama_state: bytes, + llama_state_size: int, + seed: int, + ): + self.input_ids = input_ids + self.scores = scores + self.n_tokens = n_tokens + self.llama_state = llama_state + self.llama_state_size = llama_state_size + self.seed = seed + + +LogitsProcessor = Callable[ + [npt.NDArray[np.intc], npt.NDArray[np.single]], npt.NDArray[np.single] +] + + +class LogitsProcessorList(List[LogitsProcessor]): + def __call__( + self, input_ids: npt.NDArray[np.intc], scores: npt.NDArray[np.single] + ) -> npt.NDArray[np.single]: + for processor in self: + scores = processor(input_ids, scores) + return scores + + +StoppingCriteria = Callable[[npt.NDArray[np.intc], npt.NDArray[np.single]], bool] + + +class StoppingCriteriaList(List[StoppingCriteria]): + def __call__( + self, input_ids: npt.NDArray[np.intc], logits: npt.NDArray[np.single] + ) -> bool: + return any([stopping_criteria(input_ids, logits) for stopping_criteria in self]) + + +class MinTokensLogitsProcessor(LogitsProcessor): + def __init__(self, min_tokens: int, token_eos: int): + self.min_tokens = min_tokens + self.token_eos = token_eos + self.prompt_tokens = None + + def __call__( + self, input_ids: npt.NDArray[np.intc], scores: npt.NDArray[np.single] + ) -> npt.NDArray[np.single]: + if self.prompt_tokens is None: + self.prompt_tokens = len(input_ids) + if len(input_ids) - self.prompt_tokens < self.min_tokens: + scores[self.token_eos] = -np.inf + return scores diff --git a/llama_cpp/llama_cache.py b/llama_cpp/llama_cache.py new file mode 100644 index 000000000..e059e98e1 --- /dev/null +++ b/llama_cpp/llama_cache.py @@ -0,0 +1,155 @@ +import sys +from abc import ABC, abstractmethod +from typing import ( + Optional, + Sequence, + Tuple, +) +from collections import OrderedDict + +import diskcache + +import llama_cpp.llama + +from .llama_types import * + + +class BaseLlamaCache(ABC): + """Base cache class for a llama.cpp model.""" + + def __init__(self, capacity_bytes: int = (2 << 30)): + self.capacity_bytes = capacity_bytes + + @property + @abstractmethod + def cache_size(self) -> int: + raise NotImplementedError + + def _find_longest_prefix_key( + self, + key: Tuple[int, ...], + ) -> Optional[Tuple[int, ...]]: + pass + + @abstractmethod + def __getitem__(self, key: Sequence[int]) -> "llama_cpp.llama.LlamaState": + raise NotImplementedError + + @abstractmethod + def __contains__(self, key: Sequence[int]) -> bool: + raise NotImplementedError + + @abstractmethod + def __setitem__( + self, key: Sequence[int], value: "llama_cpp.llama.LlamaState" + ) -> None: + raise NotImplementedError + + +class LlamaRAMCache(BaseLlamaCache): + """Cache for a llama.cpp model using RAM.""" + + def __init__(self, capacity_bytes: int = (2 << 30)): + super().__init__(capacity_bytes) + self.capacity_bytes = capacity_bytes + self.cache_state: OrderedDict[ + Tuple[int, ...], "llama_cpp.llama.LlamaState" + ] = OrderedDict() + + @property + def cache_size(self): + return sum([state.llama_state_size for state in self.cache_state.values()]) + + def _find_longest_prefix_key( + self, + key: Tuple[int, ...], + ) -> Optional[Tuple[int, ...]]: + min_len = 0 + min_key = None + keys = ( + (k, llama_cpp.llama.Llama.longest_token_prefix(k, key)) + for k in self.cache_state.keys() + ) + for k, prefix_len in keys: + if prefix_len > min_len: + min_len = prefix_len + min_key = k + return min_key + + def __getitem__(self, key: Sequence[int]) -> "llama_cpp.llama.LlamaState": + key = tuple(key) + _key = self._find_longest_prefix_key(key) + if _key is None: + raise KeyError("Key not found") + value = self.cache_state[_key] + self.cache_state.move_to_end(_key) + return value + + def __contains__(self, key: Sequence[int]) -> bool: + return self._find_longest_prefix_key(tuple(key)) is not None + + def __setitem__(self, key: Sequence[int], value: "llama_cpp.llama.LlamaState"): + key = tuple(key) + if key in self.cache_state: + del self.cache_state[key] + self.cache_state[key] = value + while self.cache_size > self.capacity_bytes and len(self.cache_state) > 0: + self.cache_state.popitem(last=False) + + +# Alias for backwards compatibility +LlamaCache = LlamaRAMCache + + +class LlamaDiskCache(BaseLlamaCache): + """Cache for a llama.cpp model using disk.""" + + def __init__( + self, cache_dir: str = ".cache/llama_cache", capacity_bytes: int = (2 << 30) + ): + super().__init__(capacity_bytes) + self.cache = diskcache.Cache(cache_dir) + + @property + def cache_size(self): + return int(self.cache.volume()) # type: ignore + + def _find_longest_prefix_key( + self, + key: Tuple[int, ...], + ) -> Optional[Tuple[int, ...]]: + min_len = 0 + min_key: Optional[Tuple[int, ...]] = None + for k in self.cache.iterkeys(): # type: ignore + prefix_len = llama_cpp.llama.Llama.longest_token_prefix(k, key) + if prefix_len > min_len: + min_len = prefix_len + min_key = k # type: ignore + return min_key + + def __getitem__(self, key: Sequence[int]) -> "llama_cpp.llama.LlamaState": + key = tuple(key) + _key = self._find_longest_prefix_key(key) + if _key is None: + raise KeyError("Key not found") + value: "llama_cpp.llama.LlamaState" = self.cache.pop(_key) # type: ignore + # NOTE: This puts an integer as key in cache, which breaks, + # Llama.longest_token_prefix(k, key) above since k is not a tuple of ints/tokens + # self.cache.push(_key, side="front") # type: ignore + return value + + def __contains__(self, key: Sequence[int]) -> bool: + return self._find_longest_prefix_key(tuple(key)) is not None + + def __setitem__(self, key: Sequence[int], value: "llama_cpp.llama.LlamaState"): + print("LlamaDiskCache.__setitem__: called", file=sys.stderr) + key = tuple(key) + if key in self.cache: + print("LlamaDiskCache.__setitem__: delete", file=sys.stderr) + del self.cache[key] + self.cache[key] = value + print("LlamaDiskCache.__setitem__: set", file=sys.stderr) + while self.cache_size > self.capacity_bytes and len(self.cache) > 0: + key_to_remove = next(iter(self.cache)) + del self.cache[key_to_remove] + print("LlamaDiskCache.__setitem__: trim", file=sys.stderr) diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 037f96a2d..17575c700 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -1,23 +1,74 @@ from __future__ import annotations import os +import sys import json import ctypes import dataclasses -from typing import Any, Dict, Iterator, List, Optional, Tuple, Union, Protocol +import random +import string + +from contextlib import ExitStack +from typing import ( + Any, + Dict, + Iterator, + List, + Literal, + Optional, + Tuple, + Union, + Protocol, + cast, +) + +import jinja2 +from jinja2.sandbox import ImmutableSandboxedEnvironment + +import numpy as np +import numpy.typing as npt import llama_cpp.llama as llama import llama_cpp.llama_types as llama_types import llama_cpp.llama_grammar as llama_grammar -from ._utils import suppress_stdout_stderr +from ._logger import logger +from ._utils import suppress_stdout_stderr, Singleton + +### Common Chat Templates and Special Tokens ### + +# Source: https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/blob/main/tokenizer_config.json +CHATML_CHAT_TEMPLATE = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}" +CHATML_BOS_TOKEN = "" +CHATML_EOS_TOKEN = "<|im_end|>" + +# Source: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1/blob/main/tokenizer_config.json +MISTRAL_INSTRUCT_CHAT_TEMPLATE = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token + ' ' }}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" +MISTRAL_INSTRUCT_BOS_TOKEN = "" +MISTRAL_INSTRUCT_EOS_TOKEN = "" + +# Source: https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1/blob/main/tokenizer_config.json +MIXTRAL_INSTRUCT_CHAT_TEMPLATE = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" + +# Source: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/blob/main/tokenizer_config.json +LLAMA3_INSTRUCT_CHAT_TEMPLATE = "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}" + +### Chat Completion Handler ### class LlamaChatCompletionHandler(Protocol): + """Base Protocol for a llama chat completion handler. + + Very generic protocol that can be used to implement any chat format. + The only hard requirement is that it must return a ChatCompletion when + stream=False and an iterator of ChatCompletionChunks when stream=True.""" + def __call__( self, *, + # llama.cpp instance llama: llama.Llama, + # openai api parameters messages: List[llama_types.ChatCompletionRequestMessage], functions: Optional[List[llama_types.ChatCompletionFunction]] = None, function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, @@ -26,8 +77,6 @@ def __call__( temperature: float = 0.2, top_p: float = 0.95, top_k: int = 40, - min_p: float = 0.05, - typical_p: float = 1.0, stream: bool = False, stop: Optional[Union[str, List[str]]] = [], seed: Optional[int] = None, @@ -38,165 +87,207 @@ def __call__( presence_penalty: float = 0.0, frequency_penalty: float = 0.0, repeat_penalty: float = 1.1, + model: Optional[str] = None, + logit_bias: Optional[Dict[str, float]] = None, + # llama.cpp parameters + min_p: float = 0.05, + typical_p: float = 1.0, tfs_z: float = 1.0, mirostat_mode: int = 0, mirostat_tau: float = 5.0, mirostat_eta: float = 0.1, - model: Optional[str] = None, logits_processor: Optional[llama.LogitsProcessorList] = None, grammar: Optional[llama.LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, **kwargs, # type: ignore ) -> Union[ llama_types.CreateChatCompletionResponse, Iterator[llama_types.CreateChatCompletionStreamResponse], - ]: - ... - + ]: ... -CHAT_HANDLERS: Dict[str, LlamaChatCompletionHandler] = {} - -def get_chat_completion_handler(name: str) -> LlamaChatCompletionHandler: - return CHAT_HANDLERS[name] +class LlamaChatCompletionHandlerNotFoundException(Exception): + pass -def register_chat_completion_handler(name: str): - def decorator(f: LlamaChatCompletionHandler): - CHAT_HANDLERS[name] = f - return f +class LlamaChatCompletionHandlerRegistry(Singleton): + _chat_handlers: Dict[str, LlamaChatCompletionHandler] = {} - return decorator + def register_chat_completion_handler( + self, + name: str, + chat_handler: LlamaChatCompletionHandler, + overwrite: bool = False, + ): + if not overwrite and name in self._chat_handlers: + raise ValueError( + f"Formatter with name '{name}' is already registered. Use `overwrite=True` to overwrite it." + ) + self._chat_handlers[name] = chat_handler + def unregister_chat_handler(self, name: str): + if name in self._chat_handlers: + del self._chat_handlers[name] + else: + raise ValueError(f"No formatter registered under the name '{name}'.") -def _get_system_message( - messages: List[llama_types.ChatCompletionRequestMessage], -) -> str: - """Get the first system message.""" - for message in messages: - if message["role"] == "system": - return message["content"] or "" - return "" + def get_chat_completion_handler_by_name( + self, name: str + ) -> LlamaChatCompletionHandler: + try: + chat_handler = self._chat_handlers[name] + return chat_handler + except KeyError: + raise LlamaChatCompletionHandlerNotFoundException( + f"Invalid chat handler: {name} (valid formats: {list(self._chat_handlers.keys())})" + ) -def _map_roles( - messages: List[llama_types.ChatCompletionRequestMessage], role_map: Dict[str, str] -) -> List[Tuple[str, Optional[str]]]: - """Map the message roles.""" - output: List[Tuple[str, Optional[str]]] = [] - for message in messages: - role = message["role"] - if role in role_map: - output.append((role_map[role], message["content"])) - return output +def get_chat_completion_handler(name: str) -> LlamaChatCompletionHandler: + return LlamaChatCompletionHandlerRegistry().get_chat_completion_handler_by_name( + name + ) -def _format_llama2( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str -) -> str: - """Format the prompt with the llama2 style.""" - seps = [sep, sep2] - ret = system_message + sep - for i, (role, message) in enumerate(messages): - if system_message and i == 0: - ret += message + seps[i % 2] - elif message: - ret += role + message + " " + seps[i % 2] - else: - ret += role + " " - return ret +def register_chat_completion_handler(name: str): + def decorator(f: LlamaChatCompletionHandler): + LlamaChatCompletionHandlerRegistry().register_chat_completion_handler(name, f) + return f + return decorator -def _format_add_colon_single( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str -) -> str: - """Format the prompt with the add-colon-single style.""" - ret = system_message + sep - for role, message in messages: - if message: - ret += role + ": " + message + sep - else: - ret += role + ":" - return ret +### Chat Formatter ### -def _format_add_colon_two( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str -) -> str: - """Format the prompt with the add-colon-two style.""" - seps = [sep, sep2] - ret = system_message + seps[0] - for i, (role, message) in enumerate(messages): - if message: - ret += role + ": " + message + seps[i % 2] - else: - ret += role + ":" - return ret +@dataclasses.dataclass +class ChatFormatterResponse: + """Dataclass that stores completion parameters for a given chat format and + create_chat_completion request. -def _format_no_colon_single( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str -) -> str: - """Format the prompt with the no-colon-single style.""" - ret = system_message - for role, message in messages: - if message: - ret += role + message + sep - else: - ret += role - return ret + prompt contains the formatted prompt generated from the chat format and messages. + stop contains the stop token or list of stop tokens to use for the chat format.""" + prompt: str + stop: Optional[Union[str, List[str]]] = None + stopping_criteria: Optional[llama.StoppingCriteriaList] = None + added_special: bool = False -def _format_add_colon_space_single( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str -) -> str: - """Format the prompt with the add-colon-space-single style.""" - ret = system_message + sep - for role, message in messages: - if message: - ret += role + ": " + message + sep - else: - ret += role + ": " # must be end with a space - return ret +class ChatFormatter(Protocol): + """Base Protocol for a chat formatter. A chat formatter is a function that + takes a list of messages and returns a chat format response which can be used + to generate a completion. The response can also include a stop token or list + of stop tokens to use for the completion.""" -def _format_chatml( - system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str -) -> str: - """Format the prompt with the chatml style.""" - ret = "" if system_message == "" else system_message + sep + "\n" - for role, message in messages: - if message: - ret += role + "\n" + message + sep + "\n" - else: - ret += role + "\n" - return ret + def __call__( + self, + *, + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, + ) -> ChatFormatterResponse: ... -@dataclasses.dataclass -class ChatFormatterResponse: - prompt: str - stop: Optional[Union[str, List[str]]] = None +class Jinja2ChatFormatter(ChatFormatter): + def __init__( + self, + template: str, + eos_token: str, + bos_token: str, + add_generation_prompt: bool = True, + stop_token_ids: Optional[List[int]] = None, + ): + """A chat formatter that uses jinja2 templates to format the prompt.""" + self.template = template + self.eos_token = eos_token + self.bos_token = bos_token + self.add_generation_prompt = add_generation_prompt + self.stop_token_ids = ( + set(stop_token_ids) if stop_token_ids is not None else None + ) + self._environment = ImmutableSandboxedEnvironment( + loader=jinja2.BaseLoader(), + trim_blocks=True, + lstrip_blocks=True, + ).from_string(self.template) -class ChatFormatter(Protocol): def __call__( self, *, messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, **kwargs: Any, ) -> ChatFormatterResponse: - ... + def raise_exception(message: str): + raise ValueError(message) + + prompt = self._environment.render( + messages=messages, + eos_token=self.eos_token, + bos_token=self.bos_token, + raise_exception=raise_exception, + add_generation_prompt=self.add_generation_prompt, + functions=functions, + function_call=function_call, + tools=tools, + tool_choice=tool_choice, + ) + stopping_criteria = None + if self.stop_token_ids is not None: -class BasicChatHandler: - def __init__(self, chat_format: str): - self.chat_format = chat_format + def stop_on_last_token( + tokens: npt.NDArray[np.intc], logits: npt.NDArray[np.single] + ) -> bool: + return tokens[-1] in self.stop_token_ids + stopping_criteria = llama.StoppingCriteriaList([stop_on_last_token]) + + return ChatFormatterResponse( + prompt=prompt, + stop=[self.eos_token], + stopping_criteria=stopping_criteria, + added_special=True, + ) + + def to_chat_handler(self) -> LlamaChatCompletionHandler: + return chat_formatter_to_chat_completion_handler(self) + + +def _convert_text_completion_logprobs_to_chat( + logprobs: Optional[llama_types.CompletionLogprobs], +) -> llama_types.ChatCompletionLogprobs: + if logprobs is None: + return None + + return { + "content": [ + { + "token": token, + "bytes": None, + "logprob": logprob, + "top_logprobs": [ + { + "token": top_token, + "logprob": top_logprob, + "bytes": None, + } + for top_token, top_logprob in top_logprobs.items() + ], + } for (token, logprob, top_logprobs) in zip(logprobs["tokens"], logprobs["token_logprobs"], logprobs["top_logprobs"]) + ], + "refusal": None, + } def _convert_text_completion_to_chat( completion: llama_types.Completion, ) -> llama_types.ChatCompletion: + assert "usage" in completion return { "id": "chat" + completion["id"], "object": "chat.completion", @@ -209,6 +300,7 @@ def _convert_text_completion_to_chat( "role": "assistant", "content": completion["choices"][0]["text"], }, + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), "finish_reason": completion["choices"][0]["finish_reason"], } ], @@ -232,6 +324,7 @@ def _convert_text_completion_chunks_to_chat( "delta": { "role": "assistant", }, + "logprobs": None, "finish_reason": None, } ], @@ -244,11 +337,14 @@ def _convert_text_completion_chunks_to_chat( "choices": [ { "index": 0, - "delta": { - "content": chunk["choices"][0]["text"], - } - if chunk["choices"][0]["finish_reason"] is None - else {}, + "delta": ( + { + "content": chunk["choices"][0]["text"], + } + if chunk["choices"][0]["finish_reason"] is None + else {} + ), + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), "finish_reason": chunk["choices"][0]["finish_reason"], } ], @@ -272,103 +368,330 @@ def _convert_completion_to_chat( return _convert_text_completion_to_chat(completion) -_CHAT_FORMATS: Dict[str, ChatFormatter] = {} - - -def register_chat_format(name: str): - def decorator(f: ChatFormatter): - def basic_create_chat_completion( - *, - llama: llama.Llama, - messages: List[llama_types.ChatCompletionRequestMessage], - functions: Optional[List[llama_types.ChatCompletionFunction]] = None, - function_call: Optional[ - llama_types.ChatCompletionRequestFunctionCall - ] = None, - tools: Optional[List[llama_types.ChatCompletionTool]] = None, - tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, - temperature: float = 0.2, - top_p: float = 0.95, - top_k: int = 40, - min_p: float = 0.05, - typical_p: float = 1.0, - stream: bool = False, - stop: Optional[Union[str, List[str]]] = [], - seed: Optional[int] = None, - response_format: Optional[ - llama_types.ChatCompletionRequestResponseFormat - ] = None, - max_tokens: Optional[int] = None, - presence_penalty: float = 0.0, - frequency_penalty: float = 0.0, - repeat_penalty: float = 1.1, - tfs_z: float = 1.0, - mirostat_mode: int = 0, - mirostat_tau: float = 5.0, - mirostat_eta: float = 0.1, - model: Optional[str] = None, - logits_processor: Optional[llama.LogitsProcessorList] = None, - grammar: Optional[llama.LlamaGrammar] = None, - logit_bias: Optional[Dict[str, float]] = None, - **kwargs, # type: ignore - ) -> Union[ - llama_types.CreateChatCompletionResponse, - Iterator[llama_types.CreateChatCompletionStreamResponse], - ]: - result = f( - messages=messages, - functions=functions, - function_call=function_call, - ) - prompt = result.prompt - if result.stop is not None: - stop = [] if stop is None else [stop] if isinstance(stop, str) else stop - rstop = result.stop if isinstance(result.stop, list) else [result.stop] - stop = stop + rstop - - if response_format is not None and response_format["type"] == "json_object": - grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF - ) +def _convert_completion_to_chat_function( + tool_name: str, + completion_or_chunks: Union[ + llama_types.CreateCompletionResponse, + Iterator[llama_types.CreateCompletionStreamResponse], + ], + stream: bool, +): + if not stream: + completion: llama_types.CreateCompletionResponse = completion_or_chunks # type: ignore + assert "usage" in completion + tool_id = "call_" + "_0_" + tool_name + "_" + completion["id"] + # TODO: Fix for legacy function calls + chat_completion: llama_types.CreateChatCompletionResponse = { + "id": "chat" + completion["id"], + "object": "chat.completion", + "created": completion["created"], + "model": completion["model"], + "choices": [ + { + "index": 0, + "message": { + "role": "assistant", + "content": None, + "function_call": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + "tool_calls": [ + { + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + } + ], + }, + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), + "finish_reason": "tool_calls", + } + ], + "usage": completion["usage"], + } + return chat_completion + else: + chunks: Iterator[llama_types.CreateCompletionStreamResponse] = completion_or_chunks # type: ignore - completion_or_chunks = llama.create_completion( - prompt=prompt, - temperature=temperature, - top_p=top_p, - top_k=top_k, - min_p=min_p, - typical_p=typical_p, - stream=stream, - stop=stop, - seed=seed, - max_tokens=max_tokens, - presence_penalty=presence_penalty, - frequency_penalty=frequency_penalty, - repeat_penalty=repeat_penalty, - tfs_z=tfs_z, - mirostat_mode=mirostat_mode, - mirostat_tau=mirostat_tau, - mirostat_eta=mirostat_eta, - model=model, - logits_processor=logits_processor, - grammar=grammar, - logit_bias=logit_bias, - ) - return _convert_completion_to_chat(completion_or_chunks, stream=stream) + def _stream_response_to_function_stream( + chunks: Iterator[llama_types.CreateCompletionStreamResponse], + ) -> Iterator[llama_types.CreateChatCompletionStreamResponse]: + # blank first message + first = True + id_ = None + created = None + model = None + tool_id = None + for chunk in chunks: + if first: + id_ = "chat" + chunk["id"] + created = chunk["created"] + model = chunk["model"] + tool_id = "call_" + "_0_" + tool_name + "_" + chunk["id"] + yield { + "id": id_, + "object": "chat.completion.chunk", + "created": created, + "model": model, + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": "assistant", + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + } + yield { + "id": "chat" + chunk["id"], + "object": "chat.completion.chunk", + "created": chunk["created"], + "model": chunk["model"], + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), + "delta": { + "role": None, + "content": None, + "function_call": { + "name": tool_name, + "arguments": chunk["choices"][0]["text"], + }, + "tool_calls": [ + { + "index": 0, + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": chunk["choices"][0][ + "text" + ], + }, + } + ], + }, + } + ], + } + first = False + continue + assert tool_id is not None + yield { + "id": "chat" + chunk["id"], + "object": "chat.completion.chunk", + "created": chunk["created"], + "model": chunk["model"], + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), + "delta": { + "role": None, + "content": None, + "function_call": { + "name": tool_name, + "arguments": chunk["choices"][0]["text"], + }, + "tool_calls": [ + { + "index": 0, + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": chunk["choices"][0]["text"], + }, + } + ], + }, + } + ], + } - register_chat_completion_handler(name)(basic_create_chat_completion) - return f + if id_ is not None and created is not None and model is not None: + yield { + "id": id_, + "object": "chat.completion.chunk", + "created": created, + "model": model, + "choices": [ + { + "index": 0, + "finish_reason": "tool_calls", + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + } - return decorator + return _stream_response_to_function_stream(chunks) -def get_chat_format(name: str): - try: - return _CHAT_FORMATS[name] - except KeyError: - raise ValueError( - f"Invalid chat format: {name} (valid formats: {list(_CHAT_FORMATS.keys())})" +def chat_formatter_to_chat_completion_handler( + chat_formatter: ChatFormatter, +) -> LlamaChatCompletionHandler: + def chat_completion_handler( + *, + llama: llama.Llama, + messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, + temperature: float = 0.2, + top_p: float = 0.95, + top_k: int = 40, + min_p: float = 0.05, + typical_p: float = 1.0, + stream: bool = False, + stop: Optional[Union[str, List[str]]] = [], + seed: Optional[int] = None, + response_format: Optional[ + llama_types.ChatCompletionRequestResponseFormat + ] = None, + max_tokens: Optional[int] = None, + presence_penalty: float = 0.0, + frequency_penalty: float = 0.0, + repeat_penalty: float = 1.1, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_tau: float = 5.0, + mirostat_eta: float = 0.1, + model: Optional[str] = None, + logits_processor: Optional[llama.LogitsProcessorList] = None, + grammar: Optional[llama.LlamaGrammar] = None, + logit_bias: Optional[Dict[str, float]] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, + **kwargs, # type: ignore + ) -> Union[ + llama_types.CreateChatCompletionResponse, + Iterator[llama_types.CreateChatCompletionStreamResponse], + ]: + result = chat_formatter( + messages=messages, + functions=functions, + function_call=function_call, + tools=tools, + tool_choice=tool_choice, + ) + prompt = llama.tokenize( + result.prompt.encode("utf-8"), + add_bos=not result.added_special, + special=True, ) + if result.stop is not None: + stop = [] if stop is None else [stop] if isinstance(stop, str) else stop + rstop = result.stop if isinstance(result.stop, list) else [result.stop] + stop = stop + rstop + + stopping_criteria = None + if result.stopping_criteria is not None: + stopping_criteria = result.stopping_criteria + + if response_format is not None and response_format["type"] == "json_object": + grammar = _grammar_for_response_format( + response_format, verbose=llama.verbose + ) + + # Convert legacy functions to tools + if functions is not None: + tools = [ + { + "type": "function", + "function": function, + } + for function in functions + ] + + # Convert legacy function_call to tool_choice + if function_call is not None: + if isinstance(function_call, str) and ( + function_call == "none" or function_call == "auto" + ): + tool_choice = function_call + if isinstance(function_call, dict) and "name" in function_call: + tool_choice = { + "type": "function", + "function": { + "name": function_call["name"], + }, + } + + tool = None + if ( + tool_choice is not None + and isinstance(tool_choice, dict) + and tools is not None + ): + name = tool_choice["function"]["name"] + tool = next((t for t in tools if t["function"]["name"] == name), None) + if tool is None: + raise ValueError(f"Tool choice '{name}' not found in tools.") + schema = tool["function"]["parameters"] + try: + # create grammar from json schema + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(schema), verbose=llama.verbose + ) + except Exception as e: + if llama.verbose: + print(str(e), file=sys.stderr) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + logprobs=top_logprobs if logprobs else None, + stream=stream, + stop=stop, + seed=seed, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + stopping_criteria=stopping_criteria, + grammar=grammar, + logit_bias=logit_bias, + ) + if tool is not None: + tool_name = tool["function"]["name"] + return _convert_completion_to_chat_function( + tool_name, completion_or_chunks, stream + ) + return _convert_completion_to_chat(completion_or_chunks, stream=stream) + + return chat_completion_handler def hf_autotokenizer_to_chat_formatter( @@ -377,22 +700,284 @@ def hf_autotokenizer_to_chat_formatter( # https://huggingface.co/docs/transformers/main/chat_templating # https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1#instruction-format # https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1/blob/main/tokenizer_config.json - from transformers import AutoTokenizer + from transformers import AutoTokenizer # type: ignore - tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path) + tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path) # type: ignore def format_autotokenizer( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - tokenizer.use_default_system_prompt = False - _prompt = tokenizer.apply_chat_template(messages, tokenize=False) + tokenizer.use_default_system_prompt = False # type: ignore + prompt: str = tokenizer.apply_chat_template(messages, tokenize=False) # type: ignore + assert isinstance(prompt, str) # Return formatted prompt and eos token by default - return ChatFormatterResponse(prompt=_prompt, stop=tokenizer.eos_token) + return ChatFormatterResponse( + prompt=prompt, stop=tokenizer.eos_token, added_special=True + ) return format_autotokenizer +def hf_autotokenizer_to_chat_completion_handler( + pretrained_model_name_or_path: Union[str, os.PathLike[str]] +) -> LlamaChatCompletionHandler: + chat_formatter = hf_autotokenizer_to_chat_formatter(pretrained_model_name_or_path) + return chat_formatter_to_chat_completion_handler(chat_formatter) + + +def hf_tokenizer_config_to_chat_formatter( + tokenizer_config: Dict[str, Any], + add_generation_prompt: bool = True, +) -> ChatFormatter: + assert isinstance(tokenizer_config, dict) + + assert "chat_template" in tokenizer_config + assert isinstance(tokenizer_config["chat_template"], str) + chat_template = tokenizer_config["chat_template"] + + assert "bos_token" in tokenizer_config + assert isinstance(tokenizer_config["bos_token"], str) + bos_token = tokenizer_config["bos_token"] + + assert "eos_token" in tokenizer_config + assert isinstance(tokenizer_config["eos_token"], str) + eos_token = tokenizer_config["eos_token"] + + env = ImmutableSandboxedEnvironment( + trim_blocks=True, + lstrip_blocks=True, + ).from_string(chat_template) + + def format_tokenizer_config( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, + ) -> ChatFormatterResponse: + # TODO: veryify this is correct + # Add a blank assistant message to the end of the messages to prompt the model to generate a response + if add_generation_prompt: + messages = [ + *messages, + llama_types.ChatCompletionRequestAssistantMessage( + role="assistant", content="" + ), + ] + prompt = env.render( + messages=messages, + bos_token=bos_token, + eos_token=eos_token, + ) + return ChatFormatterResponse( + prompt=prompt, stop=[eos_token, bos_token], added_special=True + ) + + return format_tokenizer_config + + +def hf_tokenizer_config_to_chat_completion_handler( + tokenizer_config: Dict[str, Any], + add_generation_prompt: bool = True, +) -> LlamaChatCompletionHandler: + chat_formatter = hf_tokenizer_config_to_chat_formatter( + tokenizer_config, add_generation_prompt=add_generation_prompt + ) + return chat_formatter_to_chat_completion_handler(chat_formatter) + + +def guess_chat_format_from_gguf_metadata(metadata: Dict[str, str]) -> Optional[str]: + if "tokenizer.chat_template" not in metadata: + return None + + if metadata["tokenizer.chat_template"] == CHATML_CHAT_TEMPLATE: + return "chatml" + + if ( + metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE + or metadata["tokenizer.chat_template"] == MIXTRAL_INSTRUCT_CHAT_TEMPLATE + ): + return "mistral-instruct" + + if metadata["tokenizer.chat_template"] == LLAMA3_INSTRUCT_CHAT_TEMPLATE: + return "llama-3" + + return None + + +### Utility functions for formatting chat prompts ### +# TODO: Replace these with jinja2 templates + + +def _get_system_message( + messages: List[llama_types.ChatCompletionRequestMessage], +) -> str: + """Get the first system message.""" + for message in messages: + if message["role"] == "system": + return message["content"] or "" + return "" + + +def _map_roles( + messages: List[llama_types.ChatCompletionRequestMessage], + role_map: Dict[str, str], +) -> List[Tuple[str, Optional[str]]]: + """Map the message roles.""" + output: List[Tuple[str, Optional[str]]] = [] + for message in messages: + role = message["role"] + if role in role_map: + content: str | None = ( + message["content"] if isinstance(message["content"], str) else None + ) + output.append((role_map[role], content)) + return output + + +def _format_llama2( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str +) -> str: + """Format the prompt with the llama2 style.""" + seps = [sep, sep2] + ret = system_message + sep + for i, (role, message) in enumerate(messages): + if system_message and i == 0: + m = message or "" + ret += m + seps[i % 2] + elif message: + ret += role + message + " " + seps[i % 2] + else: + ret += role + " " + return ret + + +def _format_add_colon_single( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the add-colon-single style.""" + ret = system_message + sep + for role, message in messages: + if message: + ret += role + ": " + message + sep + else: + ret += role + ":" + return ret + + +def _format_add_colon_two( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str, sep2: str +) -> str: + """Format the prompt with the add-colon-two style.""" + seps = [sep, sep2] + ret = system_message + seps[0] + for i, (role, message) in enumerate(messages): + if message: + ret += role + ": " + message + seps[i % 2] + else: + ret += role + ":" + return ret + + +def _format_no_colon_single( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the no-colon-single style.""" + ret = system_message + for role, message in messages: + if message: + ret += role + message + sep + else: + ret += role + return ret + + +def _format_add_colon_space_single( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the add-colon-space-single style.""" + ret = system_message + sep + for role, message in messages: + if message: + ret += role + ": " + message + sep + else: + ret += role + ": " # must be end with a space + return ret + + +def _format_chatml( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the chatml style.""" + ret = "" if system_message == "" else system_message + sep + "\n" + for role, message in messages: + if message: + ret += role + "\n" + message + sep + "\n" + else: + ret += role + "\n" + return ret + + +def _format_chatglm3( + system_message: str, messages: List[Tuple[str, Optional[str]]], sep: str +) -> str: + """Format the prompt with the chatglm3 style.""" + ret = "" + if system_message: + ret += system_message + for role, message in messages: + if message: + ret += role + "\n" + " " + message + else: + ret += role + return ret + + +def _grammar_for_json(verbose: bool = False): + return llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=verbose + ) + + +def _grammar_for_json_schema( + schema: str, verbose: bool = False, fallback_to_json: bool = True +): + try: + return llama_grammar.LlamaGrammar.from_json_schema(schema, verbose=verbose) + except Exception as e: + if fallback_to_json: + return _grammar_for_json(verbose=verbose) + else: + raise e + + +def _grammar_for_response_format( + response_format: llama_types.ChatCompletionRequestResponseFormat, + verbose: bool = False, +): + if response_format["type"] != "json_object": + return None + + if "schema" in response_format: + return _grammar_for_json_schema( + json.dumps(response_format["schema"]), verbose=verbose + ) + else: + return _grammar_for_json(verbose=verbose) + + +### Chat Formats ### + + +def register_chat_format(name: str): + def decorator(f: ChatFormatter): + chat_completion_handler = chat_formatter_to_chat_completion_handler(f) + LlamaChatCompletionHandlerRegistry().register_chat_completion_handler( + name, chat_completion_handler + ) + return f + + return decorator + + # see https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/tokenization_llama.py # system prompt is "embedded" in the first message @register_chat_format("llama-2") @@ -400,7 +985,7 @@ def format_llama2( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - _system_template = "[INST] <>\n{system_message}\n<>" + _system_template = "[INST] <>\n{system_message}\n<>" _roles = dict(user="[INST]", assistant="[/INST]") _messages = _map_roles(messages, _roles) system_message = _get_system_message(messages) @@ -410,6 +995,25 @@ def format_llama2( return ChatFormatterResponse(prompt=_prompt) +# Chat format for Llama-3 models, see more details at: +# https://github.com/meta-llama/llama3/blob/main/llama/tokenizer.py#L202-L229 +@register_chat_format("llama-3") +def format_llama3( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, +) -> ChatFormatterResponse: + _roles = dict( + system="<|start_header_id|>system<|end_header_id|>\n\n", + user="<|start_header_id|>user<|end_header_id|>\n\n", + assistant="<|start_header_id|>assistant<|end_header_id|>\n\n", + ) + _sep = "<|eot_id|>" + _messages = _map_roles(messages, _roles) + _messages.append((_roles["assistant"], None)) + _prompt = _format_no_colon_single("", _messages, _sep) + return ChatFormatterResponse(prompt=_prompt, stop=_sep) + + @register_chat_format("alpaca") def format_alpaca( messages: List[llama_types.ChatCompletionRequestMessage], @@ -423,21 +1027,23 @@ def format_alpaca( _prompt = _format_add_colon_two(system_message, _messages, _sep, _sep2) return ChatFormatterResponse(prompt=_prompt) + @register_chat_format("qwen") def format_qwen( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: _roles = dict(user="<|im_start|>user", assistant="<|im_start|>assistant") - system_message="You are a helpful assistant." - system_template="<|im_start|>system\n{system_message}" - system_message=system_template.format(system_message=system_message) + system_message = _get_system_message(messages) or "You are a helpful assistant." + system_template = "<|im_start|>system\n{system_message}" + system_message = system_template.format(system_message=system_message) _messages = _map_roles(messages, _roles) _messages.append((_roles["assistant"], None)) _sep = "<|im_end|>" _prompt = _format_chatml(system_message, _messages, _sep) _sep2 = "<|endoftext|>" - return ChatFormatterResponse(prompt=_prompt,stop=_sep2) + return ChatFormatterResponse(prompt=_prompt, stop=_sep2) + @register_chat_format("vicuna") def format( @@ -508,17 +1114,14 @@ def format_openbuddy( messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: - _system_message = """Consider a conversation between User (a human) and Assistant (named Buddy). -Buddy is an INTP-T, a friendly, intelligent and multilingual AI assistant, by OpenBuddy team. GitHub: https://github.com/OpenBuddy/OpenBuddy -Buddy cannot access the Internet. -Buddy can fluently speak the user's language (e.g. English, Chinese). -Buddy can generate poems, stories, code, essays, songs, parodies, and more. -Buddy possesses vast knowledge about the world, history, and culture. -Buddy's responses are always safe, creative, high-quality, human-like, and interesting. -Buddy strictly refuses to discuss political, NSFW, or other unsafe topics. - -User: Hi. -Assistant: Hi, I'm Buddy, your AI assistant. How can I help you today?""" + _system_message = """You are a helpful, respectful and honest INTP-T AI Assistant named Buddy. You are talking to a human User. +Always answer as helpfully and logically as possible, while being safe. Your answers should not include any harmful, political, religious, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature. +If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. +You can speak fluently in many languages, for example: English, Chinese. +You cannot access the internet, but you have vast knowledge, cutoff: 2021-09. +You are trained by OpenBuddy team, (https://openbuddy.ai, https://github.com/OpenBuddy/OpenBuddy), you are based on LLaMA and Falcon transformers model, not related to GPT or OpenAI. + +""" _roles = dict(user="User", assistant="Assistant") _sep = "\n" system_message = _system_message @@ -636,6 +1239,7 @@ def format_mistrallite( _prompt = _format_no_colon_single(system_message, _messages, _sep) return ChatFormatterResponse(prompt=_prompt) + @register_chat_format("zephyr") def format_zephyr( messages: List[llama_types.ChatCompletionRequestMessage], @@ -686,6 +1290,44 @@ def format_chatml( return ChatFormatterResponse(prompt=_prompt, stop=_sep) +@register_chat_format("mistral-instruct") +def format_mistral_instruct( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, +) -> ChatFormatterResponse: + eos = "" + stop = eos + prompt = "" + for message in messages: + if ( + message["role"] == "user" + and message["content"] is not None + and isinstance(message["content"], str) + ): + prompt += "[INST] " + message["content"] + elif message["role"] == "assistant" and message["content"] is not None: + prompt += " [/INST]" + message["content"] + eos + prompt += " [/INST]" + return ChatFormatterResponse(prompt=prompt, stop=stop) + + +@register_chat_format("chatglm3") +def format_chatglm3( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, +) -> ChatFormatterResponse: + system_template = """<|system|> +{system_message}""" + system_message = _get_system_message(messages) + system_message = system_template.format(system_message=system_message) + _roles = dict(user="<|user|>", assistant="<|assistant|>") + _sep = "" + _messages = _map_roles(messages, _roles) + _messages.append((_roles["assistant"], None)) + _prompt = _format_chatglm3(system_message, _messages, _sep) + return ChatFormatterResponse(prompt=_prompt, stop=_sep) + + @register_chat_format("openchat") def format_openchat( messages: List[llama_types.ChatCompletionRequestMessage], @@ -704,6 +1346,51 @@ def format_openchat( return ChatFormatterResponse(prompt=_prompt, stop=_sep) +# Chat format for Saiga models, see more details and available models: +# https://huggingface.co/collections/IlyaGusev/saiga2-saigamistral-6505d4ccc3d1e53166b636cd +@register_chat_format("saiga") +def format_saiga( + messages: list[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, +) -> ChatFormatterResponse: + _message_template = "{role}\n{content}" + _roles = dict(user="user", bot="bot", system="system") + _messages = _map_roles(messages, _roles) + + _prompt = "" + for role, content in _messages: + if content: + _prompt += _message_template.format(role=role, content=content) + else: + _prompt += f"{role}\n" + # Response template + _prompt += "bot" + return ChatFormatterResponse(prompt=_prompt.strip()) + + +# Chat format for Google's Gemma models, see more details and available models: +# https://huggingface.co/collections/google/gemma-release-65d5efbccdbb8c4202ec078b +@register_chat_format("gemma") +def format_gemma( + messages: List[llama_types.ChatCompletionRequestMessage], + **kwargs: Any, +) -> ChatFormatterResponse: + system_message = _get_system_message(messages) + if system_message != "": + logger.debug( + "`role='system'` messages are not allowed on Google's Gemma models." + ) + _roles = dict(user="user\n", assistant="model\n") + _sep = "\n" + _messages = _map_roles(messages, _roles) + _messages.append((_roles["assistant"], None)) + _prompt = _format_no_colon_single(system_message="", messages=_messages, sep=_sep) + return ChatFormatterResponse(prompt=_prompt, stop=_sep) + + +# Tricky chat formats that require custom chat handlers + + @register_chat_completion_handler("functionary") def functionary_chat_handler( llama: llama.Llama, @@ -952,12 +1639,12 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): function_call = completion_text.split(".")[-1][:-1] new_prompt = prompt + completion_text + stop elif isinstance(function_call, str) and function_call != "none": - new_prompt = prompt + f":\n" + new_prompt = prompt + ":\n" elif isinstance(function_call, dict): new_prompt = prompt + f" to=functions.{function_call['name']}:\n" function_call = function_call["name"] else: - new_prompt = prompt + f":\n" + new_prompt = prompt + ":\n" function_body = None for function in functions or []: @@ -976,7 +1663,8 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): json.dumps(function_body) ) grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.json_schema_to_gbnf(json.dumps(function_body)) + llama_grammar.json_schema_to_gbnf(json.dumps(function_body)), + verbose=llama.verbose, ) print(grammar_text) except Exception as e: @@ -987,11 +1675,14 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): print(e) with suppress_stdout_stderr(disable=llama.verbose): grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF + llama_grammar.JSON_GBNF, + verbose=llama.verbose, ) else: with suppress_stdout_stderr(disable=llama.verbose): - grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) completion: llama_types.Completion = llama.create_completion( prompt=new_prompt, @@ -1050,6 +1741,7 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): } ], }, + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), "finish_reason": "tool_calls", } ], @@ -1057,166 +1749,293 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): ) -class Llava15ChatHandler: - _clip_free = None - - def __init__(self, clip_model_path: str, verbose: bool = False): - import llama_cpp.llava_cpp as llava_cpp +@register_chat_completion_handler("functionary-v1") +@register_chat_completion_handler("functionary-v2") +def functionary_v1_v2_chat_handler( + llama: llama.Llama, + messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, + temperature: float = 0.2, + top_p: float = 0.95, + top_k: int = 40, + min_p: float = 0.05, + typical_p: float = 1.0, + stream: bool = False, + stop: Optional[Union[str, List[str]]] = [], + response_format: Optional[llama_types.ChatCompletionRequestResponseFormat] = None, + max_tokens: Optional[int] = None, + presence_penalty: float = 0.0, + frequency_penalty: float = 0.0, + repeat_penalty: float = 1.1, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_tau: float = 5.0, + mirostat_eta: float = 0.1, + model: Optional[str] = None, + logits_processor: Optional[llama.LogitsProcessorList] = None, + grammar: Optional[llama.LlamaGrammar] = None, + **kwargs, # type: ignore +) -> Union[llama_types.ChatCompletion, Iterator[llama_types.ChatCompletionChunk]]: + SYSTEM_MESSAGE = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant calls functions with appropriate input when necessary""" - self._llava_cpp = llava_cpp - self.clip_model_path = clip_model_path - self.verbose = verbose - self._clip_free = self._llava_cpp._libllava.clip_free # type: ignore + tokenizer = llama.tokenizer_ + assert hasattr( + tokenizer, "hf_tokenizer" + ), "Please provide a valid hf_tokenizer_path from https://huggingface.co/meetkai when initializing the Llama class" + from transformers import AutoTokenizer - with suppress_stdout_stderr(disable=self.verbose): - self.clip_ctx = self._llava_cpp.clip_model_load( - self.clip_model_path.encode(), 0 - ) + if "<|START_OF_FUNCTION_CALL|>" in tokenizer.hf_tokenizer.additional_special_tokens: + version = "v1" + END_SYSTEM_TOKEN = "<|END_OF_SYSTEM|>" + END_USER_TOKEN = "<|END_OF_USER|>" + END_ASSISTANT_TOKEN = "<|END_OF_ASSISTANT|>" + END_FUNCTION_RESULT_TOKEN = "<|END_OF_FUNCTION_RESULT|>" + START_FUNCTION_CALL_TOKEN = "<|START_OF_FUNCTION_CALL|>" + END_FUNCTION_CALL_TOKEN = "<|END_OF_FUNCTION_CALL|>" + else: + version = "v2" + RECIPIENT_TOKEN = "<|recipient|>" + FROM_TOKEN = "<|from|>" + STOP_TOKEN = "<|stop|>" + CONTENT_TOKEN = "<|content|>" - def __del__(self): - with suppress_stdout_stderr(disable=self.verbose): - if self.clip_ctx is not None and self._clip_free is not None: - self._clip_free(self.clip_ctx) - self.clip_ctx = None + def generate_type_definition( + param: Dict[str, llama_types.JsonType], indent_level: int, shared_defs + ) -> str: + indent = " " * indent_level + if "$ref" in param: + # Reference to a shared definition + ref_name = param["$ref"].split("/")[ + -1 + ] # Extract the type name from the reference + return ref_name + elif param.get("type") == "array": + items = param.get("items", {}) + item_type = generate_type_definition(items, indent_level + 1, shared_defs) + return f"Array<{item_type}>" + elif param.get("type") == "object": + properties = param.get("properties", {}) + nested_schema = "{\n" + for nested_param_name, nested_param in properties.items(): + nested_param_type = generate_type_definition( + nested_param, indent_level + 1, shared_defs + ) + nested_schema += ( + f"{indent} {nested_param_name}: {nested_param_type},\n" + ) + nested_schema += indent + "}" + return nested_schema + elif "enum" in param: + # Enum type + return " | ".join([f'"{enum_value}"' for enum_value in param["enum"]]) + else: + # Simple type + return param.get("type", "any") - def load_image(self, image_url: str) -> bytes: - if image_url.startswith("data:"): - import base64 + def generate_shared_definitions(shared_defs, indent_level: int) -> str: + indent = " " * indent_level + shared_definitions = "" + for def_name, def_properties in shared_defs.items(): + shared_definitions += f"{indent}type {def_name} = " + if def_properties.get("type") == "object": + shared_definitions += generate_type_definition( + def_properties, indent_level, shared_defs + ) + elif "enum" in def_properties: + # Enum type + shared_definitions += " | ".join( + [f'"{enum_value}"' for enum_value in def_properties["enum"]] + ) + shared_definitions += ";\n" + return shared_definitions - image_bytes = base64.b64decode(image_url.split(",")[1]) - return image_bytes - else: - import urllib.request + def generate_schema_from_functions(functions, namespace="functions") -> str: + schema = ( + "// Supported function definitions that should be called when necessary.\n" + ) + schema += f"namespace {namespace} {{\n\n" - with urllib.request.urlopen(image_url) as f: - image_bytes = f.read() - return image_bytes + # Generate shared definitions + shared_definitions = {} + for function in functions: + parameters = function.get("parameters", {}) + shared_definitions.update(parameters.get("$defs", {})) - def __call__( - self, - *, - llama: llama.Llama, + schema += generate_shared_definitions(shared_definitions, 1) + + for function in functions: + function_name = function["name"] + description = function.get("description", "") + parameters = function.get("parameters", {}) + required_params = parameters.get("required", []) + + schema += f"// {description}\n" + schema += f"type {function_name} = (_: {{\n" + + for param_name, param in parameters.get("properties", {}).items(): + param_description = param.get("description", "") + param_type = generate_type_definition(param, 2, shared_definitions) + optional_indicator = "" if param_name in required_params else "?" + schema += f"// {param_description}\n" + schema += f"{param_name}{optional_indicator}: {param_type},\n" + schema += "}) => any;\n\n" + + schema += "}} // namespace {}".format(namespace) + return schema + + def prepare_messages_for_inference( messages: List[llama_types.ChatCompletionRequestMessage], - functions: Optional[List[llama_types.ChatCompletionFunction]] = None, - function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tokenizer: AutoTokenizer, + version: Literal["v1", "v2"], + functions: Optional[List[llama_types.ChatCompletionFunctions]] = None, tools: Optional[List[llama_types.ChatCompletionTool]] = None, - tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, - temperature: float = 0.2, - top_p: float = 0.95, - top_k: int = 40, - min_p: float = 0.05, - typical_p: float = 1.0, - stream: bool = False, - stop: Optional[Union[str, List[str]]] = [], - response_format: Optional[ - llama_types.ChatCompletionRequestResponseFormat - ] = None, - max_tokens: Optional[int] = None, - presence_penalty: float = 0.0, - frequency_penalty: float = 0.0, - repeat_penalty: float = 1.1, - tfs_z: float = 1.0, - mirostat_mode: int = 0, - mirostat_tau: float = 5.0, - mirostat_eta: float = 0.1, - model: Optional[str] = None, - logits_processor: Optional[llama.LogitsProcessorList] = None, - grammar: Optional[llama.LlamaGrammar] = None, - **kwargs, # type: ignore - ) -> Union[ - llama_types.CreateChatCompletionResponse, - Iterator[llama_types.CreateChatCompletionStreamResponse], - ]: - assert ( - llama.context_params.logits_all is True - ) # BUG: logits_all=True is required for llava - assert self.clip_ctx is not None - system_prompt = _get_system_message(messages) - system_prompt = ( - system_prompt - if system_prompt != "" - else "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." - ) - user_role = "\nUSER:" - assistant_role = "\nASSISTANT:" - llama.reset() - llama.eval(llama.tokenize(system_prompt.encode("utf8"), add_bos=True)) - for message in messages: - if message["role"] == "user" and message["content"] is not None: - if isinstance(message["content"], str): - llama.eval( - llama.tokenize( - f"{user_role} {message['content']}".encode("utf8"), - add_bos=False, - ) - ) - else: - assert isinstance(message["content"], list) - llama.eval( - llama.tokenize(f"{user_role} ".encode("utf8"), add_bos=False) + tool_choice: Union[Dict, str] = "auto", + ): + all_messages: List[llama_types.ChatCompletionRequestMessage] = [] + if tool_choice == "none": + all_messages.append( + llama_types.ChatCompletionRequestSystemMessage( + role="system", content=generate_schema_from_functions([]) + ) + ) + else: + if functions is not None: + all_messages.append( + llama_types.ChatCompletionRequestSystemMessage( + role="system", content=generate_schema_from_functions(functions) ) - for content in message["content"]: - if content["type"] == "text": - llama.eval( - llama.tokenize( - f"{content['text']}".encode("utf8"), add_bos=False - ) - ) - if content["type"] == "image_url": - image_bytes = ( - self.load_image(content["image_url"]["url"]) - if isinstance(content["image_url"], dict) - else self.load_image(content["image_url"]) - ) - import array - - data_array = array.array("B", image_bytes) - c_ubyte_ptr = ( - ctypes.c_ubyte * len(data_array) - ).from_buffer(data_array) - with suppress_stdout_stderr(disable=self.verbose): - embed = ( - self._llava_cpp.llava_image_embed_make_with_bytes( - ctx_clip=self.clip_ctx, - n_threads=llama.context_params.n_threads, - image_bytes=c_ubyte_ptr, - image_bytes_length=len(image_bytes), - ) - ) - try: - n_past = ctypes.c_int(llama.n_tokens) - n_past_p = ctypes.pointer(n_past) - with suppress_stdout_stderr(disable=self.verbose): - self._llava_cpp.llava_eval_image_embed( - ctx_llama=llama.ctx, - embed=embed, - n_batch=llama.n_batch, - n_past=n_past_p, - ) - assert llama.n_ctx() >= n_past.value - llama.n_tokens = n_past.value - finally: - with suppress_stdout_stderr(disable=self.verbose): - self._llava_cpp.llava_image_embed_free(embed) - if message["role"] == "assistant" and message["content"] is not None: - llama.eval( - llama.tokenize( - f"ASSISTANT: {message['content']}".encode("utf8"), add_bos=False + ) + elif tools is not None and tool_choice != "none": + all_messages.append( + llama_types.ChatCompletionRequestSystemMessage( + role="system", + content=generate_schema_from_functions( + [ + tool["function"] + for tool in tools + if tool["type"] == "function" + ] + ), ) ) - assert llama.n_ctx() >= llama.n_tokens - llama.eval(llama.tokenize(f"{assistant_role}".encode("utf8"), add_bos=False)) - assert llama.n_ctx() >= llama.n_tokens - prompt = llama.input_ids[: llama.n_tokens].tolist() + all_messages.append( + llama_types.ChatCompletionRequestSystemMessage( + role="system", content=SYSTEM_MESSAGE + ) + ) - if response_format is not None and response_format["type"] == "json_object": - with suppress_stdout_stderr(disable=self.verbose): + for message in messages: + # Function call responses + if message["role"] == "function" and "name" in message: + message["name"] = f"functions.{message['name']}" + # Function call requests by assistant + if "function_call" in message: + message["function_call"][ + "name" + ] = f"functions.{message['function_call']['name']}" + all_messages.append(message) + + if version == "v1": + suffix = "assistant:\n" + else: + suffix = "<|from|>assistant\n<|recipient|>" + + return ( + tokenizer.hf_tokenizer.apply_chat_template(all_messages, tokenize=False) + + suffix + ) + + if tools is not None: + functions = [tool["function"] for tool in tools if tool["type"] == "function"] + + if tool_choice is not None: + function_call = ( + tool_choice if isinstance(tool_choice, str) else tool_choice["function"] + ) + elif function_call is not None: + pass + else: + function_call = "auto" + + prompt = prepare_messages_for_inference( + messages, tokenizer, version, functions, tools, function_call + ) + + # If no tools/functions are provided + if function_call == "none" or functions is None or len(functions) == 0: + if version == "v1": + stop = END_ASSISTANT_TOKEN + else: + stop = STOP_TOKEN + prompt += "all\n<|content|>" + + completion_or_completion_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ) + if stream is False: + completion_or_completion_chunks["choices"][0]["text"] = ( + completion_or_completion_chunks["choices"][0]["text"].lstrip() + ) + return _convert_completion_to_chat(completion_or_completion_chunks, stream=stream) # type: ignore + + def get_grammar(function_call): + function_body = None + for function in functions or []: + if function["name"] == function_call: + function_body = function["parameters"] + break + for tool in tools or []: + if tool["type"] == "function" and tool["function"]["name"] == function_call: + function_body = tool["function"]["parameters"] + break + + try: + with suppress_stdout_stderr(disable=llama.verbose): + grammar_text = llama_grammar.json_schema_to_gbnf( + json.dumps(function_body) + ) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.json_schema_to_gbnf(json.dumps(function_body)) + ) + print(grammar_text) + except Exception as e: + if llama.verbose: + print( + "Failed to parse function body as JSON schema, falling back to default grammar" + ) + print(e) + with suppress_stdout_stderr(disable=llama.verbose): grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF + llama_grammar.JSON_GBNF, verbose=llama.verbose ) - return _convert_completion_to_chat( + return grammar + + def create_completion(prompt, stop, grammar): + completion = cast( + llama_types.Completion, llama.create_completion( prompt=prompt, temperature=temperature, @@ -1238,5 +2057,1760 @@ def __call__( logits_processor=logits_processor, grammar=grammar, ), - stream=stream, ) + + return completion + + content = "" + function_calls, function_bodies = [], [] + completion_tokens = 0 + + def generate_streaming(tools, functions, function_call, prompt): + assert version == "v2", "Streaming for v1 is not supported" + + chunk_id, chunk_created = None, None + + # If tool_choice/function_call is provided + if isinstance(function_call, dict): + prompt += f"{function_call['name']}\n{CONTENT_TOKEN}" + grammar = get_grammar(function_call["name"]) + stops = [STOP_TOKEN, FROM_TOKEN] + tool_id = "".join( + [random.choice(string.ascii_letters + string.digits) for _ in range(24)] + ) + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion_text = "" + first = True + for chunk in completion: + # Yield the tool/function name first + if first: + if tools is not None: + func_call_dict = { + "tool_calls": [ + { + "index": 0, + "id": "call_" + tool_id, + "type": "function", + "function": { + "name": function_call["name"], + "arguments": "", + }, + } + ] + } + else: + func_call_dict = { + "function_call": { + "name": function_call["name"], + "arguments": "", + } + } + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk["id"], + object="chat.completion.chunk", + created=chunk["created"], + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": None, + "delta": { + "role": None, + "content": None, + **func_call_dict, + }, + } + ], + ) + first = False + if tools is not None: + func_call_dict = { + "tool_calls": [ + { + "index": 0, + "id": "call_" + tool_id, + "type": "function", + "function": { + "name": None, + "arguments": chunk["choices"][0]["text"].rstrip(), + }, + } + ] + } + else: + func_call_dict = { + "function_call": { + "name": None, + "arguments": chunk["choices"][0]["text"].rstrip(), + } + } + if len(chunk["choices"][0]["text"].rstrip()) > 0: + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk["id"], + object="chat.completion.chunk", + created=chunk["created"], + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), + "delta": { + "role": None, + "content": None, + **func_call_dict, + }, + } + ], + ) + # Yield tool_call/function_call stop message + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk["id"], + object="chat.completion.chunk", + created=chunk["created"], + model=chunk["model"], + choices=[ + { + "index": 0, + "finish_reason": ( + "tool_calls" if tools is not None else "function_call" + ), + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + ) + # If "auto" or no tool_choice/function_call + elif isinstance(function_call, str) and function_call == "auto": + tool_index = 0 + while True: + # Generate function name first + grammar = None + stops = CONTENT_TOKEN + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) + completion_text = "" + for chunk in completion: + completion_text += chunk["choices"][0]["text"] + if chunk_id is None: + chunk_id = chunk["id"] + if chunk_created is None: + chunk_created = chunk["created"] + function_name = completion_text.strip() + if function_name == "all": + prompt += "all\n<|content|>" + # Yield the first empty message for content + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + model=chunk["model"], + created=chunk_created, + object="chat.completion.chunk", + choices=[ + { + "index": 0, + "delta": {"role": "assistant", "content": ""}, + "logprobs": None, + "finish_reason": None, + } + ], + ) + else: + prompt += f"{function_name}\n<|content|>" + grammar = get_grammar(function_name) + tool_id = "".join( + [ + random.choice(string.ascii_letters + string.digits) + for _ in range(24) + ] + ) + if tools is not None: + func_call_dict = { + "tool_calls": [ + { + "index": tool_index, + "id": "call_" + tool_id, + "type": "function", + "function": { + "name": function_name, + "arguments": "", + }, + } + ] + } + else: + func_call_dict = { + "function_call": {"name": function_name, "arguments": ""} + } + # Stream function name + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), + "delta": { + "role": "assistant", + "content": None, + **func_call_dict, + }, + } + ], + ) + # Generate content + stops = [RECIPIENT_TOKEN, STOP_TOKEN] + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) + if function_name == "all": + completion_text = "" + stop_sequence, buffer, is_end = ( + "\n<|from|>assistant\n<|recipient|>", + [], + False, + ) + for i, chunk in enumerate(completion): + completion_text += chunk["choices"][0]["text"] + if is_end: + buffer.append(chunk["choices"][0]["text"].strip(" ")) + if stop_sequence.startswith("".join(buffer)): + continue + else: + buffer.pop() + while len(buffer) > 0: + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), + "delta": { + "role": "assistant", + "content": buffer.pop(0), + }, + } + ], + ) + is_end = False + elif chunk["choices"][0]["text"] == "\n": + is_end = True + buffer.append(chunk["choices"][0]["text"].strip(" ")) + continue + + if len(buffer) == 0 and len(chunk["choices"][0]["text"]) > 0: + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), + "delta": { + "role": "assistant", + "content": ( + chunk["choices"][0]["text"] + if i > 0 + else chunk["choices"][0][ + "text" + ].lstrip() + ), + }, + } + ], + ) + # Check whether the model wants to generate another turn + if ( + "<|from|> assistant" in completion_text + or "<|from|>assistant" in completion_text + ): + if completion_text.endswith("\n<|from|>assistant\n"): + cleaned_completion_text = completion_text[ + : -len("\n<|from|>assistant\n") + ].strip() + elif completion_text.endswith("\n<|from|> assistant\n"): + cleaned_completion_text = completion_text[ + : -len("\n<|from|> assistant\n") + ].strip() + else: + cleaned_completion_text = completion_text.strip() + prompt += f"{cleaned_completion_text}\n<|from|>assistant\n<|recipient|>" + else: + # Yield stop message + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + model=chunk["model"], + created=chunk_created, + object="chat.completion.chunk", + choices=[ + { + "index": 0, + "delta": {}, + "logprobs": None, + "finish_reason": "stop", + } + ], + ) + break + else: + # Check whether the model wants to generate another turn + completion_text = "" + for chunk in completion: + completion_text += chunk["choices"][0]["text"] + if len(chunk["choices"][0]["text"].rstrip()) > 0: + if tools is not None: + func_call_dict = { + "tool_calls": [ + { + "index": tool_index, + "id": "call_" + tool_id, + "type": "function", + "function": { + "name": None, + "arguments": chunk["choices"][0][ + "text" + ].rstrip(), + }, + } + ] + } + else: + func_call_dict = { + "function_call": { + "name": None, + "arguments": chunk["choices"][0][ + "text" + ].rstrip(), + } + } + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "logprobs": _convert_text_completion_logprobs_to_chat(chunk["choices"][0]["logprobs"]), + "delta": { + "role": None, + "content": None, + **func_call_dict, + }, + } + ], + ) + prompt += completion_text.strip() + grammar = None + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) + completion_text += "".join( + [chunk["choices"][0]["text"] for chunk in completion] + ) + if ( + "<|from|> assistant" in completion_text + or "<|from|>assistant" in completion_text + ) and tools is not None: + prompt += "\n<|from|>assistant\n<|recipient|>" + tool_index += 1 + else: + # Yield tool_call/function_call stop message + yield llama_types.CreateChatCompletionStreamResponse( + id="chat" + chunk_id, + object="chat.completion.chunk", + created=chunk_created, + model=chunk["model"], + choices=[ + { + "index": 0, + "finish_reason": ( + "tool_calls" + if tools is not None + else "function_call" + ), + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + ) + break + + if stream is not False: + return generate_streaming( + tools=tools, functions=functions, function_call=function_call, prompt=prompt + ) + else: + if version == "v1": + # If no or "auto" tool_choice/function_call + if isinstance(function_call, str) and function_call == "auto": + stops = ["\n", END_ASSISTANT_TOKEN] + # If tool_choice/function_call is provided + elif isinstance(function_call, dict): + prompt += f"{START_FUNCTION_CALL_TOKEN}{function_call['name']}:\n" + stops = END_FUNCTION_CALL_TOKEN + function_call = function_call["name"] + function_calls.append(function_call) + grammar = get_grammar(function_call) + else: + prompt = prompt + stops = ["\n", END_ASSISTANT_TOKEN] + + completion = create_completion(prompt=prompt, stop=stops, grammar=grammar) + completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] + + # If the generation does not involve a function call + if ( + START_FUNCTION_CALL_TOKEN not in prompt + and START_FUNCTION_CALL_TOKEN not in completion_text + ): + completion["usage"]["completion_tokens"] = completion_tokens + return _convert_completion_to_chat(completion, stream=stream) # type: ignore + # If the generation involves a function call in completion, generate the parameters + elif ( + START_FUNCTION_CALL_TOKEN not in prompt + and START_FUNCTION_CALL_TOKEN in completion_text + ): + prompt += ( + completion_text.replace( + f"{START_FUNCTION_CALL_TOKEN} ", START_FUNCTION_CALL_TOKEN + ) + + "\n" + ) + function_calls.append( + completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip() + ) + grammar = get_grammar(function_calls[-1]) + completion = create_completion( + prompt=prompt, stop=END_FUNCTION_CALL_TOKEN, grammar=grammar + ) + completion_tokens += completion["usage"]["completion_tokens"] + function_bodies.append(completion["choices"][0]["text"].strip()) + # If the prompt involves a function call, just append generated parameters to function_bodies + else: + function_bodies.append(completion_text.strip()) + else: + # If tool_choice/function_call is provided + if isinstance(function_call, dict): + prompt += f"{function_call['name']}\n{CONTENT_TOKEN}" + function_call = function_call["name"] + function_calls.append(function_call) + grammar = get_grammar(function_call) + stops = [STOP_TOKEN, FROM_TOKEN] + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) + completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] + function_bodies.append(completion_text.strip()) + # If "auto" or no tool_choice/function_call + elif isinstance(function_call, str) and function_call == "auto": + while True: + # Generate function name first + grammar = None + stops = CONTENT_TOKEN + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) + completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] + function_name = completion_text.strip() + if function_name == "all": + prompt += "all\n<|content|>" + else: + function_call = completion_text.strip() + prompt += f"{function_call}\n<|content|>" + function_calls.append(function_call) + grammar = get_grammar(function_call) + # Generate content + stops = [RECIPIENT_TOKEN, STOP_TOKEN] + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) + completion_text = completion["choices"][0]["text"] + completion_tokens += completion["usage"]["completion_tokens"] + if function_name == "all": + if completion_text.endswith("\n<|from|>assistant\n"): + content += completion_text[: -len("\n<|from|>assistant\n")] + if completion_text.endswith("\n<|from|> assistant\n"): + content += completion_text[-len("\n<|from|> assistant\n")] + else: + content += completion_text + content = content.lstrip() + # Check whether the model wants to generate another turn + if ( + "<|from|> assistant" in completion_text + or "<|from|>assistant" in completion_text + ): + if completion_text.endswith("\n<|from|>assistant\n"): + cleaned_completion_text = completion_text[ + : -len("\n<|from|>assistant\n") + ].strip() + elif completion_text.endswith("\n<|from|> assistant\n"): + cleaned_completion_text = completion_text[ + -len("\n<|from|> assistant\n") + ].strip() + else: + cleaned_completion_text = completion_text.strip() + prompt += f"{cleaned_completion_text}\n<|from|>assistant\n<|recipient|>" + else: + break + else: + function_bodies.append(completion_text.strip()) + # Check whether the model wants to generate another turn + prompt += completion_text.strip() + grammar = None + completion = create_completion( + prompt=prompt, stop=stops, grammar=grammar + ) + completion_tokens += completion["usage"]["completion_tokens"] + if ( + "<|from|> assistant" in completion["choices"][0]["text"] + or "<|from|>assistant" in completion["choices"][0]["text"] + ): + prompt += "\n<|from|>assistant\n<|recipient|>" + else: + break + + assert "usage" in completion + assert len(function_calls) == len(function_bodies) + + tool_calls: List[llama_types.ChatCompletionMessageToolCall] = [] + for function_call, function_body in zip(function_calls, function_bodies): + tool_calls.append( + { + "id": "call_" + + "".join( + [ + random.choice(string.ascii_letters + string.digits) + for _ in range(24) + ] + ), + "type": "function", + "function": { + "name": function_call, + "arguments": function_body, + }, + } + ) + + # TODO: support stream mode + function_call_dict: Union[ + Dict[str, str], + Dict[ + Literal["function_call"], + llama_types.ChatCompletionRequestAssistantMessageFunctionCall, + ], + ] = {} + if len(tool_calls) > 0: + if tools is not None: + function_call_dict["tool_calls"] = tool_calls + else: + function_call_dict["function_call"] = { + "name": tool_calls[0]["function"]["name"], + "arguments": tool_calls[0]["function"]["arguments"], + } + completion["usage"]["completion_tokens"] = completion_tokens + return llama_types.CreateChatCompletionResponse( + id="chat" + completion["id"], + object="chat.completion", + created=completion["created"], + model=completion["model"], + choices=[ + { + "index": 0, + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), + "message": { + "role": "assistant", + "content": None if content == "" else content, + **function_call_dict, + }, + "finish_reason": "tool_calls" if len(tool_calls) > 0 else "stop", + } + ], + usage=completion["usage"], + ) + + +class Llava15ChatHandler: + DEFAULT_SYSTEM_MESSAGE: Optional[str] = ( + "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." + ) + + CHAT_FORMAT = ( + "{% for message in messages %}" + "{% if message.role == 'system' %}" + "{{ message.content }}" + "{% endif %}" + "{% if message.role == 'user' %}" + "{% if message.content is string %}" + "\nUSER: {{ message.content }}" + "{% endif %}" + "{% if message.content is iterable %}" + "\nUSER: " + "{% for content in message.content %}" + "{% if content.type == 'image_url' and content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.type == 'image_url' and content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endfor %}" + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + "{% endif %}" + "{% if message.role == 'assistant' and message.content is not none %}" + "\nASSISTANT: {{ message.content }}" + "{% endif %}" + "{% endfor %}" + "{% if add_generation_prompt %}" + "\nASSISTANT: " + "{% endif %}" + ) + + def __init__(self, clip_model_path: str, verbose: bool = True): + import llama_cpp.llava_cpp as llava_cpp + + self.clip_model_path = clip_model_path + self.verbose = verbose + + self._llava_cpp = llava_cpp # TODO: Fix + self._exit_stack = ExitStack() + self._last_image_embed: Optional[ + llava_cpp.CtypesPointer[llava_cpp.llava_image_embed] + ] = None + self._last_image_hash: Optional[int] = None + + if not os.path.exists(clip_model_path): + raise ValueError(f"Clip model path does not exist: {clip_model_path}") + + with suppress_stdout_stderr(disable=self.verbose): + clip_ctx = self._llava_cpp.clip_model_load(self.clip_model_path.encode(), 0) + + if clip_ctx is None: + raise ValueError(f"Failed to load clip model: {clip_model_path}") + + self.clip_ctx = clip_ctx + + def clip_free(): + with suppress_stdout_stderr(disable=self.verbose): + self._llava_cpp.clip_free(self.clip_ctx) + + self._exit_stack.callback(clip_free) + + def last_image_embed_free(): + with suppress_stdout_stderr(disable=self.verbose): + if self._last_image_embed is not None: + self._llava_cpp.llava_image_embed_free(self._last_image_embed) + self._last_image_embed = None + + self._exit_stack.callback(last_image_embed_free) + + def load_image(self, image_url: str) -> bytes: + return self._load_image(image_url) + + def _embed_image_bytes(self, image_bytes: bytes, n_threads_batch: int = 1): + if ( + self._last_image_embed is not None + and self._last_image_hash is not None + and hash(image_bytes) == self._last_image_hash + ): + return self._last_image_embed + with suppress_stdout_stderr(disable=self.verbose): + # Free the previous image embed + if self._last_image_embed is not None: + self._llava_cpp.llava_image_embed_free(self._last_image_embed) + self._last_image_embed = None + self._last_image_hash = None + embed = self._llava_cpp.llava_image_embed_make_with_bytes( + self.clip_ctx, + n_threads_batch, + (ctypes.c_uint8 * len(image_bytes)).from_buffer( + bytearray(image_bytes) + ), + len(image_bytes), + ) + self._last_image_embed = embed + self._last_image_hash = hash(image_bytes) + return embed + + def __call__( + self, + *, + llama: llama.Llama, + messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, + temperature: float = 0.2, + top_p: float = 0.95, + top_k: int = 40, + min_p: float = 0.05, + typical_p: float = 1.0, + stream: bool = False, + stop: Optional[Union[str, List[str]]] = [], + seed: Optional[int] = None, + response_format: Optional[ + llama_types.ChatCompletionRequestResponseFormat + ] = None, + max_tokens: Optional[int] = None, + presence_penalty: float = 0.0, + frequency_penalty: float = 0.0, + repeat_penalty: float = 1.1, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_tau: float = 5.0, + mirostat_eta: float = 0.1, + model: Optional[str] = None, + logits_processor: Optional[llama.LogitsProcessorList] = None, + grammar: Optional[llama.LlamaGrammar] = None, + logit_bias: Optional[Dict[str, float]] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, + **kwargs, # type: ignore + ) -> Union[ + llama_types.CreateChatCompletionResponse, + Iterator[llama_types.CreateChatCompletionStreamResponse], + ]: + assert self.clip_ctx is not None + + system_prompt = _get_system_message(messages) + if system_prompt == "" and self.DEFAULT_SYSTEM_MESSAGE is not None: + messages = [ + llama_types.ChatCompletionRequestSystemMessage( + role="system", content=self.DEFAULT_SYSTEM_MESSAGE + ) + ] + messages + + image_urls = self.get_image_urls(messages) + template = ImmutableSandboxedEnvironment( + trim_blocks=True, + lstrip_blocks=True, + ).from_string(self.CHAT_FORMAT) + text = template.render( + messages=messages, + add_generation_prompt=True, + eos_token=llama.detokenize([llama.token_eos()]), + bos_token=llama.detokenize([llama.token_bos()]), + ) + split_text = self.split_text_on_image_urls(text, image_urls) + + if self.verbose: + print(text, file=sys.stderr) + + + # Evaluate prompt + llama.reset() + llama._ctx.kv_cache_clear() + for type_, value in split_text: + if type_ == "text": + tokens = llama.tokenize( + value.encode("utf8"), add_bos=False, special=True + ) + if llama.n_tokens + len(tokens) > llama.n_ctx(): + raise ValueError( + f"Prompt exceeds n_ctx: {llama.n_tokens + len(tokens)} > {llama.n_ctx()}" + ) + llama.eval(tokens) + else: + image_bytes = self.load_image(value) + embed = self._embed_image_bytes(image_bytes, llama.context_params.n_threads_batch) + if llama.n_tokens + embed.contents.n_image_pos > llama.n_ctx(): + raise ValueError( + f"Prompt exceeds n_ctx: {llama.n_tokens + embed.contents.n_image_pos} > {llama.n_ctx()}" + ) + n_past = ctypes.c_int(llama.n_tokens) + n_past_p = ctypes.pointer(n_past) + with suppress_stdout_stderr(disable=self.verbose): + self._llava_cpp.llava_eval_image_embed( + llama.ctx, + embed, + llama.n_batch, + n_past_p, + ) + # Required to avoid issues with hf tokenizer + llama.input_ids[llama.n_tokens : n_past.value] = -1 + llama.n_tokens = n_past.value + + # Get prompt tokens to avoid a cache miss + prompt = llama.input_ids[: llama.n_tokens].tolist() + + if response_format is not None and response_format["type"] == "json_object": + grammar = _grammar_for_response_format(response_format) + + # Convert legacy functions to tools + if functions is not None: + tools = [ + { + "type": "function", + "function": function, + } + for function in functions + ] + + # Convert legacy function_call to tool_choice + if function_call is not None: + if isinstance(function_call, str) and ( + function_call == "none" or function_call == "auto" + ): + tool_choice = function_call + if isinstance(function_call, dict) and "name" in function_call: + tool_choice = { + "type": "function", + "function": { + "name": function_call["name"], + }, + } + + tool = None + if ( + tool_choice is not None + and isinstance(tool_choice, dict) + and tools is not None + ): + name = tool_choice["function"]["name"] + tool = next((t for t in tools if t["function"]["name"] == name), None) + if tool is None: + raise ValueError(f"Tool choice '{name}' not found in tools.") + schema = tool["function"]["parameters"] + try: + # create grammar from json schema + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(schema), verbose=llama.verbose + ) + except Exception as e: + if llama.verbose: + print(str(e), file=sys.stderr) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + logprobs=top_logprobs if logprobs else None, + stream=stream, + stop=stop, + seed=seed, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + logit_bias=logit_bias, + ) + if tool is not None: + tool_name = tool["function"]["name"] + return _convert_completion_to_chat_function( + tool_name, completion_or_chunks, stream + ) + return _convert_completion_to_chat(completion_or_chunks, stream=stream) + + @staticmethod + def _load_image(image_url: str) -> bytes: + # TODO: Add Pillow support for other image formats beyond (jpg, png) + if image_url.startswith("data:"): + import base64 + + image_bytes = base64.b64decode(image_url.split(",")[1]) + return image_bytes + else: + import urllib.request + + with urllib.request.urlopen(image_url) as f: + image_bytes = f.read() + return image_bytes + + @staticmethod + def get_image_urls(messages: List[llama_types.ChatCompletionRequestMessage]): + image_urls: List[str] = [] + for message in messages: + if message["role"] == "user": + if message["content"] is None: + continue + for content in message["content"]: + if isinstance(content, dict) and "type" in content: + if content["type"] == "image_url": + if ( + isinstance(content["image_url"], dict) + and "url" in content["image_url"] + ): + image_urls.append(content["image_url"]["url"]) + else: + image_urls.append(content["image_url"]) + return image_urls + + @staticmethod + def split_text_on_image_urls(text: str, image_urls: List[str]): + def find_first(s: str, substrs: List[str]): + for i, substr in enumerate(substrs): + pos = s.find(substr) + if pos != -1: + return pos, i + return None, None + + split_text: List[Tuple[Literal["text", "image_url"], str]] = [] + remaining = text + while remaining: + # Find first image_url + pos, i = find_first(remaining, image_urls) + if pos is not None and i is not None: + if pos > 0: + split_text.append(("text", remaining[:pos])) + split_text.append(("image_url", image_urls[i])) + remaining = remaining[pos + len(image_urls[i]) :] + else: + split_text.append(("text", remaining)) + remaining = "" + return split_text + + @classmethod + def from_pretrained( + cls, + repo_id: str, + filename: Optional[str], + local_dir: Optional[Union[str, os.PathLike[str]]] = None, + local_dir_use_symlinks: Union[bool, Literal["auto"]] = "auto", + cache_dir: Optional[Union[str, os.PathLike[str]]] = None, + **kwargs: Any, + ) -> "Llava15ChatHandler": + import fnmatch + from pathlib import Path + + try: + from huggingface_hub import hf_hub_download, HfFileSystem # type: ignore + from huggingface_hub.utils import validate_repo_id # type: ignore + except ImportError: + raise ImportError( + "Llama.from_pretrained requires the huggingface-hub package. " + "You can install it with `pip install huggingface-hub`." + ) + + validate_repo_id(repo_id) + + hffs = HfFileSystem() + + files = [ + file["name"] if isinstance(file, dict) else file + for file in hffs.ls(repo_id) # type: ignore + ] + + # split each file into repo_id, subfolder, filename + file_list: List[str] = [] + for file in files: + rel_path = Path(file).relative_to(repo_id) + file_list.append(str(rel_path)) + + matching_files = [file for file in file_list if fnmatch.fnmatch(file, filename)] # type: ignore + + if len(matching_files) == 0: + raise ValueError( + f"No file found in {repo_id} that match {filename}\n\n" + f"Available Files:\n{json.dumps(file_list)}" + ) + + if len(matching_files) > 1: + raise ValueError( + f"Multiple files found in {repo_id} matching {filename}\n\n" + f"Available Files:\n{json.dumps(files)}" + ) + + (matching_file,) = matching_files + + subfolder = str(Path(matching_file).parent) + filename = Path(matching_file).name + + # download the file + hf_hub_download( + repo_id=repo_id, + filename=filename, + subfolder=subfolder, + local_dir=cast(Union[str, Path, None], local_dir), + local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cast(Union[str, Path, None], cache_dir), + ) + + if local_dir is None: + model_path = hf_hub_download( + repo_id=repo_id, + filename=filename, + subfolder=subfolder, + local_dir=local_dir, + local_dir_use_symlinks=local_dir_use_symlinks, + cache_dir=cast(Union[str, Path, None], cache_dir), + local_files_only=True, + ) + else: + model_path = os.path.join(local_dir, filename) + + return cls( + clip_model_path=model_path, + **kwargs, + ) + + +class ObsidianChatHandler(Llava15ChatHandler): + # Prompt Format + # The model followed ChatML format. However, with ### as the seperator + + # <|im_start|>user + # What is this sign about?\n + # ### + # <|im_start|>assistant + # The sign is about bullying, and it is placed on a black background with a red background. + # ### + + CHAT_FORMAT = ( + "{% for message in messages %}" + # System message + "{% if message.role == 'system' %}" + "<|im_start|>system\n" + "{{ message.content }}\n" + "###\n" + "{% endif %}" + # User message + "{% if message.role == 'user' %}" + "<|im_start|>user\n" + "{% if message.content is string %}" + "{{ message.content }}" + "{% endif %}" + "{% if message.content is iterable %}" + "{% for content in message.content %}" + "{% if content.type == 'image_url' and content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.type == 'image_url' and content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endfor %}" + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + "###\n" + "{% endif %}" + # Assistant message + "{% if message.role == 'assistant' %}" + "<|im_start|>assistant\n" + "{{ message.content }}" + "###\n" + "{% endif %}" + "{% endfor %}" + # Generation prompt + "{% if add_generation_prompt %}" + "<|im_start|>assistant\n" + "{% endif %}" + ) + + +class MoondreamChatHandler(Llava15ChatHandler): + # Chat Format: + # f"\n\n{chat_history}Question: {question}\n\nAnswer:" + CHAT_FORMAT = ( + "{% for message in messages %}" + "{% if message.role == 'user' %}" + "{% if message.content is iterable %}" + # + "{% for content in message.content %}" + "{% if content.type == 'image_url' %}" + "{% if content.image_url is string %}" + "{{ content.image_url }}\n\n" + "{% endif %}" + "{% if content.image_url is mapping %}" + "{{ content.image_url.url }}\n\n" + "{% endif %}" + "{% endif %}" + "{% endfor %}" + # Question: + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "Question: {{ content.text }}\n\n" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + # Question: + "{% if message.content is string %}" + "Question: {{ message.content }}\n\n" + "{% endif %}" + "{% endif %}" + # Answer: + "{% if message.role == 'assistant' %}" + "Answer:{{ message.content }}\n\n" + "{% endif %}" + "{% endfor %}" + # Generation prompt + "{% if add_generation_prompt %}" + "Answer:" + "{% endif %}" + ) + + +class Llava16ChatHandler(Llava15ChatHandler): + DEFAULT_SYSTEM_MESSAGE = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. " + + # Example prompt + # "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. USER: \nWhat is shown in this image? ASSISTANT:" + + CHAT_FORMAT = ( + "{% for message in messages %}" + "{% if message.role == 'system' %}" + "{{ message.content }}" + "{% endif %}" + "{% if message.role == 'user' %}" + "{% if message.content is iterable %}" + # + "{% for content in message.content %}" + "{% if content.type == 'image_url' %}" + "{% if content.image_url is string %}" + "{{ content.image_url }}\n" + "{% endif %}" + "{% if content.image_url is mapping %}" + "{{ content.image_url.url }}\n" + "{% endif %}" + "{% endif %}" + "{% endfor %}" + # Question: + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + # Question: + "{% if message.content is string %}" + "{{ message.content }}" + "{% endif %}" + "{% endif %}" + # Answer: + "{% if message.role == 'assistant' %}" + "{{ message.content }}" + "{% endif %}" + "{% endfor %}" + # Generation prompt + "{% if add_generation_prompt %}" + "Answer:" + "{% endif %}" + ) + + +class NanoLlavaChatHandler(Llava15ChatHandler): + # Prompt Format + # The model follow the ChatML standard, however, without \n at the end of <|im_end|>: + + # <|im_start|>system + # Answer the question<|im_end|><|im_start|>user + # + # What is the picture about?<|im_end|><|im_start|>assistant + DEFAULT_SYSTEM_MESSAGE = "Answer the question" + + CHAT_FORMAT = ( + "{% for message in messages %}" + # System message + "{% if message.role == 'system' %}" + "<|im_start|>system\n" + "{{ message.content }}" + "<|im_end|>" + "{% endif %}" + # User message + "{% if message.role == 'user' %}" + "<|im_start|>user\n" + "{% if message.content is string %}" + "{{ message.content }}" + "{% endif %}" + "{% if message.content is iterable %}" + "{% for content in message.content %}" + "{% if content.type == 'image_url' and content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.type == 'image_url' and content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endfor %}" + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + "<|im_end|>" + "{% endif %}" + # Assistant message + "{% if message.role == 'assistant' %}" + "<|im_start|>assistant\n" + "{{ message.content }}" + "<|im_end|>" + "{% endif %}" + "{% endfor %}" + # Generation prompt + "{% if add_generation_prompt %}" + "<|im_start|>assistant\n" + "{% endif %}" + ) + + +class Llama3VisionAlphaChatHandler(Llava15ChatHandler): + # question = "" + q + + # prompt = f"<|start_header_id|>user<|end_header_id|>\n\n{question}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" + DEFAULT_SYSTEM_MESSAGE = None + + CHAT_FORMAT = ( + "{% for message in messages %}" + "<|start_header_id|>" + "{% if message.role == 'user' %}" + "user<|end_header_id|>\n\n" + "{% if message.content is iterable %}" + # + "{% for content in message.content %}" + "{% if content.type == 'image_url' %}" + "{% if content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endif %}" + "{% endfor %}" + # Question: + "{% for content in message.content %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + # Question: + "{% if message.content is string %}" + "{{ message.content }}" + "{% endif %}" + "{% endif %}" + # Answer: + "{% if message.role == 'assistant' %}" + "assistant<|end_header_id|>\n\n" + "{{ message.content }}" + "{% endif %}" + "<|eot_id|>" + "{% endfor %}" + # Generation prompt + "{% if add_generation_prompt %}" + "<|start_header_id|>assistant<|end_header_id|>\n\n" + "{% endif %}" + ) + + +# alias +Llama3VisionAlpha = Llama3VisionAlphaChatHandler + + +class MiniCPMv26ChatHandler(Llava15ChatHandler): + DEFAULT_SYSTEM_MESSAGE = "You are a helpful assistant." + + CHAT_FORMAT = ( + "{% for message in messages %}" + "{% if loop.first and messages[0]['role'] != 'system' %}" + "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n" + "{% endif %}" + "<|im_start|>{{ message['role'] }}\n" + "{% if message['content'] is iterable %}" + "{% for content in message['content'] %}" + "{% if content.type == 'image_url' %}" + "{% if content.image_url is string %}" + "{{ content.image_url }}" + "{% endif %}" + "{% if content.image_url is mapping %}" + "{{ content.image_url.url }}" + "{% endif %}" + "{% endif %}" + "{% endfor %}" + + "{% for content in message['content'] %}" + "{% if content.type == 'text' %}" + "{{ content.text }}" + "{% endif %}" + "{% endfor %}" + "{% endif %}" + "{% if message['content'] is string %}" + "{{ message['content'] }}" + "{% endif %}" + "<|im_end|>\n" + "{% endfor %}" + "{% if add_generation_prompt %}" + "<|im_start|>assistant\n" + "{% endif %}" + ) + + +@register_chat_completion_handler("chatml-function-calling") +def chatml_function_calling( + llama: llama.Llama, + messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, + temperature: float = 0.2, + top_p: float = 0.95, + top_k: int = 40, + min_p: float = 0.05, + typical_p: float = 1.0, + stream: bool = False, + stop: Optional[Union[str, List[str]]] = [], + response_format: Optional[llama_types.ChatCompletionRequestResponseFormat] = None, + max_tokens: Optional[int] = None, + presence_penalty: float = 0.0, + frequency_penalty: float = 0.0, + repeat_penalty: float = 1.1, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_tau: float = 5.0, + mirostat_eta: float = 0.1, + model: Optional[str] = None, + logits_processor: Optional[llama.LogitsProcessorList] = None, + grammar: Optional[llama.LlamaGrammar] = None, + logprobs: Optional[bool] = None, + top_logprobs: Optional[int] = None, + **kwargs, # type: ignore +) -> Union[ + llama_types.CreateChatCompletionResponse, + Iterator[llama_types.CreateChatCompletionStreamResponse], +]: + function_calling_template = ( + "{% for message in messages %}" + "<|im_start|>{{ message.role }}\n" + # System message + "{% if message.role == 'system' %}" + "{{ message.content }}" + "{% if tool_calls %}" + "\n\nYou have access to the following functions:\n" + "{% for tool in tools %}" + "\nfunctions.{{ tool.function.name }}:\n" + "{{ tool.function.parameters | tojson }}" + "\n{% endfor %}" + "\n\nYou can respond to users messages with either a single message or one or more function calls." + "\n\nTo respond with a message begin the message with 'message:', use the following format:" + "\n\nmessage:" + "\n" + "\n\nTo respond with one or more function calls begin the message with 'functions.:', use the following format:" + "\n\nfunctions.:" + '\n{ "arg1": "value1", "arg2": "value2" }' + "\nfunctions.:" + '\n{ "arg1": "value1", "arg2": "value2" }' + "{% endif %}" + "<|im_end|>\n" + "{% endif %}" + # User message + "{% if message.role == 'user' %}" + "{{ message.content }}" + "<|im_end|>\n" + "{% endif %}" + # Assistant message + "{% if message.role == 'assistant' %}" + ## Reglar message + "{% if message.content and message.content | length > 0 %}" + "{% if tool_calls %}" + "message:\n" + "{% endif %}" + "{{ message.content }}" + "<|im_end|>\n" + "{% endif %}" + ## Function calls + "{% if 'tool_calls' in message %}" + "{% for tool_call in message.tool_calls %}" + "functions.{{ tool_call.function.name }}:\n" + "{{ tool_call.function.arguments }}" + "{% endfor %}" + "<|im_end|>\n" + "{% endif %}" + "{% endif %}" + "{% endfor %}" + "{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}" + ) + template_renderer = ImmutableSandboxedEnvironment( + autoescape=jinja2.select_autoescape(["html", "xml"]), + undefined=jinja2.StrictUndefined, + ).from_string(function_calling_template) + + # Convert legacy functions to tools + if functions is not None: + tools = [ + { + "type": "function", + "function": function, + } + for function in functions + ] + + # Convert legacy function_call to tool_choice + if function_call is not None: + if isinstance(function_call, str) and ( + function_call == "none" or function_call == "auto" + ): + tool_choice = function_call + if isinstance(function_call, dict) and "name" in function_call: + tool_choice = { + "type": "function", + "function": { + "name": function_call["name"], + }, + } + + stop = ( + [stop, "<|im_end|>"] + if isinstance(stop, str) + else stop + ["<|im_end|>"] if stop else ["<|im_end|>"] + ) + + # Case 1: No tool choice by user + if ( + tool_choice is None + or (isinstance(tool_choice, str) and tool_choice == "none") + or tools is None + or len(tools) == 0 + ): + prompt = template_renderer.render( + messages=messages, + tools=[], + tool_calls=None, + add_generation_prompt=True, + ) + + if response_format is not None and response_format["type"] == "json_object": + grammar = _grammar_for_response_format(response_format) + + return _convert_completion_to_chat( + llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + logprobs=top_logprobs if logprobs else None, + ), + stream=stream, + ) + + # Case 2: Tool choice by user + if isinstance(tool_choice, dict): + tool_name = tool_choice["function"]["name"] + tool = next( + (tool for tool in tools if tool["function"]["name"] == tool_name), None + ) + if tool is None: + raise ValueError(f"Tool with name '{tool_name}' not found in tools") + prompt = template_renderer.render( + messages=messages, + tools=tools, + tool_calls=True, + add_generation_prompt=True, + ) + prompt += f"functions.{tool_name}:\n" + try: + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(tool["function"]["parameters"]), verbose=llama.verbose + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + if llama.verbose: + print( + "Failed to parse function body as JSON schema, falling back to default grammar" + ) + print(e) + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ) + return _convert_completion_to_chat_function( + tool_name, completion_or_chunks, stream + ) + + # Case 3: Automatic tool choice + assert isinstance(tool_choice, str) and tool_choice == "auto" + function_names = " | ".join( + [f'''"functions.{tool['function']['name']}:"''' for tool in tools] + ) + initial_gbnf_tool_grammar = ( + """root ::= functions | "message:"\n""" + f"""functions ::= {function_names}\n""" + ) + follow_up_gbnf_tool_grammar = ( + """root ::= functions | "<|im_end|>"\n""" + f"""functions ::= {function_names}\n""" + ) + prompt = template_renderer.render( + messages=messages, + tools=tools, + tool_calls=True, + add_generation_prompt=True, + ) + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=0, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=False, + stop=[":"], + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=llama_grammar.LlamaGrammar.from_string( + initial_gbnf_tool_grammar, verbose=llama.verbose + ), + ) + completion: llama_types.CreateCompletionResponse = completion_or_chunks # type: ignore + text = completion["choices"][0]["text"] + if "message" in text: + return _convert_completion_to_chat( + llama.create_completion( + prompt=prompt + "message:\n", + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=["<|im_end|>"], + logprobs=top_logprobs if logprobs else None, + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=llama_grammar.LlamaGrammar.from_string( + follow_up_gbnf_tool_grammar, verbose=llama.verbose + ), + ), + stream=stream, + ) + + # One or more function calls + tool_name = text[len("functions.") :] + tool = next((tool for tool in tools if tool["function"]["name"] == tool_name), None) + if not stream: + completions: List[llama_types.CreateCompletionResponse] = [] + completions_tool_name: List[str] = [] + while tool is not None: + prompt += f"functions.{tool_name}:\n" + try: + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(tool["function"]["parameters"]), verbose=llama.verbose + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + if llama.verbose: + print( + "Failed to parse function body as JSON schema, falling back to default grammar" + ) + print(e) + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=False, + stop=stop, + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ) + completion_or_chunks = cast( + llama_types.CreateCompletionResponse, completion_or_chunks + ) + completions.append(completion_or_chunks) + completions_tool_name.append(tool_name) + prompt += completion_or_chunks["choices"][0]["text"] + prompt += "\n" + + response = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=False, + stop=stop, + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=llama_grammar.LlamaGrammar.from_string( + follow_up_gbnf_tool_grammar, verbose=llama.verbose + ), + ) + response = cast(llama_types.CreateCompletionResponse, response) + + tool_name = response["choices"][0]["text"][len("functions.") :] + tool = next( + (tool for tool in tools if tool["function"]["name"] == tool_name), None + ) + + # Merge completions + function_call_dict: Union[ + Dict[str, str], + Dict[ + Literal["function_call"], + llama_types.ChatCompletionRequestAssistantMessageFunctionCall, + ], + ] = ( + { + "function_call": { + "name": tool_name, + "arguments": completions[0]["choices"][0]["text"], + } + } + if len(completions) == 1 + else {} + ) + return { + "id": "chat" + completion["id"], + "object": "chat.completion", + "created": completion["created"], + "model": completion["model"], + "choices": [ + { + "finish_reason": "tool_calls", + "index": 0, + "logprobs": _convert_text_completion_logprobs_to_chat(completion["choices"][0]["logprobs"]), + "message": { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_" + + f"_{i}_" + + tool_name + + "_" + + completion["id"], + "type": "function", + "function": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + } + for i, (tool_name, completion) in enumerate( + zip(completions_tool_name, completions) + ) + ], + **function_call_dict, + }, + } + ], + "usage": { + "completion_tokens": sum( + ( + completion["usage"]["completion_tokens"] + if "usage" in completion + else 0 + ) + for completion in completions + ), + "prompt_tokens": sum( + completion["usage"]["prompt_tokens"] if "usage" in completion else 0 + for completion in completions + ), + "total_tokens": sum( + completion["usage"]["total_tokens"] if "usage" in completion else 0 + for completion in completions + ), + }, + } + + raise ValueError("Automatic streaming tool choice is not supported") diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 700567f4a..63de3a93a 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -1,140 +1,297 @@ -import sys +from __future__ import annotations + import os import ctypes -from ctypes import ( - c_bool, - c_char_p, - c_int, - c_int8, - c_int32, - c_uint8, - c_uint32, - c_size_t, - c_float, - c_double, - c_void_p, - POINTER, - _Pointer, # type: ignore - Structure, - Array, -) import pathlib -from typing import List, Union +from typing import ( + Callable, + Union, + NewType, + Optional, + TYPE_CHECKING, +) -# Load the library -def _load_shared_library(lib_base_name: str): - # Construct the paths to the possible shared library names - _base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) - # Searching for the library in the current directory under the name "libllama" (default name - # for llamacpp) and "llama" (default name for this repo) - _lib_paths: List[pathlib.Path] = [] - # Determine the file extension based on the platform - if sys.platform.startswith("linux"): - _lib_paths += [ - _base_path / f"lib{lib_base_name}.so", - ] - elif sys.platform == "darwin": - _lib_paths += [ - _base_path / f"lib{lib_base_name}.so", - _base_path / f"lib{lib_base_name}.dylib", - ] - elif sys.platform == "win32": - _lib_paths += [ - _base_path / f"{lib_base_name}.dll", - _base_path / f"lib{lib_base_name}.dll", - ] - else: - raise RuntimeError("Unsupported platform") - - if "LLAMA_CPP_LIB" in os.environ: - lib_base_name = os.environ["LLAMA_CPP_LIB"] - _lib = pathlib.Path(lib_base_name) - _base_path = _lib.parent.resolve() - _lib_paths = [_lib.resolve()] - - cdll_args = dict() # type: ignore - # Add the library directory to the DLL search path on Windows (if needed) - if sys.platform == "win32" and sys.version_info >= (3, 8): - os.add_dll_directory(str(_base_path)) - if "CUDA_PATH" in os.environ: - os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin")) - os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib")) - cdll_args["winmode"] = ctypes.RTLD_GLOBAL - - # Try to load the shared library, handling potential errors - for _lib_path in _lib_paths: - if _lib_path.exists(): - try: - return ctypes.CDLL(str(_lib_path), **cdll_args) - except Exception as e: - raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}") - - raise FileNotFoundError( - f"Shared library with base name '{lib_base_name}' not found" +from llama_cpp._ctypes_extensions import ( + load_shared_library, + byref, + ctypes_function_for_shared_library, +) + +if TYPE_CHECKING: + from llama_cpp._ctypes_extensions import ( + CtypesCData, + CtypesArray, + CtypesPointer, + CtypesVoidPointer, + CtypesRef, + CtypesPointerOrRef, + CtypesFuncPointer, ) # Specify the base name of the shared library to load _lib_base_name = "llama" - +_override_base_path = os.environ.get("LLAMA_CPP_LIB_PATH") +_base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" if _override_base_path is None else pathlib.Path(_override_base_path) # Load the library -_lib = _load_shared_library(_lib_base_name) +_lib = load_shared_library(_lib_base_name, _base_path) + +ctypes_function = ctypes_function_for_shared_library(_lib) + + +# from ggml.h +# // NOTE: always add types at the end of the enum to keep backward compatibility +# enum ggml_type { +# GGML_TYPE_F32 = 0, +# GGML_TYPE_F16 = 1, +# GGML_TYPE_Q4_0 = 2, +# GGML_TYPE_Q4_1 = 3, +# // GGML_TYPE_Q4_2 = 4, support has been removed +# // GGML_TYPE_Q4_3 = 5, support has been removed +# GGML_TYPE_Q5_0 = 6, +# GGML_TYPE_Q5_1 = 7, +# GGML_TYPE_Q8_0 = 8, +# GGML_TYPE_Q8_1 = 9, +# GGML_TYPE_Q2_K = 10, +# GGML_TYPE_Q3_K = 11, +# GGML_TYPE_Q4_K = 12, +# GGML_TYPE_Q5_K = 13, +# GGML_TYPE_Q6_K = 14, +# GGML_TYPE_Q8_K = 15, +# GGML_TYPE_IQ2_XXS = 16, +# GGML_TYPE_IQ2_XS = 17, +# GGML_TYPE_IQ3_XXS = 18, +# GGML_TYPE_IQ1_S = 19, +# GGML_TYPE_IQ4_NL = 20, +# GGML_TYPE_IQ3_S = 21, +# GGML_TYPE_IQ2_S = 22, +# GGML_TYPE_IQ4_XS = 23, +# GGML_TYPE_I8 = 24, +# GGML_TYPE_I16 = 25, +# GGML_TYPE_I32 = 26, +# GGML_TYPE_I64 = 27, +# GGML_TYPE_F64 = 28, +# GGML_TYPE_IQ1_M = 29, +# GGML_TYPE_COUNT, +# }; +GGML_TYPE_F32 = 0 +GGML_TYPE_F16 = 1 +GGML_TYPE_Q4_0 = 2 +GGML_TYPE_Q4_1 = 3 +GGML_TYPE_Q5_0 = 6 +GGML_TYPE_Q5_1 = 7 +GGML_TYPE_Q8_0 = 8 +GGML_TYPE_Q8_1 = 9 +GGML_TYPE_Q2_K = 10 +GGML_TYPE_Q3_K = 11 +GGML_TYPE_Q4_K = 12 +GGML_TYPE_Q5_K = 13 +GGML_TYPE_Q6_K = 14 +GGML_TYPE_Q8_K = 15 +GGML_TYPE_IQ2_XXS = 16 +GGML_TYPE_IQ2_XS = 17 +GGML_TYPE_IQ3_XXS = 18 +GGML_TYPE_IQ1_S = 19 +GGML_TYPE_IQ4_NL = 20 +GGML_TYPE_IQ3_S = 21 +GGML_TYPE_IQ2_S = 22 +GGML_TYPE_IQ4_XS = 23 +GGML_TYPE_I8 = 24 +GGML_TYPE_I16 = 25 +GGML_TYPE_I32 = 26 +GGML_TYPE_I64 = 27 +GGML_TYPE_F64 = 28 +GGML_TYPE_IQ1_M = 29 +GGML_TYPE_COUNT = 30 + +# from ggml-backend.h +# typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); +ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE( + ctypes.c_bool, ctypes.c_void_p, ctypes.c_bool, ctypes.c_void_p +) -# Misc -c_float_p = POINTER(c_float) -c_uint8_p = POINTER(c_uint8) -c_size_t_p = POINTER(c_size_t) +# // Abort callback +# // If not NULL, called before ggml computation +# // If it returns true, the computation is aborted +# typedef bool (*ggml_abort_callback)(void * data); +ggml_abort_callback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_void_p) # llama.h bindings -GGML_USE_CUBLAS = hasattr(_lib, "ggml_init_cublas") -GGML_CUDA_MAX_DEVICES = 16 -LLAMA_MAX_DEVICES = GGML_CUDA_MAX_DEVICES if GGML_USE_CUBLAS else 1 +_lib.llama_max_devices.argtypes = [] +_lib.llama_max_devices.restype = ctypes.c_size_t + +LLAMA_MAX_DEVICES = _lib.llama_max_devices() # define LLAMA_DEFAULT_SEED 0xFFFFFFFF LLAMA_DEFAULT_SEED = 0xFFFFFFFF -# define LLAMA_MAX_RNG_STATE (64*1024) -LLAMA_MAX_RNG_STATE = 64 * 1024 +# define LLAMA_TOKEN_NULL -1 +LLAMA_TOKEN_NULL = -1 -#define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla' +# define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla' LLAMA_FILE_MAGIC_GGLA = 0x67676C61 # define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn' LLAMA_FILE_MAGIC_GGSN = 0x6767736E +# define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq' +LLAMA_FILE_MAGIC_GGSQ = 0x67677371 + # define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN -# define LLAMA_SESSION_VERSION 3 -LLAMA_SESSION_VERSION = 3 +# define LLAMA_SESSION_VERSION 9 +LLAMA_SESSION_VERSION = 9 +# define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ +LLAMA_STATE_SEQ_MAGIC = LLAMA_FILE_MAGIC_GGSQ +# define LLAMA_STATE_SEQ_VERSION 2 +LLAMA_STATE_SEQ_VERSION = 2 + +# struct llama_vocab; +llama_vocab_p = NewType("llama_vocab_p", int) +llama_vocab_p_ctypes = ctypes.c_void_p # struct llama_model; -llama_model_p = c_void_p +llama_model_p = NewType("llama_model_p", int) +llama_model_p_ctypes = ctypes.c_void_p # struct llama_context; -llama_context_p = c_void_p +llama_context_p = NewType("llama_context_p", int) +llama_context_p_ctypes = ctypes.c_void_p + +# # struct llama_sampler; +# llama_sampler_p = NewType("llama_sampler_p", int) +# llama_sampler_p_ctypes = ctypes.c_void_p +# struct llama_kv_cache; +llama_kv_cache_p = NewType("llama_kv_cache_p", int) +llama_kv_cache_p_ctypes = ctypes.c_void_p # typedef int32_t llama_pos; -llama_pos = c_int32 +llama_pos = ctypes.c_int32 # typedef int32_t llama_token; -llama_token = c_int32 -llama_token_p = POINTER(llama_token) +llama_token = ctypes.c_int32 +llama_token_p = ctypes.POINTER(llama_token) # typedef int32_t llama_seq_id; -llama_seq_id = c_int32 +llama_seq_id = ctypes.c_int32 # enum llama_vocab_type { -# LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece -# LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding +# LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab +# LLAMA_VOCAB_TYPE_SPM = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback +# LLAMA_VOCAB_TYPE_BPE = 2, // GPT-2 tokenizer based on byte-level BPE +# LLAMA_VOCAB_TYPE_WPM = 3, // BERT tokenizer based on WordPiece +# LLAMA_VOCAB_TYPE_UGM = 4, // T5 tokenizer based on Unigram +# LLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization +# }; +LLAMA_VOCAB_TYPE_NONE = 0 +"""For models without vocab""" +LLAMA_VOCAB_TYPE_SPM = 1 +"""LLaMA tokenizer based on byte-level BPE with byte fallback""" +LLAMA_VOCAB_TYPE_BPE = 2 +"""GPT-2 tokenizer based on byte-level BPE""" +LLAMA_VOCAB_TYPE_WPM = 3 +"""BERT tokenizer based on WordPiece""" +LLAMA_VOCAB_TYPE_UGM = 4 +"""T5 tokenizer based on Unigram""" +LLAMA_VOCAB_TYPE_RWKV = 5 +"""RWKV tokenizer based on greedy tokenization""" + + +# // pre-tokenization types +# enum llama_vocab_pre_type { +# LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0, +# LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1, +# LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM = 2, +# LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3, +# LLAMA_VOCAB_PRE_TYPE_FALCON = 4, +# LLAMA_VOCAB_PRE_TYPE_MPT = 5, +# LLAMA_VOCAB_PRE_TYPE_STARCODER = 6, +# LLAMA_VOCAB_PRE_TYPE_GPT2 = 7, +# LLAMA_VOCAB_PRE_TYPE_REFACT = 8, +# LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9, +# LLAMA_VOCAB_PRE_TYPE_STABLELM2 = 10, +# LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11, +# LLAMA_VOCAB_PRE_TYPE_OLMO = 12, +# LLAMA_VOCAB_PRE_TYPE_DBRX = 13, +# LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, +# LLAMA_VOCAB_PRE_TYPE_PORO = 15, +# LLAMA_VOCAB_PRE_TYPE_CHATGLM3 = 16, +# LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17, +# LLAMA_VOCAB_PRE_TYPE_VIKING = 18, +# LLAMA_VOCAB_PRE_TYPE_JAIS = 19, +# LLAMA_VOCAB_PRE_TYPE_TEKKEN = 20, +# LLAMA_VOCAB_PRE_TYPE_SMOLLM = 21, +# LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22, +# LLAMA_VOCAB_PRE_TYPE_BLOOM = 23, +# LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24, +# LLAMA_VOCAB_PRE_TYPE_EXAONE = 25, +# LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, +# LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, +# LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, +# LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, +# LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, +# LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, +# LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, +# LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33, +# LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34, +# }; +LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 +LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 +LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM = 2 +LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3 +LLAMA_VOCAB_PRE_TYPE_FALCON = 4 +LLAMA_VOCAB_PRE_TYPE_MPT = 5 +LLAMA_VOCAB_PRE_TYPE_STARCODER = 6 +LLAMA_VOCAB_PRE_TYPE_GPT2 = 7 +LLAMA_VOCAB_PRE_TYPE_REFACT = 8 +LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9 +LLAMA_VOCAB_PRE_TYPE_STABLELM2 = 10 +LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11 +LLAMA_VOCAB_PRE_TYPE_OLMO = 12 +LLAMA_VOCAB_PRE_TYPE_DBRX = 13 +LLAMA_VOCAB_PRE_TYPE_SMAUG = 14 +LLAMA_VOCAB_PRE_TYPE_PORO = 15 +LLAMA_VOCAB_PRE_TYPE_CHATGLM3 = 16 +LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17 +LLAMA_VOCAB_PRE_TYPE_VIKING = 18 +LLAMA_VOCAB_PRE_TYPE_JAIS = 19 +LLAMA_VOCAB_PRE_TYPE_TEKKEN = 20 +LLAMA_VOCAB_PRE_TYPE_SMOLLM = 21 +LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22 +LLAMA_VOCAB_PRE_TYPE_BLOOM = 23 +LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24 +LLAMA_VOCAB_PRE_TYPE_EXAONE = 25 +LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26 +LLAMA_VOCAB_PRE_TYPE_MINERVA = 27 +LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28 +LLAMA_VOCAB_PRE_TYPE_GPT4O = 29 +LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30 +LLAMA_VOCAB_PRE_TYPE_TRILLION = 31 +LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32 +LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33 +LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34 + + +# // note: these values should be synchronized with ggml_rope +# // TODO: maybe move this enum to ggml.h (ggml_rope_type) +# enum llama_rope_type { +# LLAMA_ROPE_TYPE_NONE = -1, +# LLAMA_ROPE_TYPE_NORM = 0, +# LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX, +# LLAMA_ROPE_TYPE_MROPE = GGML_ROPE_TYPE_MROPE, +# LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION, # }; -LLAMA_VOCAB_TYPE_SPM = 0 -LLAMA_VOCAB_TYPE_BPE = 1 +LLAMA_ROPE_TYPE_NONE = -1 +LLAMA_ROPE_TYPE_NORM = 0 +LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX = 2 +LLAMA_ROPE_TYPE_MROPE = GGML_ROPE_TYPE_MROPE = 8 +LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION = 24 -# enum llama_token_type { +# enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file # LLAMA_TOKEN_TYPE_UNDEFINED = 0, # LLAMA_TOKEN_TYPE_NORMAL = 1, # LLAMA_TOKEN_TYPE_UNKNOWN = 2, @@ -152,13 +309,39 @@ def _load_shared_library(lib_base_name: str): LLAMA_TOKEN_TYPE_BYTE = 6 +# enum llama_token_attr { +# LLAMA_TOKEN_ATTR_UNDEFINED = 0, +# LLAMA_TOKEN_ATTR_UNKNOWN = 1 << 0, +# LLAMA_TOKEN_ATTR_UNUSED = 1 << 1, +# LLAMA_TOKEN_ATTR_NORMAL = 1 << 2, +# LLAMA_TOKEN_ATTR_CONTROL = 1 << 3, // SPECIAL? +# LLAMA_TOKEN_ATTR_USER_DEFINED = 1 << 4, +# LLAMA_TOKEN_ATTR_BYTE = 1 << 5, +# LLAMA_TOKEN_ATTR_NORMALIZED = 1 << 6, +# LLAMA_TOKEN_ATTR_LSTRIP = 1 << 7, +# LLAMA_TOKEN_ATTR_RSTRIP = 1 << 8, +# LLAMA_TOKEN_ATTR_SINGLE_WORD = 1 << 9, +# }; +LLAMA_TOKEN_ATTR_UNDEFINED = 0 +LLAMA_TOKEN_ATTR_UNKNOWN = 1 << 0 +LLAMA_TOKEN_ATTR_UNUSED = 1 << 1 +LLAMA_TOKEN_ATTR_NORMAL = 1 << 2 +LLAMA_TOKEN_ATTR_CONTROL = 1 << 3 +LLAMA_TOKEN_ATTR_USER_DEFINED = 1 << 4 +LLAMA_TOKEN_ATTR_BYTE = 1 << 5 +LLAMA_TOKEN_ATTR_NORMALIZED = 1 << 6 +LLAMA_TOKEN_ATTR_LSTRIP = 1 << 7 +LLAMA_TOKEN_ATTR_RSTRIP = 1 << 8 +LLAMA_TOKEN_ATTR_SINGLE_WORD = 1 << 9 + + # // model file types # enum llama_ftype { # LLAMA_FTYPE_ALL_F32 = 0, # LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors -# LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 +# // LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 # // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed # // LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed # LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors @@ -173,14 +356,32 @@ def _load_shared_library(lib_base_name: str): # LLAMA_FTYPE_MOSTLY_Q5_K_S = 16, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q5_K_M = 17, // except 1d tensors # LLAMA_FTYPE_MOSTLY_Q6_K = 18, // except 1d tensors - +# LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ3_XS = 22, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ1_S = 24, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ4_NL = 25, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ3_S = 26, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ3_M = 27, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ2_S = 28, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ2_M = 29, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_BF16 = 32, // except 1d tensors +# //LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // removed from gguf files, use Q4_0 and runtime repack +# //LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // removed from gguf files, use Q4_0 and runtime repack +# //LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // removed from gguf files, use Q4_0 and runtime repack +# LLAMA_FTYPE_MOSTLY_TQ1_0 = 36, // except 1d tensors +# LLAMA_FTYPE_MOSTLY_TQ2_0 = 37, // except 1d tensors +# # LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file # }; LLAMA_FTYPE_ALL_F32 = 0 LLAMA_FTYPE_MOSTLY_F16 = 1 LLAMA_FTYPE_MOSTLY_Q4_0 = 2 LLAMA_FTYPE_MOSTLY_Q4_1 = 3 -LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4 LLAMA_FTYPE_MOSTLY_Q8_0 = 7 LLAMA_FTYPE_MOSTLY_Q5_0 = 8 LLAMA_FTYPE_MOSTLY_Q5_1 = 9 @@ -193,20 +394,75 @@ def _load_shared_library(lib_base_name: str): LLAMA_FTYPE_MOSTLY_Q5_K_S = 16 LLAMA_FTYPE_MOSTLY_Q5_K_M = 17 LLAMA_FTYPE_MOSTLY_Q6_K = 18 +LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19 +LLAMA_FTYPE_MOSTLY_IQ2_XS = 20 +LLAMA_FTYPE_MOSTLY_Q2_K_S = 21 +LLAMA_FTYPE_MOSTLY_IQ3_XS = 22 +LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23 +LLAMA_FTYPE_MOSTLY_IQ1_S = 24 +LLAMA_FTYPE_MOSTLY_IQ4_NL = 25 +LLAMA_FTYPE_MOSTLY_IQ3_S = 26 +LLAMA_FTYPE_MOSTLY_IQ3_M = 27 +LLAMA_FTYPE_MOSTLY_IQ2_S = 28 +LLAMA_FTYPE_MOSTLY_IQ2_M = 29 +LLAMA_FTYPE_MOSTLY_IQ4_XS = 30 +LLAMA_FTYPE_MOSTLY_IQ1_M = 31 +LLAMA_FTYPE_MOSTLY_BF16 = 32 +# LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33 +# LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34 +# LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35 +LLAMA_FTYPE_MOSTLY_TQ1_0 = 36 +LLAMA_FTYPE_MOSTLY_TQ2_0 = 37 LLAMA_FTYPE_GUESSED = 1024 # enum llama_rope_scaling_type { -# LLAMA_ROPE_SCALING_UNSPECIFIED = -1, -# LLAMA_ROPE_SCALING_NONE = 0, -# LLAMA_ROPE_SCALING_LINEAR = 1, -# LLAMA_ROPE_SCALING_YARN = 2, -# LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN, +# LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1, +# LLAMA_ROPE_SCALING_TYPE_NONE = 0, +# LLAMA_ROPE_SCALING_TYPE_LINEAR = 1, +# LLAMA_ROPE_SCALING_TYPE_YARN = 2, +# LLAMA_ROPE_SCALING_TYPE_LONGROPE = 3, +# LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN, +# }; +LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1 +LLAMA_ROPE_SCALING_TYPE_NONE = 0 +LLAMA_ROPE_SCALING_TYPE_LINEAR = 1 +LLAMA_ROPE_SCALING_TYPE_YARN = 2 +LLAMA_ROPE_SCALING_TYPE_LONGROPE = 3 +LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN + +# enum llama_pooling_type { +# LLAMA_POOLING_TYPE_UNSPECIFIED = -1, +# LLAMA_POOLING_TYPE_NONE = 0, +# LLAMA_POOLING_TYPE_MEAN = 1, +# LLAMA_POOLING_TYPE_CLS = 2, +# LLAMA_POOLING_TYPE_LAST = 3, +# LLAMA_POOLING_TYPE_RANK = 4, // used by reranking models to attach the classification head to the graph +# }; +LLAMA_POOLING_TYPE_UNSPECIFIED = -1 +LLAMA_POOLING_TYPE_NONE = 0 +LLAMA_POOLING_TYPE_MEAN = 1 +LLAMA_POOLING_TYPE_CLS = 2 +LLAMA_POOLING_TYPE_LAST = 3 +LLAMA_POOLING_TYPE_RANK = 4 + +# enum llama_attention_type { +# LLAMA_ATTENTION_TYPE_UNSPECIFIED = -1, +# LLAMA_ATTENTION_TYPE_CAUSAL = 0, +# LLAMA_ATTENTION_TYPE_NON_CAUSAL = 1, +# }; +LLAMA_ATTENTION_TYPE_UNSPECIFIED = -1 +LLAMA_ATTENTION_TYPE_CAUSAL = 0 +LLAMA_ATTENTION_TYPE_NON_CAUSAL = 1 + + +# enum llama_split_mode { +# LLAMA_SPLIT_MODE_NONE = 0, // single GPU +# LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs +# LLAMA_SPLIT_MODE_ROW = 2, // split rows across GPUs # }; -LLAMA_ROPE_SCALING_UNSPECIFIED = -1 -LLAMA_ROPE_SCALING_NONE = 0 -LLAMA_ROPE_SCALING_LINEAR = 1 -LLAMA_ROPE_SCALING_YARN = 2 -LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN +LLAMA_SPLIT_MODE_NONE = 0 +LLAMA_SPLIT_MODE_LAYER = 1 +LLAMA_SPLIT_MODE_ROW = 2 # typedef struct llama_token_data { @@ -214,46 +470,66 @@ def _load_shared_library(lib_base_name: str): # float logit; // log-odds of the token # float p; // probability of the token # } llama_token_data; -class llama_token_data(Structure): +class llama_token_data(ctypes.Structure): """Used to store token data - + Attributes: id (llama_token): token id logit (float): log-odds of the token p (float): probability of the token""" + + if TYPE_CHECKING: + id: llama_token + logit: float + p: float + _fields_ = [ ("id", llama_token), - ("logit", c_float), - ("p", c_float), + ("logit", ctypes.c_float), + ("p", ctypes.c_float), ] -llama_token_data_p = POINTER(llama_token_data) +llama_token_data_p = ctypes.POINTER(llama_token_data) # typedef struct llama_token_data_array { +# // TODO: consider SoA +# // NOTE: this pointer can be modified by the samplers # llama_token_data * data; # size_t size; +# int64_t selected; // this is the index in the data array (i.e. not the token id) # bool sorted; # } llama_token_data_array; -class llama_token_data_array(Structure): +class llama_token_data_array(ctypes.Structure): """Used to sample tokens given logits - + Attributes: data (ctypes.Array[llama_token_data]): token data size (int): size of the array + selected (int): index in the data array (i.e. not the token id) sorted (bool): whether the array is sorted""" + + if TYPE_CHECKING: + data: CtypesArray[llama_token_data] + size: int + selected: int + sorted: bool + _fields_ = [ ("data", llama_token_data_p), - ("size", c_size_t), - ("sorted", c_bool), + ("size", ctypes.c_size_t), + ("selected", ctypes.c_int64), + ("sorted", ctypes.c_bool), ] -llama_token_data_array_p = POINTER(llama_token_data_array) +llama_token_data_array_p = ctypes.POINTER(llama_token_data_array) -# typedef void (*llama_progress_callback)(float progress, void *ctx); -llama_progress_callback = ctypes.CFUNCTYPE(None, c_float, c_void_p) +# typedef bool (*llama_progress_callback)(float progress, void * user_data); +llama_progress_callback = ctypes.CFUNCTYPE( + ctypes.c_bool, ctypes.c_float, ctypes.c_void_p +) # // Input data for llama_decode @@ -263,8 +539,11 @@ class llama_token_data_array(Structure): # // - token : the token ids of the input (used when embd is NULL) # // - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL) # // - pos : the positions of the respective token in the sequence +# // (if set to NULL, the token position will be tracked automatically by llama_decode) # // - seq_id : the sequence to which the respective token belongs -# // - logits : if zero, the logits for the respective token will not be output +# // (if set to NULL, the sequence ID will be assumed to be 0) +# // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output +# // (if set to NULL, only the logits for last token will be returned) # // # typedef struct llama_batch { # int32_t n_tokens; @@ -274,19 +553,9 @@ class llama_token_data_array(Structure): # llama_pos * pos; # int32_t * n_seq_id; # llama_seq_id ** seq_id; -# int8_t * logits; - - -# // NOTE: helpers for smooth API transition - can be deprecated in the future -# // for future-proof code, use the above fields instead and ignore everything below -# // -# // pos[i] = all_pos_0 + i*all_pos_1 -# // -# llama_pos all_pos_0; // used if pos == NULL -# llama_pos all_pos_1; // used if pos == NULL -# llama_seq_id all_seq_id; // used if seq_id == NULL +# int8_t * logits; // TODO: rename this to "output" # } llama_batch; -class llama_batch(Structure): +class llama_batch(ctypes.Structure): """Input data for llama_decode A llama_batch object can contain input about one or many sequences @@ -294,105 +563,193 @@ class llama_batch(Structure): The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens Attributes: + n_tokens (int): number of tokens token (ctypes.Array[llama_token]): the token ids of the input (used when embd is NULL) - embd (ctypes.Array[ctypes.c_float]): token embeddings (i.e. float vector of size n_embd) (used when token is NULL) + embd (ctypes.Array[ctypes.ctypes.c_float]): token embeddings (i.e. float vector of size n_embd) (used when token is NULL) pos (ctypes.Array[ctypes.Array[llama_pos]]): the positions of the respective token in the sequence - seq_id (ctypes.Array[ctypes.Array[llama_seq_id]]): the sequence to which the respective token belongs""" + seq_id (ctypes.Array[ctypes.Array[llama_seq_id]]): the sequence to which the respective token belongs + logits (ctypes.Array[ctypes.ctypes.c_int8]): if zero, the logits for the respective token will not be output + """ + + if TYPE_CHECKING: + n_tokens: int + token: CtypesArray[llama_token] + embd: CtypesArray[ctypes.c_float] + pos: CtypesArray[CtypesArray[llama_pos]] + n_seq_id: CtypesArray[ctypes.c_int] + seq_id: CtypesArray[CtypesArray[llama_seq_id]] + logits: CtypesArray[ctypes.c_int8] _fields_ = [ - ("n_tokens", c_int32), - ("token", POINTER(llama_token)), - ("embd", c_float_p), - ("pos", POINTER(llama_pos)), - ("n_seq_id", POINTER(c_int32)), - ("seq_id", POINTER(POINTER(llama_seq_id))), - ("logits", POINTER(c_int8)), - ("all_pos_0", llama_pos), - ("all_pos_1", llama_pos), - ("all_seq_id", llama_seq_id), + ("n_tokens", ctypes.c_int32), + ("token", ctypes.POINTER(llama_token)), + ("embd", ctypes.POINTER(ctypes.c_float)), + ("pos", ctypes.POINTER(llama_pos)), + ("n_seq_id", ctypes.POINTER(ctypes.c_int32)), + ("seq_id", ctypes.POINTER(ctypes.POINTER(llama_seq_id))), + ("logits", ctypes.POINTER(ctypes.c_int8)), ] + # enum llama_model_kv_override_type { -# LLAMA_KV_OVERRIDE_INT, -# LLAMA_KV_OVERRIDE_FLOAT, -# LLAMA_KV_OVERRIDE_BOOL, +# LLAMA_KV_OVERRIDE_TYPE_INT, +# LLAMA_KV_OVERRIDE_TYPE_FLOAT, +# LLAMA_KV_OVERRIDE_TYPE_BOOL, +# LLAMA_KV_OVERRIDE_TYPE_STR, # }; -class llama_model_kv_override_type(Structure): - _fields_ = [ - ("LLAMA_KV_OVERRIDE_INT", c_int), - ("LLAMA_KV_OVERRIDE_FLOAT", c_int), - ("LLAMA_KV_OVERRIDE_BOOL", c_int), - ] +LLAMA_KV_OVERRIDE_TYPE_INT = 0 +LLAMA_KV_OVERRIDE_TYPE_FLOAT = 1 +LLAMA_KV_OVERRIDE_TYPE_BOOL = 2 +LLAMA_KV_OVERRIDE_TYPE_STR = 3 + # struct llama_model_kv_override { -# char key[128]; # enum llama_model_kv_override_type tag; + +# char key[128]; + + # union { -# int64_t int_value; -# double float_value; -# bool bool_value; +# int64_t val_i64; +# double val_f64; +# bool val_bool; +# char val_str[128]; # }; # }; -class llama_model_kv_override(Structure): +class llama_model_kv_override_value(ctypes.Union): + _fields_ = [ + ("val_i64", ctypes.c_int64), + ("val_f64", ctypes.c_double), + ("val_bool", ctypes.c_bool), + ("val_str", ctypes.c_char * 128), + ] + + if TYPE_CHECKING: + val_i64: int + val_f64: float + val_bool: bool + val_str: bytes + + +class llama_model_kv_override(ctypes.Structure): _fields_ = [ + ("tag", ctypes.c_int), ("key", ctypes.c_char * 128), - ("tag", llama_model_kv_override_type), - ("int_value", ctypes.c_int64), - ("float_value", c_double), - ("bool_value", c_bool), + ("value", llama_model_kv_override_value), ] + if TYPE_CHECKING: + tag: int + key: bytes + value: Union[int, float, bool, bytes] + + +# struct llama_model_tensor_buft_override { +# const char * pattern; +# ggml_backend_buffer_type_t buft; +# }; + + # struct llama_model_params { +# // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) +# ggml_backend_dev_t * devices; + +# // NULL-terminated list of buffer types to use for tensors that match a pattern +# const struct llama_model_tensor_buft_override * tensor_buft_overrides; + # int32_t n_gpu_layers; // number of layers to store in VRAM -# int32_t main_gpu; // the GPU that is used for scratch and small tensors -# const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES) +# enum llama_split_mode split_mode; // how to split the model across multiple GPUs + +# // main_gpu interpretation depends on split_mode: +# // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model +# // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results +# // LLAMA_SPLIT_MODE_LAYER: ignored +# int32_t main_gpu; -# // called with a progress value between 0 and 1, pass NULL to disable +# // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() +# const float * tensor_split; + +# // Called with a progress value between 0.0 and 1.0. Pass NULL to disable. +# // If the provided progress_callback returns true, model loading continues. +# // If it returns false, model loading is immediately aborted. # llama_progress_callback progress_callback; + # // context pointer passed to the progress callback # void * progress_callback_user_data; # // override key-value pairs of the model meta data # const struct llama_model_kv_override * kv_overrides; + # // Keep the booleans together to avoid misalignment during copy-by-value. -# bool vocab_only; // only load the vocabulary, no weights -# bool use_mmap; // use mmap if possible -# bool use_mlock; // force system to keep model in RAM +# bool vocab_only; // only load the vocabulary, no weights +# bool use_mmap; // use mmap if possible +# bool use_mlock; // force system to keep model in RAM +# bool check_tensors; // validate model tensor data # }; -class llama_model_params(Structure): +class llama_model_params(ctypes.Structure): """Parameters for llama_model - + Attributes: + devices (ctypes.Array[ggml_backend_dev_t]): NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) + tensor_buft_overrides (ctypes.Array[llama_model_tensor_buft_override]): NULL-terminated list of buffer types to use for tensors that match a pattern n_gpu_layers (int): number of layers to store in VRAM - main_gpu (int): the GPU that is used for scratch and small tensors - tensor_split (ctypes.Array[ctypes.c_float]): how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES) - progress_callback (llama_progress_callback): called with a progress value between 0 and 1, pass NULL to disable - progress_callback_user_data (ctypes.c_void_p): context pointer passed to the progress callback + split_mode (int): how to split the model across multiple GPUs + main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored + tensor_split (ctypes.Array[ctypes.ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() + progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted. + progress_callback_user_data (ctypes.ctypes.c_void_p): context pointer passed to the progress callback kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data vocab_only (bool): only load the vocabulary, no weights use_mmap (bool): use mmap if possible - use_mlock (bool): force system to keep model in RAM""" + use_mlock (bool): force system to keep model in RAM + check_tensors (bool): validate model tensor data""" + + if TYPE_CHECKING: + devices: CtypesArray[ctypes.c_void_p] # NOTE: unused + tensor_buft_overrides: CtypesArray[llama_model_tensor_buft_override] # NOTE: unused + n_gpu_layers: int + split_mode: int + main_gpu: int + tensor_split: CtypesArray[ctypes.c_float] + progress_callback: Callable[[float, ctypes.c_void_p], bool] + progress_callback_user_data: ctypes.c_void_p + kv_overrides: CtypesArray[llama_model_kv_override] + vocab_only: bool + use_mmap: bool + use_mlock: bool + check_tensors: bool + _fields_ = [ - ("n_gpu_layers", c_int32), - ("main_gpu", c_int32), - ("tensor_split", c_float_p), + ("devices", ctypes.c_void_p), # NOTE: unnused + ("tensor_buft_overrides", ctypes.c_void_p), # NOTE: unused + ("n_gpu_layers", ctypes.c_int32), + ("split_mode", ctypes.c_int), + ("main_gpu", ctypes.c_int32), + ("tensor_split", ctypes.POINTER(ctypes.c_float)), ("progress_callback", llama_progress_callback), - ("progress_callback_user_data", c_void_p), - ("kv_overrides", POINTER(llama_model_kv_override)), - ("vocab_only", c_bool), - ("use_mmap", c_bool), - ("use_mlock", c_bool), + ("progress_callback_user_data", ctypes.c_void_p), + ("kv_overrides", ctypes.POINTER(llama_model_kv_override)), + ("vocab_only", ctypes.c_bool), + ("use_mmap", ctypes.c_bool), + ("use_mlock", ctypes.c_bool), + ("check_tensors", ctypes.c_bool), ] +# // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations +# // https://github.com/ggerganov/llama.cpp/pull/7544 # struct llama_context_params { -# uint32_t seed; // RNG seed, -1 for random # uint32_t n_ctx; // text context, 0 = from model -# uint32_t n_batch; // prompt processing maximum batch size -# uint32_t n_threads; // number of threads to use for generation -# uint32_t n_threads_batch; // number of threads to use for batch processing -# int8_t rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` +# uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode +# uint32_t n_ubatch; // physical maximum batch size +# uint32_t n_seq_max; // max number of sequences (i.e. distinct states for recurrent models) +# int32_t n_threads; // number of threads to use for generation +# int32_t n_threads_batch; // number of threads to use for batch processing + +# enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` +# enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id +# enum llama_attention_type attention_type; // attention type to use for embeddings # // ref: https://github.com/ggerganov/llama.cpp/pull/2054 # float rope_freq_base; // RoPE base frequency, 0 = from model @@ -402,26 +759,41 @@ class llama_model_params(Structure): # float yarn_beta_fast; // YaRN low correction dim # float yarn_beta_slow; // YaRN high correction dim # uint32_t yarn_orig_ctx; // YaRN original context size +# float defrag_thold; // defragment the KV cache if holes/size > thold, < 0 disabled (default) -# enum ggml_type type_k; // data type for K cache -# enum ggml_type type_v; // data type for V cache +# ggml_backend_sched_eval_callback cb_eval; +# void * cb_eval_user_data; + +# enum ggml_type type_k; // data type for K cache [EXPERIMENTAL] +# enum ggml_type type_v; // data type for V cache [EXPERIMENTAL] # // Keep the booleans together to avoid misalignment during copy-by-value. -# bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true) -# bool logits_all; // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) -# bool embedding; // embedding mode only +# bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) +# bool embeddings; // if true, extract embeddings (together with logits) # bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU +# bool flash_attn; // whether to use flash attention [EXPERIMENTAL] +# bool no_perf; // whether to measure performance timings + + +# // Abort callback +# // if it returns true, execution of llama_decode() will be aborted +# // currently works only with CPU execution +# ggml_abort_callback abort_callback; +# void * abort_callback_data; # }; -class llama_context_params(Structure): +class llama_context_params(ctypes.Structure): """Parameters for llama_context - + Attributes: - seed (int): RNG seed, -1 for random n_ctx (int): text context, 0 = from model - n_batch (int): prompt processing maximum batch size + n_batch (int): logical maximum batch size that can be submitted to llama_decode + n_ubatch (int): physical maximum batch size + n_seq_max (int): max number of sequences (i.e. distinct states for recurrent models) n_threads (int): number of threads to use for generation n_threads_batch (int): number of threads to use for batch processing rope_scaling_type (int): RoPE scaling type, from `enum llama_rope_scaling_type` + pooling_type (int): whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) + attention_type (int): attention type to use for embeddings rope_freq_base (float): RoPE base frequency, 0 = from model rope_freq_scale (float): RoPE frequency scaling factor, 0 = from model yarn_ext_factor (float): YaRN extrapolation mix factor, negative = from model @@ -429,32 +801,79 @@ class llama_context_params(Structure): yarn_beta_fast (float): YaRN low correction dim yarn_beta_slow (float): YaRN high correction dim yarn_orig_ctx (int): YaRN original context size + defrag_thold (float): defragment the KV cache if holes/size > thold, < 0 disabled (default) + cb_eval (ggml_backend_sched_eval_callback): callback for scheduling eval + cb_eval_user_data (ctypes.ctypes.c_void_p): user data for cb_eval type_k (int): data type for K cache type_v (int): data type for V cache - mul_mat_q (bool): if true, use experimental mul_mat_q kernels (DEPRECATED - always true) - logits_all (bool): the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) - embedding (bool): embedding mode only - offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU""" + logits_all (bool): the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) + embeddings (bool): if true, extract embeddings (together with logits) + offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU + flash_attn (bool): whether to use flash attention + no_perf (bool): whether to measure performance timings + abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted + abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback + """ + + if TYPE_CHECKING: + n_ctx: int + n_batch: int + n_ubatch: int + n_seq_max: int + n_threads: int + n_threads_batch: int + rope_scaling_type: int + pooling_type: int + attention_type: int + rope_freq_base: float + rope_freq_scale: float + yarn_ext_factor: float + yarn_attn_factor: float + yarn_beta_fast: float + yarn_beta_slow: float + yarn_orig_ctx: int + defrag_thold: float + cb_eval: Callable[[ctypes.c_void_p, bool], bool] + cb_eval_user_data: ctypes.c_void_p + type_k: int + type_v: int + logits_all: bool + embeddings: bool + offload_kqv: bool + flash_attn: bool + no_perf: bool + abort_callback: Callable[[ctypes.c_void_p], bool] + abort_callback_data: ctypes.c_void_p + _fields_ = [ - ("seed", c_uint32), - ("n_ctx", c_uint32), - ("n_batch", c_uint32), - ("n_threads", c_uint32), - ("n_threads_batch", c_uint32), - ("rope_scaling_type", c_int8), - ("rope_freq_base", c_float), - ("rope_freq_scale", c_float), - ("yarn_ext_factor", c_float), - ("yarn_attn_factor", c_float), - ("yarn_beta_fast", c_float), - ("yarn_beta_slow", c_float), - ("yarn_orig_ctx", c_uint32), - ("type_k", c_int), - ("type_v", c_int), - ("mul_mat_q", c_bool), - ("logits_all", c_bool), - ("embedding", c_bool), - ("offload_kqv", c_bool), + ("n_ctx", ctypes.c_uint32), + ("n_batch", ctypes.c_uint32), + ("n_ubatch", ctypes.c_uint32), + ("n_seq_max", ctypes.c_uint32), + ("n_threads", ctypes.c_int32), + ("n_threads_batch", ctypes.c_int32), + ("rope_scaling_type", ctypes.c_int), + ("pooling_type", ctypes.c_int), + ("attention_type", ctypes.c_int), + ("rope_freq_base", ctypes.c_float), + ("rope_freq_scale", ctypes.c_float), + ("yarn_ext_factor", ctypes.c_float), + ("yarn_attn_factor", ctypes.c_float), + ("yarn_beta_fast", ctypes.c_float), + ("yarn_beta_slow", ctypes.c_float), + ("yarn_orig_ctx", ctypes.c_uint32), + ("defrag_thold", ctypes.c_float), + ("cb_eval", ggml_backend_sched_eval_callback), + ("cb_eval_user_data", ctypes.c_void_p), + ("type_k", ctypes.c_int), + ("type_v", ctypes.c_int), + ("logits_all", ctypes.c_bool), + ("embeddings", ctypes.c_bool), + ("offload_kqv", ctypes.c_bool), + ("flash_attn", ctypes.c_bool), + ("no_perf", ctypes.c_bool), + ("abort_callback", ggml_abort_callback), + ("abort_callback_data", ctypes.c_void_p), ] @@ -464,7 +883,9 @@ class llama_context_params(Structure): # // if it exists. # // It might not exist for progress report where '.' is output repeatedly. # typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data); -llama_log_callback = ctypes.CFUNCTYPE(None, c_int, c_char_p, c_void_p) +llama_log_callback = ctypes.CFUNCTYPE( + None, ctypes.c_int, ctypes.c_char_p, ctypes.c_void_p +) """Signature for logging events Note that text includes the new line character at the end for most events. If your logging mechanism cannot handle that, check if the last character is '\n' and strip it @@ -474,512 +895,847 @@ class llama_context_params(Structure): # // model quantization parameters # typedef struct llama_model_quantize_params { -# int nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() -# enum llama_ftype ftype; // quantize to this llama_ftype -# bool allow_requantize; // allow quantizing non-f32/f16 tensors -# bool quantize_output_tensor; // quantize output.weight -# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored -# bool pure; // disable k-quant mixtures and quantize all tensors to the same type +# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() +# enum llama_ftype ftype; // quantize to this llama_ftype +# enum ggml_type output_tensor_type; // output tensor type +# enum ggml_type token_embedding_type; // token embeddings tensor type +# bool allow_requantize; // allow quantizing non-f32/f16 tensors +# bool quantize_output_tensor; // quantize output.weight +# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored +# bool pure; // quantize all tensors to the default type +# bool keep_split; // quantize to the same number of shards +# void * imatrix; // pointer to importance matrix data +# void * kv_overrides; // pointer to vector containing overrides +# void * tensor_types; // pointer to vector containing tensor types # } llama_model_quantize_params; -class llama_model_quantize_params(Structure): +class llama_model_quantize_params(ctypes.Structure): """Parameters for llama_model_quantize - + Attributes: nthread (int): number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() ftype (int): quantize to this llama_ftype + output_tensor_type (int): output tensor type + token_embedding_type (int): token embeddings tensor type allow_requantize (bool): allow quantizing non-f32/f16 tensors quantize_output_tensor (bool): quantize output.weight only_copy (bool): only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored - pure (bool): disable k-quant mixtures and quantize all tensors to the same type""" + pure (bool): quantize all tensors to the default type + keep_split (bool): quantize to the same number of shards + imatrix (ctypes.c_void_p): pointer to importance matrix data + kv_overrides (ctypes.c_void_p): pointer to vector containing overrides + tensor_types (ctypes.c_void_p): pointer to vector containing tensor types + """ + + if TYPE_CHECKING: + nthread: int + ftype: int + output_tensor_type: int + token_embedding_type: int + allow_requantize: bool + quantize_output_tensor: bool + only_copy: bool + pure: bool + keep_split: bool + imatrix: ctypes.c_void_p + kv_overrides: ctypes.c_void_p + tensor_types: ctypes.c_void_p + _fields_ = [ - ("nthread", c_int), - ("ftype", c_int), - ("allow_requantize", c_bool), - ("quantize_output_tensor", c_bool), - ("only_copy", c_bool), + ("nthread", ctypes.c_int32), + ("ftype", ctypes.c_int), + ("output_tensor_type", ctypes.c_int), + ("token_embedding_type", ctypes.c_int), + ("allow_requantize", ctypes.c_bool), + ("quantize_output_tensor", ctypes.c_bool), + ("only_copy", ctypes.c_bool), + ("pure", ctypes.c_bool), + ("keep_split", ctypes.c_bool), + ("imatrix", ctypes.c_void_p), + ("kv_overrides", ctypes.c_void_p), + ("tensor_types", ctypes.c_void_p), ] -# // grammar types -# struct llama_grammar; -llama_grammar_p = c_void_p +# typedef struct llama_logit_bias { +# llama_token token; +# float bias; +# } llama_logit_bias; +class llama_logit_bias(ctypes.Structure): + """Used to store logit bias -# // grammar element type -# enum llama_gretype { -# // end of rule definition -# LLAMA_GRETYPE_END = 0, + Attributes: + token (llama_token): token id + bias (float): bias""" -# // start of alternate definition for rule -# LLAMA_GRETYPE_ALT = 1, + if TYPE_CHECKING: + token: llama_token + bias: float -# // non-terminal element: reference to rule -# LLAMA_GRETYPE_RULE_REF = 2, + _fields_ = [ + ("token", llama_token), + ("bias", ctypes.c_float), + ] -# // terminal element: character (code point) -# LLAMA_GRETYPE_CHAR = 3, -# // inverse char(s) ([^a], [^a-b] [^abc]) -# LLAMA_GRETYPE_CHAR_NOT = 4, +llama_logit_bias_p = ctypes.POINTER(llama_logit_bias) -# // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to -# // be an inclusive range ([a-z]) -# LLAMA_GRETYPE_CHAR_RNG_UPPER = 5, -# // modifies a preceding LLAMA_GRETYPE_CHAR or -# // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA]) -# LLAMA_GRETYPE_CHAR_ALT = 6, -# }; -LLAMA_GRETYPE_END = 0 -LLAMA_GRETYPE_ALT = 1 -LLAMA_GRETYPE_RULE_REF = 2 -LLAMA_GRETYPE_CHAR = 3 -LLAMA_GRETYPE_CHAR_NOT = 4 -LLAMA_GRETYPE_CHAR_RNG_UPPER = 5 -LLAMA_GRETYPE_CHAR_ALT = 6 - - -# typedef struct llama_grammar_element { -# enum llama_gretype type; -# uint32_t value; // Unicode code point or rule ID -# } llama_grammar_element; -class llama_grammar_element(Structure): - _fields_ = [ - ("type", c_int), - ("value", c_uint32), - ] +# typedef struct llama_sampler_chain_params { +# bool no_perf; // whether to measure performance timings +# } llama_sampler_chain_params; +class llama_sampler_chain_params(ctypes.Structure): + """Parameters for llama_sampler_chain + Attributes: + no_perf (bool): whether to measure performance timings""" -llama_grammar_element_p = POINTER(llama_grammar_element) + if TYPE_CHECKING: + no_perf: bool -# // performance timing information -# struct llama_timings { -# double t_start_ms; -# double t_end_ms; -# double t_load_ms; -# double t_sample_ms; -# double t_p_eval_ms; -# double t_eval_ms; + _fields_ = [ + ("no_perf", ctypes.c_bool), + ] -# int32_t n_sample; -# int32_t n_p_eval; -# int32_t n_eval; -# }; -class llama_timings(Structure): +# // used in chat template +# typedef struct llama_chat_message { +# const char * role; +# const char * content; +# } llama_chat_message; +class llama_chat_message(ctypes.Structure): _fields_ = [ - ("t_start_ms", c_double), - ("t_end_ms", c_double), - ("t_load_ms", c_double), - ("t_sample_ms", c_double), - ("t_p_eval_ms", c_double), - ("t_eval_ms", c_double), - ("n_sample", c_int32), - ("n_p_eval", c_int32), - ("n_eval", c_int32), + ("role", ctypes.c_char_p), + ("content", ctypes.c_char_p), ] +# // lora adapter +# struct llama_adapter_lora; +llama_adapter_lora_p = ctypes.c_void_p +llama_adapter_lora_p_ctypes = ctypes.POINTER(ctypes.c_void_p) + + # // Helpers for getting default parameters -# LLAMA_API struct llama_model_params llama_model_default_params(void); +# LLAMA_API struct llama_model_params llama_model_default_params(void); +@ctypes_function( + "llama_model_default_params", + [], + llama_model_params, +) def llama_model_default_params() -> llama_model_params: """Get default parameters for llama_model""" - return _lib.llama_model_default_params() - + ... -_lib.llama_model_default_params.argtypes = [] -_lib.llama_model_default_params.restype = llama_model_params - -# LLAMA_API struct llama_context_params llama_context_default_params(void); +# LLAMA_API struct llama_context_params llama_context_default_params(void); +@ctypes_function( + "llama_context_default_params", + [], + llama_context_params, +) def llama_context_default_params() -> llama_context_params: """Get default parameters for llama_context""" - return _lib.llama_context_default_params() + ... -_lib.llama_context_default_params.argtypes = [] -_lib.llama_context_default_params.restype = llama_context_params +# LLAMA_API struct llama_sampler_chain_params llama_sampler_chain_default_params(void); +@ctypes_function( + "llama_sampler_chain_default_params", + [], + llama_sampler_chain_params, +) +def llama_sampler_chain_default_params() -> llama_sampler_chain_params: + """Get default parameters for llama_sampler_chain""" + ... # LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void); +@ctypes_function( + "llama_model_quantize_default_params", + [], + llama_model_quantize_params, +) def llama_model_quantize_default_params() -> llama_model_quantize_params: """Get default parameters for llama_model_quantize""" - return _lib.llama_model_quantize_default_params() - - -_lib.llama_model_quantize_default_params.argtypes = [] -_lib.llama_model_quantize_default_params.restype = llama_model_quantize_params + ... # // Initialize the llama + ggml backend # // If numa is true, use NUMA optimizations # // Call once at the start of the program # LLAMA_API void llama_backend_init(bool numa); -def llama_backend_init(numa: Union[c_bool, bool]): +# LLAMA_API void llama_backend_init(void); +@ctypes_function( + "llama_backend_init", + [], + None, +) +def llama_backend_init(): """Initialize the llama + ggml backend If numa is true, use NUMA optimizations Call once at the start of the program""" - return _lib.llama_backend_init(numa) + ... -_lib.llama_backend_init.argtypes = [c_bool] -_lib.llama_backend_init.restype = None +# // numa strategies +# enum ggml_numa_strategy { +# GGML_NUMA_STRATEGY_DISABLED = 0, +# GGML_NUMA_STRATEGY_DISTRIBUTE = 1, +# GGML_NUMA_STRATEGY_ISOLATE = 2, +# GGML_NUMA_STRATEGY_NUMACTL = 3, +# GGML_NUMA_STRATEGY_MIRROR = 4, +# GGML_NUMA_STRATEGY_COUNT +# }; +GGML_NUMA_STRATEGY_DISABLED = 0 +GGML_NUMA_STRATEGY_DISTRIBUTE = 1 +GGML_NUMA_STRATEGY_ISOLATE = 2 +GGML_NUMA_STRATEGY_NUMACTL = 3 +GGML_NUMA_STRATEGY_MIRROR = 4 +GGML_NUMA_STRATEGY_COUNT = 5 # // Call once at the end of the program - currently only used for MPI # LLAMA_API void llama_backend_free(void); +@ctypes_function( + "llama_backend_free", + [], + None, +) def llama_backend_free(): """Call once at the end of the program - currently only used for MPI""" - return _lib.llama_backend_free() + ... + + +# //optional: +# LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa); +@ctypes_function( + "llama_numa_init", + [ctypes.c_int], + None, +) +def llama_numa_init(numa: int, /): + ... + + +# // Optional: an auto threadpool gets created in ggml if not passed explicitly +# LLAMA_API void llama_attach_threadpool( +# struct llama_context * ctx, +# ggml_threadpool_t threadpool, +# ggml_threadpool_t threadpool_batch); +# TODO: Add llama_attach_threadpool -_lib.llama_backend_free.argtypes = [] -_lib.llama_backend_free.restype = None +# LLAMA_API void llama_detach_threadpool(struct llama_context * ctx); +# TODO: Add llama_detach_threadpool -# LLAMA_API struct llama_model * llama_load_model_from_file( +# DEPRECATED(LLAMA_API struct llama_model * llama_load_model_from_file( # const char * path_model, -# struct llama_model_params params); +# struct llama_model_params params), +# "use llama_model_load_from_file instead"); +@ctypes_function( + "llama_load_model_from_file", + [ctypes.c_char_p, llama_model_params], + llama_model_p_ctypes, +) def llama_load_model_from_file( - path_model: bytes, params: llama_model_params -) -> llama_model_p: - return _lib.llama_load_model_from_file(path_model, params) + path_model: bytes, params: llama_model_params, / +) -> Optional[llama_model_p]: + ... -_lib.llama_load_model_from_file.argtypes = [c_char_p, llama_model_params] -_lib.llama_load_model_from_file.restype = llama_model_p +# // Load the model from a file +# // If the file is split into multiple parts, the file name must follow this pattern: -%05d-of-%05d.gguf +# // If the split file name does not follow this pattern, use llama_model_load_from_splits +# LLAMA_API struct llama_model * llama_model_load_from_file( +# const char * path_model, +# struct llama_model_params params); +@ctypes_function( + "llama_model_load_from_file", + [ctypes.c_char_p, llama_model_params], + llama_model_p_ctypes, +) +def llama_model_load_from_file( + path_model: bytes, params: llama_model_params, / +) -> Optional[llama_model_p]: + """Load the model from a file + + If the file is split into multiple parts, the file name must follow this pattern: -%05d-of-%05d.gguf + + If the split file name does not follow this pattern, use llama_model_load_from_splits""" + ... + + +# // Load the model from multiple splits (support custom naming scheme) +# // The paths must be in the correct order +# LLAMA_API struct llama_model * llama_model_load_from_splits( +# const char ** paths, +# size_t n_paths, +# struct llama_model_params params); +@ctypes_function( + "llama_model_load_from_splits", + [ctypes.POINTER(ctypes.c_char_p), ctypes.c_size_t, llama_model_params], + llama_model_p_ctypes, +) +def llama_model_load_from_splits( + paths: List[bytes], n_paths: int, params: llama_model_params, / +) -> Optional[llama_model_p]: + """Load the model from multiple splits (support custom naming scheme) + + The paths must be in the correct order""" + ... # LLAMA_API void llama_free_model(struct llama_model * model); -def llama_free_model(model: llama_model_p): - return _lib.llama_free_model(model) +@ctypes_function( + "llama_free_model", + [llama_model_p_ctypes], + None, +) +def llama_free_model(model: llama_model_p, /): + ... -_lib.llama_free_model.argtypes = [llama_model_p] -_lib.llama_free_model.restype = None +# LLAMA_API void llama_model_free(struct llama_model * model); +@ctypes_function( + "llama_model_free", + [llama_model_p_ctypes], + None, +) +def llama_model_free(model: llama_model_p, /): + ... -# LLAMA_API struct llama_context * llama_new_context_with_model( +# LLAMA_API struct llama_context * llama_init_from_model( # struct llama_model * model, # struct llama_context_params params); -def llama_new_context_with_model( - model: llama_model_p, params: llama_context_params -) -> llama_context_p: - return _lib.llama_new_context_with_model(model, params) +@ctypes_function( + "llama_init_from_model", + [llama_model_p_ctypes, llama_context_params], + llama_context_p_ctypes, +) +def llama_init_from_model( + model: llama_model_p, params: llama_context_params, / +) -> Optional[llama_context_p]: + ... -_lib.llama_new_context_with_model.argtypes = [llama_model_p, llama_context_params] -_lib.llama_new_context_with_model.restype = llama_context_p +# DEPRECATED(LLAMA_API struct llama_context * llama_new_context_with_model( +# struct llama_model * model, +# struct llama_context_params params), +# "use llama_init_from_model instead"); +@ctypes_function( + "llama_new_context_with_model", + [llama_model_p_ctypes, llama_context_params], + llama_context_p_ctypes, +) +def llama_new_context_with_model( + model: llama_model_p, params: llama_context_params, / +) -> Optional[llama_context_p]: + ... # // Frees all allocated memory # LLAMA_API void llama_free(struct llama_context * ctx); -def llama_free(ctx: llama_context_p): +@ctypes_function( + "llama_free", + [llama_context_p_ctypes], + None, +) +def llama_free(ctx: llama_context_p, /): """Frees all allocated memory""" - return _lib.llama_free(ctx) - - -_lib.llama_free.argtypes = [llama_context_p] -_lib.llama_free.restype = None + ... # LLAMA_API int64_t llama_time_us(void); +@ctypes_function( + "llama_time_us", + [], + ctypes.c_int64, +) def llama_time_us() -> int: - return _lib.llama_time_us() + ... -_lib.llama_time_us.argtypes = [] -_lib.llama_time_us.restype = ctypes.c_int64 +# LLAMA_API size_t llama_max_devices(void); +@ctypes_function("llama_max_devices", [], ctypes.c_size_t) +def llama_max_devices() -> int: + ... -# LLAMA_API int llama_max_devices (void); -def llama_max_devices() -> int: - return _lib.llama_max_devices() +# LLAMA_API bool llama_supports_mmap (void); +@ctypes_function("llama_supports_mmap", [], ctypes.c_bool) +def llama_supports_mmap() -> bool: + ... -_lib.llama_max_devices.argtypes = [] -_lib.llama_max_devices.restype = c_int +# LLAMA_API bool llama_supports_mlock (void); +@ctypes_function("llama_supports_mlock", [], ctypes.c_bool) +def llama_supports_mlock() -> bool: + ... + + +# LLAMA_API bool llama_supports_gpu_offload(void); +@ctypes_function("llama_supports_gpu_offload", [], ctypes.c_bool) +def llama_supports_gpu_offload() -> bool: + ... + + +# LLAMA_API bool llama_supports_rpc (void); +@ctypes_function("llama_supports_rpc", [], ctypes.c_bool) +def llama_supports_rpc() -> bool: + ... + + +# LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); +@ctypes_function("llama_n_ctx", [llama_context_p_ctypes], ctypes.c_uint32) +def llama_n_ctx(ctx: llama_context_p, /) -> int: + ... + + +# LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); +@ctypes_function("llama_n_batch", [llama_context_p_ctypes], ctypes.c_uint32) +def llama_n_batch(ctx: llama_context_p, /) -> int: + ... + + +# LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx); +@ctypes_function("llama_n_ubatch", [llama_context_p_ctypes], ctypes.c_uint32) +def llama_n_ubatch(ctx: llama_context_p, /) -> int: + ... + + +# LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx); +@ctypes_function("llama_n_seq_max", [llama_context_p_ctypes], ctypes.c_uint32) +def llama_n_seq_max(ctx: llama_context_p, /) -> int: + ... -# LLAMA_API bool llama_mmap_supported (void); -def llama_mmap_supported() -> bool: - return _lib.llama_mmap_supported() -_lib.llama_mmap_supported.argtypes = [] -_lib.llama_mmap_supported.restype = c_bool +# DEPRECATED(LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model), "use llama_model_n_ctx_train instead"); +@ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) +def llama_n_ctx_train(model: llama_model_p, /) -> int: + ... -# LLAMA_API bool llama_mlock_supported(void); -def llama_mlock_supported() -> bool: - return _lib.llama_mlock_supported() +# DEPRECATED(LLAMA_API int32_t llama_n_embd (const struct llama_model * model), "use llama_model_n_embd instead"); +@ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) +def llama_n_embd(model: llama_model_p, /) -> int: + ... -_lib.llama_mlock_supported.argtypes = [] -_lib.llama_mlock_supported.restype = c_bool +# DEPRECATED(LLAMA_API int32_t llama_n_layer (const struct llama_model * model), "use llama_model_n_layer instead"); +@ctypes_function("llama_n_layer", [llama_model_p_ctypes], ctypes.c_int32) +def llama_n_layer(model: llama_model_p, /) -> int: + ... -# LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); -def llama_get_model(ctx: llama_context_p) -> llama_model_p: - return _lib.llama_get_model(ctx) +# DEPRECATED(LLAMA_API int32_t llama_n_head (const struct llama_model * model), "use llama_model_n_head instead"); +@ctypes_function("llama_n_head", [llama_model_p_ctypes], ctypes.c_int32) +def llama_n_head(model: llama_model_p, /) -> int: + ... -_lib.llama_get_model.argtypes = [llama_context_p] -_lib.llama_get_model.restype = llama_model_p +# DEPRECATED(LLAMA_API int32_t llama_n_vocab (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead"); +@ctypes_function("llama_n_vocab", [llama_vocab_p_ctypes], ctypes.c_int32) +def llama_n_vocab(model: llama_vocab_p, /) -> int: + ... -# LLAMA_API int llama_n_ctx (const struct llama_context * ctx); -def llama_n_ctx(ctx: llama_context_p) -> int: - return _lib.llama_n_ctx(ctx) +# LLAMA_API const struct llama_model * llama_get_model (const struct llama_context * ctx); +@ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) +def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: + ... -_lib.llama_n_ctx.argtypes = [llama_context_p] -_lib.llama_n_ctx.restype = c_int +# LLAMA_API struct llama_kv_cache * llama_get_kv_self ( struct llama_context * ctx); +@ctypes_function( + "llama_get_kv_self", + [llama_context_p_ctypes], + llama_kv_cache_p_ctypes, +) +def llama_get_kv_self(ctx: llama_context_p, /) -> Optional[llama_kv_cache_p]: + """Get the KV cache for self-attention""" + ... -# LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); -def llama_vocab_type(model: llama_model_p) -> int: - return _lib.llama_vocab_type(model) +# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); +@ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) +def llama_pooling_type(ctx: llama_context_p, /) -> int: + ... -_lib.llama_vocab_type.argtypes = [llama_model_p] -_lib.llama_vocab_type.restype = c_int +# LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model); +@ctypes_function("llama_model_get_vocab", [llama_model_p_ctypes], llama_vocab_p_ctypes) +def llama_model_get_vocab(model: llama_model_p, /) -> Optional[llama_vocab_p]: + ... -# LLAMA_API int llama_n_vocab (const struct llama_model * model); -def llama_n_vocab(model: llama_model_p) -> int: - return _lib.llama_n_vocab(model) +# LLAMA_API enum llama_rope_type llama_model_rope_type(const struct llama_model * model); +@ctypes_function("llama_model_rope_type", [llama_model_p_ctypes], ctypes.c_int) +def llama_model_rope_type(model: llama_model_p, /) -> int: + ... -_lib.llama_n_vocab.argtypes = [llama_model_p] -_lib.llama_n_vocab.restype = c_int +# LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model); +@ctypes_function("llama_model_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_ctx_train(model: llama_model_p, /) -> int: + ... -# LLAMA_API int llama_n_ctx_train(const struct llama_model * model); -def llama_n_ctx_train(model: llama_model_p) -> int: - return _lib.llama_n_ctx_train(model) +# LLAMA_API int32_t llama_model_n_embd (const struct llama_model * model); +@ctypes_function("llama_model_n_embd", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_embd(model: llama_model_p, /) -> int: + ... -_lib.llama_n_ctx_train.argtypes = [llama_model_p] -_lib.llama_n_ctx_train.restype = c_int +# LLAMA_API int32_t llama_model_n_layer (const struct llama_model * model); +@ctypes_function("llama_model_n_layer", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_layer(model: llama_model_p, /) -> int: + ... -# LLAMA_API int llama_n_embd (const struct llama_model * model); -def llama_n_embd(model: llama_model_p) -> int: - return _lib.llama_n_embd(model) +# LLAMA_API int32_t llama_model_n_head (const struct llama_model * model); +@ctypes_function("llama_model_n_head", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_head(model: llama_model_p, /) -> int: + ... -_lib.llama_n_embd.argtypes = [llama_model_p] -_lib.llama_n_embd.restype = c_int +# LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model); +@ctypes_function("llama_model_n_head_kv", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_head_kv(model: llama_model_p, /) -> int: + ... # // Get the model's RoPE frequency scaling factor -# LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); -def llama_rope_freq_scale_train(model: llama_model_p) -> float: - """Get the model's RoPE frequency scaling factor""" - return _lib.llama_rope_freq_scale_train(model) +# LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model); +@ctypes_function("llama_model_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) +def llama_model_rope_freq_scale_train(model: llama_model_p, /) -> float: + ... + + +# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); +@ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) +def llama_vocab_type(model: llama_model_p, /) -> int: + ... + +# LLAMA_API int32_t llama_vocab_n_tokens(const struct llama_vocab * vocab); +@ctypes_function("llama_vocab_n_tokens", [llama_vocab_p_ctypes], ctypes.c_int32) +def llama_vocab_n_tokens(vocab: llama_vocab_p, /) -> int: + ... -_lib.llama_rope_freq_scale_train.argtypes = [llama_model_p] -_lib.llama_rope_freq_scale_train.restype = c_float # // Functions to access the model's GGUF metadata scalar values # // - The functions return the length of the string on success, or -1 on failure # // - The output string is always null-terminated and cleared on failure +# // - When retrieving a string, an extra byte must be allocated to account for the null terminator # // - GGUF array values are not supported by these functions # // Get metadata value as a string by key name -# LLAMA_API int llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); +# LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); +@ctypes_function( + "llama_model_meta_val_str", + [ + llama_model_p_ctypes, + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.c_size_t, + ], + ctypes.c_int32, +) def llama_model_meta_val_str( - model: llama_model_p, key: Union[c_char_p, bytes], buf: bytes, buf_size: int + model: llama_model_p, + key: Union[ctypes.c_char_p, bytes], + buf: bytes, + buf_size: int, + /, ) -> int: """Get metadata value as a string by key name""" - return _lib.llama_model_meta_val_str(model, key, buf, buf_size) - - -_lib.llama_model_meta_val_str.argtypes = [llama_model_p, c_char_p, c_char_p, c_size_t] -_lib.llama_model_meta_val_str.restype = c_int + ... # // Get the number of metadata key/value pairs -# LLAMA_API int llama_model_meta_count(const struct llama_model * model); -def llama_model_meta_count(model: llama_model_p) -> int: +# LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model); +@ctypes_function("llama_model_meta_count", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_meta_count(model: llama_model_p, /) -> int: """Get the number of metadata key/value pairs""" - return _lib.llama_model_meta_count(model) - - -_lib.llama_model_meta_count.argtypes = [llama_model_p] -_lib.llama_model_meta_count.restype = c_int + ... # // Get metadata key name by index -# LLAMA_API int llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size); +# LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); +@ctypes_function( + "llama_model_meta_key_by_index", + [ + llama_model_p_ctypes, + ctypes.c_int32, + ctypes.c_char_p, + ctypes.c_size_t, + ], + ctypes.c_int32, +) def llama_model_meta_key_by_index( - model: llama_model_p, i: Union[c_int, int], buf: bytes, buf_size: int + model: llama_model_p, + i: Union[ctypes.c_int, int], + buf: Union[bytes, CtypesArray[ctypes.c_char]], + buf_size: int, + /, ) -> int: """Get metadata key name by index""" - return _lib.llama_model_meta_key_by_index(model, i, buf, buf_size) - - -_lib.llama_model_meta_key_by_index.argtypes = [llama_model_p, c_int, c_char_p, c_size_t] -_lib.llama_model_meta_key_by_index.restype = c_int + ... # // Get metadata value as a string by index -# LLAMA_API int llama_model_meta_val_str_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size); +# LLAMA_API int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); +@ctypes_function( + "llama_model_meta_val_str_by_index", + [ + llama_model_p_ctypes, + ctypes.c_int32, + ctypes.c_char_p, + ctypes.c_size_t, + ], + ctypes.c_int32, +) def llama_model_meta_val_str_by_index( - model: llama_model_p, i: Union[c_int, int], buf: bytes, buf_size: int + model: llama_model_p, + i: Union[ctypes.c_int, int], + buf: Union[bytes, CtypesArray[ctypes.c_char]], + buf_size: int, + /, ) -> int: """Get metadata value as a string by index""" - return _lib.llama_model_meta_val_str_by_index(model, i, buf, buf_size) - - -_lib.llama_model_meta_val_str_by_index.argtypes = [ - llama_model_p, - c_int, - c_char_p, - c_size_t, -] -_lib.llama_model_meta_val_str_by_index.restype = c_int + ... # // Get a string describing the model type -# LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); +# LLAMA_API int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); +@ctypes_function( + "llama_model_desc", + [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_size_t], + ctypes.c_int32, +) def llama_model_desc( - model: llama_model_p, buf: bytes, buf_size: Union[c_size_t, int] + model: llama_model_p, + buf: Union[bytes, CtypesArray[ctypes.c_char]], + buf_size: Union[ctypes.c_size_t, int], + /, ) -> int: """Get a string describing the model type""" - return _lib.llama_model_desc(model, buf, buf_size) - - -_lib.llama_model_desc.argtypes = [llama_model_p, c_char_p, c_size_t] -_lib.llama_model_desc.restype = c_int + ... # // Returns the total size of all the tensors in the model in bytes # LLAMA_API uint64_t llama_model_size(const struct llama_model * model); -def llama_model_size(model: llama_model_p) -> int: +@ctypes_function("llama_model_size", [llama_model_p_ctypes], ctypes.c_uint64) +def llama_model_size(model: llama_model_p, /) -> int: """Returns the total size of all the tensors in the model in bytes""" - return _lib.llama_model_size(model) + ... -_lib.llama_model_size.argtypes = [llama_model_p] -_lib.llama_model_size.restype = ctypes.c_uint64 +# // Get the default chat template. Returns nullptr if not available +# // If name is NULL, returns the default chat template +# LLAMA_API const char * llama_model_chat_template(const struct llama_model * model, const char * name); +@ctypes_function("llama_model_chat_template", [llama_model_p_ctypes, ctypes.c_char_p], ctypes.c_char_p) +def llama_model_chat_template(model: llama_model_p, name: Optional[bytes], /) -> Optional[bytes]: + """Get the default chat template. Returns None if not available + If name is None, returns the default chat template""" + ... # // Returns the total number of parameters in the model # LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model); -def llama_model_n_params(model: llama_model_p) -> int: +@ctypes_function("llama_model_n_params", [llama_model_p_ctypes], ctypes.c_uint64) +def llama_model_n_params(model: llama_model_p, /) -> int: """Returns the total number of parameters in the model""" - return _lib.llama_model_n_params(model) + ... + +# // Returns true if the model contains an encoder that requires llama_encode() call +# LLAMA_API bool llama_model_has_encoder(const struct llama_model * model); +@ctypes_function("llama_model_has_encoder", [llama_model_p_ctypes], ctypes.c_bool) +def llama_model_has_encoder(model: llama_model_p, /) -> bool: + """Returns true if the model contains an encoder that requires llama_encode() call""" + ... -_lib.llama_model_n_params.argtypes = [llama_model_p] -_lib.llama_model_n_params.restype = ctypes.c_uint64 +# // Returns true if the model contains a decoder that requires llama_decode() call +# LLAMA_API bool llama_model_has_decoder(const struct llama_model * model); +@ctypes_function("llama_model_has_decoder", [llama_model_p_ctypes], ctypes.c_bool) +def llama_model_has_decoder(model: llama_model_p, /) -> bool: + """Returns true if the model contains a decoder that requires llama_decode() call""" + ... -# // Get a llama model tensor -# LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name); -def llama_get_model_tensor( - model: llama_model_p, name: Union[c_char_p, bytes] -) -> c_void_p: - """Get a llama model tensor""" - return _lib.llama_get_model_tensor(model, name) +# // For encoder-decoder models, this function returns id of the token that must be provided +# // to the decoder to start generating output sequence. For other models, it returns -1. +# LLAMA_API llama_token llama_model_decoder_start_token(const struct llama_model * model); +@ctypes_function( + "llama_model_decoder_start_token", [llama_model_p_ctypes], ctypes.c_int32 +) +def llama_model_decoder_start_token(model: llama_model_p, /) -> int: + """For encoder-decoder models, this function returns id of the token that must be provided + to the decoder to start generating output sequence. For other models, it returns -1. + """ + ... -_lib.llama_get_model_tensor.argtypes = [llama_model_p, c_char_p] -_lib.llama_get_model_tensor.restype = c_void_p + +# // Returns true if the model is recurrent (like Mamba, RWKV, etc.) +# LLAMA_API bool llama_model_is_recurrent(const struct llama_model * model); +@ctypes_function("llama_model_is_recurrent", [llama_model_p_ctypes], ctypes.c_bool) +def llama_model_is_recurrent(model: llama_model_p, /) -> bool: + """Returns true if the model is recurrent (like Mamba, RWKV, etc.)""" + ... # // Returns 0 on success -# LLAMA_API int llama_model_quantize( +# LLAMA_API uint32_t llama_model_quantize( # const char * fname_inp, # const char * fname_out, # const llama_model_quantize_params * params); +@ctypes_function( + "llama_model_quantize", + [ + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.POINTER(llama_model_quantize_params), + ], + ctypes.c_uint32, +) def llama_model_quantize( fname_inp: bytes, fname_out: bytes, - params, # type: POINTER(llama_model_quantize_params) # type: ignore + params: CtypesPointerOrRef[llama_model_quantize_params], + /, ) -> int: """Returns 0 on success""" - return _lib.llama_model_quantize(fname_inp, fname_out, params) + ... -_lib.llama_model_quantize.argtypes = [ - c_char_p, - c_char_p, - POINTER(llama_model_quantize_params), -] -_lib.llama_model_quantize.restype = c_int +# // Load a LoRA adapter from file +# LLAMA_API struct llama_adapter_lora * llama_adapter_lora_init( +# struct llama_model * model, +# const char * path_lora); +@ctypes_function( + "llama_adapter_lora_init", + [llama_model_p_ctypes, ctypes.c_char_p], + llama_adapter_lora_p_ctypes, +) +def llama_adapter_lora_init( + model: llama_model_p, path_lora: bytes, / +) -> Optional[llama_adapter_lora_p]: + ... + + +# // Manually free a LoRA adapter +# // Note: loaded adapters will be free when the associated model is deleted +# LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter); +@ctypes_function( + "llama_adapter_lora_free", + [llama_adapter_lora_p_ctypes], + None, +) +def llama_adapter_lora_free(adapter: llama_adapter_lora_p, /): + ... -# // Apply a LoRA adapter to a loaded model -# // path_base_model is the path to a higher quality model to use as a base for -# // the layers modified by the adapter. Can be NULL to use the current loaded model. -# // The model needs to be reloaded before applying a new adapter, otherwise the adapter -# // will be applied on top of the previous one -# // Returns 0 on success -# LLAMA_API DEPRECATED(int llama_apply_lora_from_file( +# // The following functions operate on a llama_context, hence the naming: llama_verb_... + + +# // Add a loaded LoRA adapter to given context +# // This will not modify model's weight +# LLAMA_API int32_t llama_set_adapter_lora( # struct llama_context * ctx, -# const char * path_lora, -# float scale, -# const char * path_base_model, -# int n_threads), -# "use llama_model_apply_lora_from_file instead"); -def llama_apply_lora_from_file( - ctx: llama_context_p, - path_lora: Union[c_char_p, bytes], - scale: Union[c_float, float], - path_base_model: Union[c_char_p, bytes], - n_threads: Union[c_int, int], +# struct llama_adapter_lora * adapter, +# float scale); +@ctypes_function( + "llama_set_adapter_lora", + [llama_context_p_ctypes, llama_adapter_lora_p_ctypes, ctypes.c_float], + ctypes.c_int32, +) +def llama_set_adapter_lora( + ctx: llama_context_p, adapter: llama_adapter_lora_p, scale: float, / ) -> int: - """Apply a LoRA adapter to a loaded model - path_base_model is the path to a higher quality model to use as a base for - the layers modified by the adapter. Can be NULL to use the current loaded model. - The model needs to be reloaded before applying a new adapter, otherwise the adapter - will be applied on top of the previous one - Returns 0 on success""" - return _lib.llama_apply_lora_from_file( - ctx, path_lora, scale, path_base_model, n_threads - ) + """Add a loaded LoRA adapter to given context + This will not modify model's weight""" + ... -_lib.llama_apply_lora_from_file.argtypes = [ - llama_context_p, - c_char_p, - c_float, - c_char_p, - c_int, -] -_lib.llama_apply_lora_from_file.restype = c_int +# // Remove a specific LoRA adapter from given context +# // Return -1 if the adapter is not present in the context +# LLAMA_API int32_t llama_rm_adapter_lora( +# struct llama_context * ctx, +# struct llama_adapter_lora * adapter); +@ctypes_function( + "llama_rm_adapter_lora", + [llama_context_p_ctypes, llama_adapter_lora_p_ctypes], + ctypes.c_int32, +) +def llama_rm_adapter_lora( + ctx: llama_context_p, adapter: llama_adapter_lora_p, / +) -> int: + """Remove a specific LoRA adapter from given context + Return -1 if the adapter is not present in the context""" + ... -# LLAMA_API int llama_model_apply_lora_from_file( -# const struct llama_model * model, -# const char * path_lora, -# float scale, -# const char * path_base_model, -# int n_threads); -def llama_model_apply_lora_from_file( - model: llama_model_p, - path_lora: Union[c_char_p, bytes], - scale: Union[c_float, float], - path_base_model: Union[c_char_p, bytes], - n_threads: Union[c_int, int], +# // Remove all LoRA adapters from given context +# LLAMA_API void llama_clear_adapter_lora(struct llama_context * ctx); +@ctypes_function( + "llama_clear_adapter_lora", + [llama_context_p_ctypes], + None, +) +def llama_clear_adapter_lora(ctx: llama_context_p, /): + """Remove all LoRA adapters from given context""" + ... + + +# // Apply a loaded control vector to a llama_context, or if data is NULL, clear +# // the currently loaded vector. +# // n_embd should be the size of a single layer's control, and data should point +# // to an n_embd x n_layers buffer starting from layer 1. +# // il_start and il_end are the layer range the vector should apply to (both inclusive) +# // See llama_control_vector_load in common to load a control vector. +# LLAMA_API int32_t llama_apply_adapter_cvec( +# struct llama_context * ctx, +# const float * data, +# size_t len, +# int32_t n_embd, +# int32_t il_start, +# int32_t il_end); +@ctypes_function( + "llama_apply_adapter_cvec", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_float), + ctypes.c_size_t, + ctypes.c_int32, + ctypes.c_int32, + ctypes.c_int32, + ], + ctypes.c_int32, +) +def llama_apply_adapter_cvec( + ctx: llama_context_p, + data: CtypesPointerOrRef[ctypes.c_float], + len: int, + n_embd: int, + il_start: int, + il_end: int, + /, ) -> int: - return _lib.llama_model_apply_lora_from_file( - model, path_lora, scale, path_base_model, n_threads - ) - + """Apply a loaded control vector to a llama_context, or if data is NULL, clear + the currently loaded vector. + n_embd should be the size of a single layer's control, and data should point + to an n_embd x n_layers buffer starting from layer 1. + il_start and il_end are the layer range the vector should apply to (both inclusive) + See llama_control_vector_load in common to load a control vector.""" + ... -_lib.llama_model_apply_lora_from_file.argtypes = [ - llama_model_p, - c_char_p, - c_float, - c_char_p, - c_int, -] -_lib.llama_model_apply_lora_from_file.restype = c_int # // # // KV cache @@ -992,7 +1748,16 @@ def llama_model_apply_lora_from_file( # // May be negative if the cell is not populated. # llama_pos pos; # }; -class llama_kv_cache_view_cell(Structure): +class llama_kv_cache_view_cell(ctypes.Structure): + """Information associated with an individual cell in the KV cache view. + + Attributes: + pos (llama_pos): The position for this cell. Takes KV cache shifts into account. + May be negative if the cell is not populated.""" + + if TYPE_CHECKING: + pos: llama_pos + _fields_ = [("pos", llama_pos)] @@ -1004,7 +1769,7 @@ class llama_kv_cache_view_cell(Structure): # // Maximum number of sequences that can exist in a cell. It's not an error # // if there are more sequences in a cell than this value, however they will # // not be visible in the view cells_sequences. -# int32_t n_max_seq; +# int32_t n_seq_max; # // Number of tokens in the cache. For example, if there are two populated # // cells, the first with 1 sequence id in it and the second with 2 sequence @@ -1025,401 +1790,863 @@ class llama_kv_cache_view_cell(Structure): # struct llama_kv_cache_view_cell * cells; -# // The sequences for each cell. There will be n_max_seq items per cell. +# // The sequences for each cell. There will be n_seq_max items per cell. # llama_seq_id * cells_sequences; # }; -class llama_kv_cache_view(Structure): +class llama_kv_cache_view(ctypes.Structure): + if TYPE_CHECKING: + n_cells: int + n_max_seq: int + token_count: int + used_cells: int + max_contiguous: int + max_contiguous_idx: int + cells: CtypesArray[llama_kv_cache_view_cell] + cells_sequences: CtypesArray[llama_seq_id] + _fields_ = [ - ("n_cells", c_int32), - ("n_max_seq", c_int32), - ("token_count", c_int32), - ("used_cells", c_int32), - ("max_contiguous", c_int32), - ("max_contiguous_idx", c_int32), - ("cells", POINTER(llama_kv_cache_view_cell)), - ("cells_sequences", POINTER(llama_seq_id)), + ("n_cells", ctypes.c_int32), + ("n_max_seq", ctypes.c_int32), + ("token_count", ctypes.c_int32), + ("used_cells", ctypes.c_int32), + ("max_contiguous", ctypes.c_int32), + ("max_contiguous_idx", ctypes.c_int32), + ("cells", ctypes.POINTER(llama_kv_cache_view_cell)), + ("cells_sequences", ctypes.POINTER(llama_seq_id)), ] -llama_kv_cache_view_p = POINTER(llama_kv_cache_view) +llama_kv_cache_view_p = ctypes.POINTER(llama_kv_cache_view) # // Create an empty KV cache view. (use only for debugging purposes) -# LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq); +# LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max); +@ctypes_function( + "llama_kv_cache_view_init", + [llama_context_p_ctypes, ctypes.c_int32], + llama_kv_cache_view, +) def llama_kv_cache_view_init( - ctx: llama_context_p, n_max_seq: Union[c_int32, int] + ctx: llama_context_p, n_seq_max: Union[ctypes.c_int32, int], / ) -> llama_kv_cache_view: """Create an empty KV cache view. (use only for debugging purposes)""" - return _lib.llama_kv_cache_view_init(ctx, n_max_seq) - - -_lib.llama_kv_cache_view_init.argtypes = [llama_context_p, c_int32] -_lib.llama_kv_cache_view_init.restype = llama_kv_cache_view + ... # // Free a KV cache view. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view); -def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]"): # type: ignore +@ctypes_function("llama_kv_cache_view_free", [llama_kv_cache_view_p], None) +def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]", /): # type: ignore """Free a KV cache view. (use only for debugging purposes)""" - return _lib.llama_kv_cache_view_free(view) - - -_lib.llama_kv_cache_view_free.argtypes = [llama_kv_cache_view_p] -_lib.llama_kv_cache_view_free.restype = None + ... # // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes) # LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view); -def llama_kv_cache_view_update(ctx: llama_context_p, view: "ctypes.pointer[llama_kv_cache_view]"): # type: ignore +@ctypes_function( + "llama_kv_cache_view_update", [llama_context_p_ctypes, llama_kv_cache_view_p], None +) +def llama_kv_cache_view_update(ctx: llama_context_p, view: CtypesPointerOrRef[llama_kv_cache_view], /): # type: ignore """Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)""" - return _lib.llama_kv_cache_view_update(ctx, view) - - -_lib.llama_kv_cache_view_update.argtypes = [llama_context_p, llama_kv_cache_view_p] -_lib.llama_kv_cache_view_update.restype = None + ... # // Returns the number of tokens in the KV cache (slow, use only for debug) # // If a KV cell has multiple sequences assigned to it, it will be counted multiple times -# LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx); -def llama_get_kv_cache_token_count(ctx: llama_context_p) -> int: +# LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx); +@ctypes_function( + "llama_kv_self_n_tokens", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_kv_self_n_tokens(ctx: llama_context_p, /) -> int: """Returns the number of tokens in the KV cache (slow, use only for debug) If a KV cell has multiple sequences assigned to it, it will be counted multiple times """ - return _lib.llama_get_kv_cache_token_count(ctx) + ... -_lib.llama_get_kv_cache_token_count.argtypes = [llama_context_p] -_lib.llama_get_kv_cache_token_count.restype = c_int +# DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx), +# "use llama_kv_self_n_tokens instead"); +@ctypes_function( + "llama_get_kv_cache_token_count", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_get_kv_cache_token_count(ctx: llama_context_p, /) -> int: + """Returns the number of tokens in the KV cache (slow, use only for debug) + If a KV cell has multiple sequences assigned to it, it will be counted multiple times + """ + ... # // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) -# LLAMA_API int llama_get_kv_cache_used_cells(const struct llama_context * ctx); -def llama_get_kv_cache_used_cells(ctx: llama_context_p) -> int: +# LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx); +@ctypes_function( + "llama_kv_self_used_cells", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_kv_self_used_cells(ctx: llama_context_p, /) -> int: """Returns the number of used KV cells (i.e. have at least one sequence assigned to them)""" - return _lib.llama_get_kv_cache_used_cells(ctx) + ... -_lib.llama_get_kv_cache_used_cells.argtypes = [llama_context_p] -_lib.llama_get_kv_cache_used_cells.restype = c_int +# DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx), +# "use llama_kv_self_used_cells instead"); +@ctypes_function( + "llama_get_kv_cache_used_cells", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_get_kv_cache_used_cells(ctx: llama_context_p, /) -> int: + """Returns the number of used KV cells (i.e. have at least one sequence assigned to them)""" + ... -# // Clear the KV cache -# LLAMA_API void llama_kv_cache_clear( +# // Clear the KV cache - both cell info is erased and KV data is zeroed +# LLAMA_API void llama_kv_self_clear( # struct llama_context * ctx); -def llama_kv_cache_clear(ctx: llama_context_p): - """Clear the KV cache""" - return _lib.llama_kv_cache_clear(ctx) - +@ctypes_function( + "llama_kv_self_clear", [llama_context_p_ctypes], None +) +def llama_kv_self_clear(ctx: llama_context_p, /): + """Clear the KV cache - both cell info is erased and KV data is zeroed""" + ... -_lib.llama_kv_cache_clear.argtypes = [llama_context_p] -_lib.llama_kv_cache_clear.restype = None +# NOTE: Deprecated +@ctypes_function("llama_kv_self_clear", [llama_context_p_ctypes], None) +def llama_kv_cache_clear(ctx: llama_context_p, /): + """Clear the KV cache""" + ... # // Removes all tokens that belong to the specified sequence and have positions in [p0, p1) +# // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails # // seq_id < 0 : match any sequence # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_rm( +# LLAMA_API bool llama_kv_cache_seq_rm( # struct llama_context * ctx, # llama_seq_id seq_id, # llama_pos p0, # llama_pos p1); +@ctypes_function( + "llama_kv_cache_seq_rm", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + ], + ctypes.c_bool, +) def llama_kv_cache_seq_rm( ctx: llama_context_p, seq_id: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], -): + /, +) -> bool: """Removes all tokens that belong to the specified sequence and have positions in [p0, p1) + + Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails + seq_id < 0 : match any sequence p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" - return _lib.llama_kv_cache_seq_rm(ctx, seq_id, p0, p1) - - -_lib.llama_kv_cache_seq_rm.argtypes = [ - llama_context_p, - llama_seq_id, - llama_pos, - llama_pos, -] -_lib.llama_kv_cache_seq_rm.restype = None + ... # // Copy all tokens that belong to the specified sequence to another sequence # // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_cp( +# LLAMA_API void llama_kv_self_seq_cp( # struct llama_context * ctx, # llama_seq_id seq_id_src, # llama_seq_id seq_id_dst, # llama_pos p0, # llama_pos p1); +@ctypes_function( + "llama_kv_self_seq_cp", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_seq_id, + llama_pos, + llama_pos, + ], + None, +) +def llama_kv_self_seq_cp( + ctx: llama_context_p, + seq_id_src: Union[llama_seq_id, int], + seq_id_dst: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + /, +): + """Copy all tokens that belong to the specified sequence to another sequence + Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# NOTE: Deprecated +@ctypes_function( + "llama_kv_self_seq_cp", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_seq_id, + llama_pos, + llama_pos, + ], + None, +) def llama_kv_cache_seq_cp( ctx: llama_context_p, seq_id_src: Union[llama_seq_id, int], seq_id_dst: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], + /, ): """Copy all tokens that belong to the specified sequence to another sequence Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" - return _lib.llama_kv_cache_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1) - - -_lib.llama_kv_cache_seq_cp.argtypes = [ - llama_context_p, - llama_seq_id, - llama_seq_id, - llama_pos, - llama_pos, -] -_lib.llama_kv_cache_seq_cp.restype = None + ... # // Removes all tokens that do not belong to the specified sequence -# LLAMA_API void llama_kv_cache_seq_keep( +# LLAMA_API void llama_kv_self_seq_keep( # struct llama_context * ctx, # llama_seq_id seq_id); -def llama_kv_cache_seq_keep( - ctx: llama_context_p, - seq_id: Union[llama_seq_id, int], -): +@ctypes_function( + "llama_kv_self_seq_keep", [llama_context_p_ctypes, llama_seq_id], None +) +def llama_kv_self_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): """Removes all tokens that do not belong to the specified sequence""" - return _lib.llama_kv_cache_seq_keep(ctx, seq_id) + ... -_lib.llama_kv_cache_seq_keep.argtypes = [llama_context_p, llama_seq_id] -_lib.llama_kv_cache_seq_keep.restype = None +# NOTE: Deprecated +@ctypes_function( + "llama_kv_self_seq_keep", [llama_context_p_ctypes, llama_seq_id], None +) +def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): + """Removes all tokens that do not belong to the specified sequence""" + ... + # // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) -# // If the KV cache is RoPEd, the KV data is updated accordingly +# // If the KV cache is RoPEd, the KV data is updated accordingly: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_cache_update() # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_shift( +# LLAMA_API void llama_kv_cache_seq_add( # struct llama_context * ctx, # llama_seq_id seq_id, # llama_pos p0, # llama_pos p1, # llama_pos delta); -def llama_kv_cache_seq_shift( +@ctypes_function( + "llama_kv_self_seq_add", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + llama_pos, + ], + None, +) +def llama_kv_self_seq_add( ctx: llama_context_p, seq_id: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], delta: Union[llama_pos, int], + /, ): """Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) - If the KV cache is RoPEd, the KV data is updated accordingly + If the KV cache is RoPEd, the KV data is updated accordingly: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update() p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" - return _lib.llama_kv_cache_seq_shift(ctx, seq_id, p0, p1, delta) - - -_lib.llama_kv_cache_seq_shift.argtypes = [ - llama_context_p, - llama_seq_id, - llama_pos, - llama_pos, - llama_pos, -] -_lib.llama_kv_cache_seq_shift.restype = None + ... -# // -# // State / sessions -# // + +# // NOTE: Deprecated +# // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) +# // If the KV cache is RoPEd, the KV data is updated accordingly: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_cache_update() +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# LLAMA_API void llama_kv_cache_seq_add( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# llama_pos delta); +@ctypes_function( + "llama_kv_self_seq_add", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + llama_pos, + ], + None, +) +def llama_kv_cache_seq_add( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + delta: Union[llama_pos, int], + /, +): + """Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) + If the KV cache is RoPEd, the KV data is updated accordingly: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update() + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# // Integer division of the positions by factor of `d > 1` +# // If the KV cache is RoPEd, the KV data is updated accordingly +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# LLAMA_API void llama_kv_cache_seq_div( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# int d); +@ctypes_function( + "llama_kv_self_seq_div", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + ctypes.c_int, + ], + None, +) +def llama_kv_self_seq_div( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + d: Union[ctypes.c_int, int], + /, +): + """Integer division of the positions by factor of `d > 1` + If the KV cache is RoPEd, the KV data is updated accordingly + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# // NOTE: Deprecated +# // Integer division of the positions by factor of `d > 1` +# // If the KV cache is RoPEd, the KV data is updated accordingly +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# LLAMA_API void llama_kv_cache_seq_div( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# int d); +@ctypes_function( + "llama_kv_self_seq_div", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + ctypes.c_int, + ], + None, +) +def llama_kv_cache_seq_div( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + d: Union[ctypes.c_int, int], + /, +): + """Integer division of the positions by factor of `d > 1` + If the KV cache is RoPEd, the KV data is updated accordingly + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# // Returns the largest position present in the KV cache for the specified sequence +# LLAMA_API llama_pos llama_kv_self_seq_pos_max( +# struct llama_context * ctx, +# llama_seq_id seq_id); +@ctypes_function( + "llama_kv_self_seq_pos_max", [llama_context_p_ctypes, llama_seq_id], llama_pos +) +def llama_kv_self_seq_pos_max( + ctx: llama_context_p, seq_id: Union[llama_seq_id, int], / +) -> int: + """Returns the largest position present in the KV cache for the specified sequence""" + ... + + +# // Defragment the KV cache +# // This will be applied: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_self_update() +# LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx); +@ctypes_function("llama_kv_self_defrag", [llama_context_p_ctypes], None) +def llama_kv_self_defrag(ctx: llama_context_p, /): + """Defragment the KV cache + This will be applied: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update()""" + ... + + +# NOTE: Deprecated +# // Defragment the KV cache +# // This will be applied: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_self_update() +# LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx); +@ctypes_function("llama_kv_cache_defrag", [llama_context_p_ctypes], None) +def llama_kv_cache_defrag(ctx: llama_context_p, /): + """Defragment the KV cache + This will be applied: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update()""" + ... + + +# // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) +# LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); +@ctypes_function("llama_kv_self_update", [llama_context_p_ctypes], None) +def llama_kv_self_update(ctx: llama_context_p, /): + """Apply the KV cache updates (such as K-shifts, defragmentation, etc.)""" + ... + +# // NOTE: Deprecated +# // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) +# LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); +@ctypes_function("llama_kv_self_update", [llama_context_p_ctypes], None) +def llama_kv_cache_update(ctx: llama_context_p, /): + """Apply the KV cache updates (such as K-shifts, defragmentation, etc.)""" + ... + + +# // Check if the context supports KV cache shifting +# LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); +@ctypes_function("llama_kv_self_can_shift", [llama_context_p_ctypes], ctypes.c_bool) +def llama_kv_self_can_shift(ctx: llama_context_p, /) -> bool: + """Check if the context supports KV cache shifting""" + ... + + +# // NOTE: Deprecated +# // Check if the context supports KV cache shifting +# LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); +@ctypes_function("llama_kv_self_can_shift", [llama_context_p_ctypes], ctypes.c_bool) +def llama_kv_cache_can_shift(ctx: llama_context_p, /) -> bool: + """Check if the context supports KV cache shifting""" + ... -# Returns the maximum size in bytes of the state (rng, logits, embedding -# and kv_cache) - will often be smaller after compacting tokens -# LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx); -def llama_get_state_size(ctx: llama_context_p) -> int: +# // +# // State / sessions +# // + + +# // Returns the *actual* size in bytes of the state +# // (logits, embedding and kv_cache) +# // Only use when saving the state, not when restoring it, otherwise the size may be too small. +# LLAMA_API size_t llama_state_get_size(struct llama_context * ctx); +@ctypes_function("llama_state_get_size", [llama_context_p_ctypes], ctypes.c_size_t) +def llama_state_get_size(ctx: llama_context_p, /) -> int: + """Returns the *actual* size in bytes of the state (rng, logits, embedding and kv_cache) - will often be smaller after compacting tokens""" + ... + + +# LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx), +# "use llama_state_get_size instead"); +@ctypes_function("llama_get_state_size", [llama_context_p_ctypes], ctypes.c_size_t) +def llama_get_state_size(ctx: llama_context_p, /) -> int: """Returns the maximum size in bytes of the state (rng, logits, embedding and kv_cache) - will often be smaller after compacting tokens""" - return _lib.llama_get_state_size(ctx) + ... -_lib.llama_get_state_size.argtypes = [llama_context_p] -_lib.llama_get_state_size.restype = c_size_t +# // Copies the state to the specified destination address. +# // Destination needs to have allocated enough memory. +# // Returns the number of bytes copied +# LLAMA_API size_t llama_state_get_data( +# struct llama_context * ctx, +# uint8_t * dst, +# size_t size); +@ctypes_function( + "llama_state_get_data", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_uint8), + ctypes.c_size_t, + ], + ctypes.c_size_t, +) +def llama_state_get_data( + ctx: llama_context_p, + dst: CtypesArray[ctypes.c_uint8], + size: Union[ctypes.c_size_t, int], + /, +) -> int: + """Copies the state to the specified destination address. + Destination needs to have allocated enough memory. + Returns the number of bytes copied""" + ... -# Copies the state to the specified destination address. -# Destination needs to have allocated enough memory. -# Returns the number of bytes copied -# LLAMA_API size_t llama_copy_state_data( +# LLAMA_API DEPRECATED(size_t llama_copy_state_data( # struct llama_context * ctx, -# uint8_t * dst); +# uint8_t * dst), +# "use llama_state_get_data instead"); +@ctypes_function( + "llama_copy_state_data", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_uint8), + ], + ctypes.c_size_t, +) def llama_copy_state_data( - ctx: llama_context_p, dst # type: Array[c_uint8] + ctx: llama_context_p, dst: CtypesArray[ctypes.c_uint8], / ) -> int: """Copies the state to the specified destination address. Destination needs to have allocated enough memory. Returns the number of bytes copied""" - return _lib.llama_copy_state_data(ctx, dst) + ... -_lib.llama_copy_state_data.argtypes = [llama_context_p, c_uint8_p] -_lib.llama_copy_state_data.restype = c_size_t +# // Set the state reading from the specified address +# // Returns the number of bytes read +# LLAMA_API size_t llama_state_set_data( +# struct llama_context * ctx, +# const uint8_t * src, +# size_t size); +@ctypes_function( + "llama_state_set_data", + [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8), ctypes.c_size_t], + ctypes.c_size_t, +) +def llama_state_set_data( + ctx: llama_context_p, + src: CtypesArray[ctypes.c_uint8], + size: Union[ctypes.c_size_t, int], + /, +) -> int: + """Set the state reading from the specified address + Returns the number of bytes read""" + ... -# Set the state reading from the specified address -# Returns the number of bytes read -# LLAMA_API size_t llama_set_state_data( +# LLAMA_API DEPRECATED(size_t llama_set_state_data( # struct llama_context * ctx, -# uint8_t * src); +# const uint8_t * src), +# "use llama_state_set_data instead"); +@ctypes_function( + "llama_set_state_data", + [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_uint8)], + ctypes.c_size_t, +) def llama_set_state_data( - ctx: llama_context_p, src # type: Array[c_uint8] + ctx: llama_context_p, src: CtypesArray[ctypes.c_uint8], / ) -> int: """Set the state reading from the specified address""" - return _lib.llama_set_state_data(ctx, src) - - -_lib.llama_set_state_data.argtypes = [llama_context_p, c_uint8_p] -_lib.llama_set_state_data.restype = c_size_t + ... # Save/load session file -# LLAMA_API bool llama_load_session_file( +# LLAMA_API bool llama_state_load_file( # struct llama_context * ctx, # const char * path_session, # llama_token * tokens_out, # size_t n_token_capacity, # size_t * n_token_count_out); +@ctypes_function( + "llama_state_load_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_token_p, + ctypes.c_size_t, + ctypes.POINTER(ctypes.c_size_t), + ], + ctypes.c_bool, +) +def llama_state_load_file( + ctx: llama_context_p, + path_session: bytes, + tokens_out: CtypesArray[llama_token], + n_token_capacity: Union[ctypes.c_size_t, int], + n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], + /, +) -> bool: + ... + + +# LLAMA_API DEPRECATED(bool llama_load_session_file( +# struct llama_context * ctx, +# const char * path_session, +# llama_token * tokens_out, +# size_t n_token_capacity, +# size_t * n_token_count_out), +# "use llama_state_load_file instead"); +@ctypes_function( + "llama_load_session_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_token_p, + ctypes.c_size_t, + ctypes.POINTER(ctypes.c_size_t), + ], + ctypes.c_size_t, +) def llama_load_session_file( ctx: llama_context_p, path_session: bytes, - tokens_out, # type: Array[llama_token] - n_token_capacity: Union[c_size_t, int], - n_token_count_out, # type: _Pointer[c_size_t] + tokens_out: CtypesArray[llama_token], + n_token_capacity: Union[ctypes.c_size_t, int], + n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], + /, ) -> int: - return _lib.llama_load_session_file( - ctx, path_session, tokens_out, n_token_capacity, n_token_count_out - ) + ... -_lib.llama_load_session_file.argtypes = [ - llama_context_p, - c_char_p, - llama_token_p, - c_size_t, - c_size_t_p, -] -_lib.llama_load_session_file.restype = c_size_t +# LLAMA_API bool llama_state_save_file( +# struct llama_context * ctx, +# const char * path_session, +# const llama_token * tokens, +# size_t n_token_count); +@ctypes_function( + "llama_state_save_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_token_p, + ctypes.c_size_t, + ], + ctypes.c_bool, +) +def llama_state_save_file( + ctx: llama_context_p, + path_session: bytes, + tokens: CtypesArray[llama_token], + n_token_count: Union[ctypes.c_size_t, int], + /, +) -> bool: + ... -# LLAMA_API bool llama_save_session_file( +# LLAMA_API DEPRECATED(bool llama_save_session_file( # struct llama_context * ctx, # const char * path_session, # const llama_token * tokens, -# size_t n_token_count); +# size_t n_token_count), +# "use llama_state_save_file instead"); +@ctypes_function( + "llama_save_session_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_token_p, + ctypes.c_size_t, + ], + ctypes.c_size_t, +) def llama_save_session_file( ctx: llama_context_p, path_session: bytes, - tokens, # type: Array[llama_token] - n_token_count: Union[c_size_t, int], + tokens: CtypesArray[llama_token], + n_token_count: Union[ctypes.c_size_t, int], + /, ) -> int: - return _lib.llama_save_session_file(ctx, path_session, tokens, n_token_count) + ... -_lib.llama_save_session_file.argtypes = [ - llama_context_p, - c_char_p, - llama_token_p, - c_size_t, -] -_lib.llama_save_session_file.restype = c_size_t +# // Get the exact size needed to copy the KV cache of a single sequence +# LLAMA_API size_t llama_state_seq_get_size( +# struct llama_context * ctx, +# llama_seq_id seq_id); +@ctypes_function( + "llama_state_seq_get_size", + [llama_context_p_ctypes, llama_seq_id], + ctypes.c_size_t, +) +def llama_state_seq_get_size(ctx: llama_context_p, seq_id: llama_seq_id, /) -> int: + """Get the exact size needed to copy the KV cache of a single sequence""" + ... -# // -# // Decoding -# // +# // Copy the KV cache of a single sequence into the specified buffer +# LLAMA_API size_t llama_state_seq_get_data( +# struct llama_context * ctx, +# uint8_t * dst, +# size_t size, +# llama_seq_id seq_id); +@ctypes_function( + "llama_state_seq_get_data", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_uint8), + ctypes.c_size_t, + llama_seq_id, + ], + ctypes.c_size_t, +) +def llama_state_seq_get_data( + ctx: llama_context_p, + dst: CtypesArray[ctypes.c_uint8], + size: Union[ctypes.c_size_t, int], + seq_id: llama_seq_id, + /, +) -> int: + """Copy the KV cache of a single sequence into the specified buffer""" + ... -# // Run the llama inference to obtain the logits and probabilities for the next token(s). -# // tokens + n_tokens is the provided batch of new tokens to process -# // n_past is the number of tokens to use from previous eval calls -# // Returns 0 on success -# // DEPRECATED: use llama_decode() instead -# LLAMA_API DEPRECATED(int llama_eval( + +# // Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence +# // Returns: +# // - Positive: Ok +# // - Zero: Failed to load +# LLAMA_API size_t llama_state_seq_set_data( # struct llama_context * ctx, -# llama_token * tokens, -# int32_t n_tokens, -# int n_past), -# "use llama_decode() instead"); -def llama_eval( +# const uint8_t * src, +# size_t size, +# llama_seq_id dest_seq_id); +@ctypes_function( + "llama_state_seq_set_data", + [ + llama_context_p_ctypes, + ctypes.POINTER(ctypes.c_uint8), + ctypes.c_size_t, + llama_seq_id, + ], + ctypes.c_size_t, +) +def llama_state_seq_set_data( ctx: llama_context_p, - tokens, # type: Array[llama_token] - n_tokens: Union[c_int, int], - n_past: Union[c_int, int], + src: CtypesArray[ctypes.c_uint8], + size: Union[ctypes.c_size_t, int], + dest_seq_id: llama_seq_id, + /, ) -> int: - """Run the llama inference to obtain the logits and probabilities for the next token(s). - tokens + n_tokens is the provided batch of new tokens to process - n_past is the number of tokens to use from previous eval calls - Returns 0 on success - DEPRECATED: use llama_decode() instead""" - return _lib.llama_eval(ctx, tokens, n_tokens, n_past) + """Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence""" + ... -_lib.llama_eval.argtypes = [llama_context_p, llama_token_p, c_int, c_int] -_lib.llama_eval.restype = c_int +# LLAMA_API size_t llama_state_seq_save_file( +# struct llama_context * ctx, +# const char * filepath, +# llama_seq_id seq_id, +# const llama_token * tokens, +# size_t n_token_count); +@ctypes_function( + "llama_state_seq_save_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_seq_id, + llama_token_p, + ctypes.c_size_t, + ], + ctypes.c_size_t, +) +def llama_state_seq_save_file( + ctx: llama_context_p, + filepath: bytes, + seq_id: llama_seq_id, + tokens: CtypesArray[llama_token], + n_token_count: Union[ctypes.c_size_t, int], + /, +) -> int: + ... -# // Same as llama_eval, but use float matrix input directly. -# // DEPRECATED: use llama_decode() instead -# LLAMA_API DEPRECATED(int llama_eval_embd( +# LLAMA_API size_t llama_state_seq_load_file( # struct llama_context * ctx, -# float * embd, -# int32_t n_tokens, -# int n_past), -# "use llama_decode() instead"); -def llama_eval_embd( +# const char * filepath, +# llama_seq_id dest_seq_id, +# llama_token * tokens_out, +# size_t n_token_capacity, +# size_t * n_token_count_out); +@ctypes_function( + "llama_state_seq_load_file", + [ + llama_context_p_ctypes, + ctypes.c_char_p, + llama_seq_id, + llama_token_p, + ctypes.c_size_t, + ctypes.POINTER(ctypes.c_size_t), + ], + ctypes.c_size_t, +) +def llama_state_seq_load_file( ctx: llama_context_p, - embd, # type: Array[c_float] - n_tokens: Union[c_int, int], - n_past: Union[c_int, int], + filepath: bytes, + dest_seq_id: llama_seq_id, + tokens_out: CtypesArray[llama_token], + n_token_capacity: Union[ctypes.c_size_t, int], + n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], + /, ) -> int: - """Same as llama_eval, but use float matrix input directly. - DEPRECATED: use llama_decode() instead""" - return _lib.llama_eval_embd(ctx, embd, n_tokens, n_past) + ... -_lib.llama_eval_embd.argtypes = [llama_context_p, c_float_p, c_int, c_int] -_lib.llama_eval_embd.restype = c_int +# // +# // Decoding +# // -# // Return batch for single sequence of tokens starting at pos_0 +# // Return batch for single sequence of tokens +# // The sequence ID will be fixed to 0 +# // The position of the tokens will be tracked automatically by llama_decode # // # // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it # // # LLAMA_API struct llama_batch llama_batch_get_one( # llama_token * tokens, -# int32_t n_tokens, -# llama_pos pos_0, -# llama_seq_id seq_id); +# int32_t n_tokens); +@ctypes_function( + "llama_batch_get_one", + [ + llama_token_p, + ctypes.c_int32, + ], + llama_batch, +) def llama_batch_get_one( - tokens, # type: Array[llama_token] - n_tokens: Union[c_int, int], - pos_0: Union[llama_pos, int], - seq_id: llama_seq_id, + tokens: CtypesArray[llama_token], + n_tokens: Union[ctypes.c_int, int], + /, ) -> llama_batch: """Return batch for single sequence of tokens starting at pos_0 NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it """ - return _lib.llama_batch_get_one(tokens, n_tokens, pos_0, seq_id) - - -_lib.llama_batch_get_one.argtypes = [ - llama_token_p, - c_int, - llama_pos, - llama_seq_id, -] -_lib.llama_batch_get_one.restype = llama_batch + ... # // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens @@ -1433,10 +2660,14 @@ def llama_batch_get_one( # int32_t n_tokens, # int32_t embd, # int32_t n_seq_max); +@ctypes_function( + "llama_batch_init", [ctypes.c_int32, ctypes.c_int32, ctypes.c_int32], llama_batch +) def llama_batch_init( - n_tokens: Union[c_int32, int], - embd: Union[c_int32, int], - n_seq_max: Union[c_int32, int], + n_tokens: Union[ctypes.c_int32, int], + embd: Union[ctypes.c_int32, int], + n_seq_max: Union[ctypes.c_int32, int], + /, ) -> llama_batch: """Allocates a batch of tokens on the heap that can hold a maximum of n_tokens Each token can be assigned up to n_seq_max sequence ids @@ -1445,112 +2676,241 @@ def llama_batch_init( Otherwise, llama_batch.token will be allocated to store n_tokens llama_token The rest of the llama_batch members are allocated with size n_tokens All members are left uninitialized""" - return _lib.llama_batch_init(n_tokens, embd, n_seq_max) - - -_lib.llama_batch_init.argtypes = [c_int32, c_int32, c_int32] -_lib.llama_batch_init.restype = llama_batch + ... # // Frees a batch of tokens allocated with llama_batch_init() # LLAMA_API void llama_batch_free(struct llama_batch batch); -def llama_batch_free(batch: llama_batch): +@ctypes_function("llama_batch_free", [llama_batch], None) +def llama_batch_free(batch: llama_batch, /): """Frees a batch of tokens allocated with llama_batch_init()""" - return _lib.llama_batch_free(batch) + ... -_lib.llama_batch_free.argtypes = [llama_batch] -_lib.llama_batch_free.restype = None +# // Processes a batch of tokens with the ecoder part of the encoder-decoder model. +# // Stores the encoder output internally for later use by the decoder cross-attention layers. +# // 0 - success +# // < 0 - error +# LLAMA_API int32_t llama_encode( +# struct llama_context * ctx, +# struct llama_batch batch); +@ctypes_function("llama_encode", [llama_context_p_ctypes, llama_batch], ctypes.c_int32) +def llama_encode(ctx: llama_context_p, batch: llama_batch, /) -> int: + """Processes a batch of tokens with the ecoder part of the encoder-decoder model. + Stores the encoder output internally for later use by the decoder cross-attention layers. + 0 - success + < 0 - error""" + ... # // Positive return values does not mean a fatal error, but rather a warning. # // 0 - success # // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) # // < 0 - error -# LLAMA_API int llama_decode( +# LLAMA_API int32_t llama_decode( # struct llama_context * ctx, # struct llama_batch batch); -def llama_decode(ctx: llama_context_p, batch: llama_batch) -> int: +@ctypes_function("llama_decode", [llama_context_p_ctypes, llama_batch], ctypes.c_int32) +def llama_decode(ctx: llama_context_p, batch: llama_batch, /) -> int: """Positive return values does not mean a fatal error, but rather a warning. 0 - success 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) < 0 - error""" - return _lib.llama_decode(ctx, batch) - - -_lib.llama_decode.argtypes = [llama_context_p, llama_batch] -_lib.llama_decode.restype = c_int + ... # // Set the number of threads used for decoding # // n_threads is the number of threads used for generation (single token) # // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) -# LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); +# LLAMA_API void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch); +@ctypes_function( + "llama_set_n_threads", + [ + llama_context_p_ctypes, + ctypes.c_int32, + ctypes.c_int32, + ], + None, +) def llama_set_n_threads( ctx: llama_context_p, - n_threads: Union[c_uint32, int], - n_threads_batch: Union[c_uint32, int], + n_threads: Union[ctypes.c_int32, int], + n_threads_batch: Union[ctypes.c_int32, int], + /, ): """Set the number of threads used for decoding n_threads is the number of threads used for generation (single token) n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) """ - return _lib.llama_set_n_threads(ctx, n_threads, n_threads_batch) - - -_lib.llama_set_n_threads.argtypes = [llama_context_p, c_uint32, c_uint32] -_lib.llama_set_n_threads.restype = None - - -# // Token logits obtained from the last call to llama_eval() -# // The logits for the last token are stored in the last row -# // Logits for which llama_batch.logits[i] == 0 are undefined -# // Rows: n_tokens provided with llama_batch + ... + + +# // Get the number of threads used for generation of a single token. +# LLAMA_API int32_t llama_n_threads(struct llama_context * ctx); +@ctypes_function("llama_n_threads", [llama_context_p_ctypes], ctypes.c_int32) +def llama_n_threads(ctx: llama_context_p, /) -> int: + """Get the number of threads used for generation of a single token""" + ... + + +# // Get the number of threads used for prompt and batch processing (multiple token). +# LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx); +@ctypes_function("llama_n_threads_batch", [llama_context_p_ctypes], ctypes.c_int32) +def llama_n_threads_batch(ctx: llama_context_p, /) -> int: + """Get the number of threads used for prompt and batch processing (multiple token)""" + ... + + +# // Set whether the model is in embeddings mode or not +# // If true, embeddings will be returned but logits will not +# LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings); +@ctypes_function("llama_set_embeddings", [llama_context_p_ctypes, ctypes.c_bool], None) +def llama_set_embeddings(ctx: llama_context_p, embeddings: bool, /): + """Set whether the model is in embeddings model or not + If true, embeddings will be returned but logits will not""" + ... + + +# // Set whether to use causal attention or not +# // If set to true, the model will only attend to the past tokens +# LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn); +@ctypes_function("llama_set_causal_attn", [llama_context_p_ctypes, ctypes.c_bool], None) +def llama_set_causal_attn(ctx: llama_context_p, causal_attn: bool, /): + """Set whether to use causal attention or not + If set to true, the model will only attend to the past tokens""" + ... + + +# // Set whether the model is in warmup mode or not +# // If true, all model tensors are activated during llama_decode() to load and cache their weights. +# LLAMA_API void llama_set_warmup(struct llama_context * ctx, bool warmup); +@ctypes_function("llama_set_warmup", [llama_context_p_ctypes, ctypes.c_bool], None) +def llama_set_warmup(ctx: llama_context_p, warmup: bool, /): + """Set whether the model is in warmup mode or not + If true, all model tensors are activated during llama_decode() to load and cache their weights.""" + ... + + +# // Set abort callback +# LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data); +@ctypes_function( + "llama_set_abort_callback", + [llama_context_p_ctypes, ggml_abort_callback, ctypes.c_void_p], + None, +) +def llama_set_abort_callback( + ctx: llama_context_p, + abort_callback: Callable[[ctypes.c_void_p], None], + abort_callback_data: ctypes.c_void_p, + /, +): + """Set abort callback""" + ... + + +# // Wait until all computations are finished +# // This is automatically done when using one of the functions below to obtain the computation results +# // and is not necessary to call it explicitly in most cases +# LLAMA_API void llama_synchronize(struct llama_context * ctx); +@ctypes_function("llama_synchronize", [llama_context_p_ctypes], None) +def llama_synchronize(ctx: llama_context_p, /): + """Wait until all computations are finished + This is automatically done when using one of the functions below to obtain the computation results + and is not necessary to call it explicitly in most cases""" + ... + + +# // Token logits obtained from the last call to llama_decode() +# // The logits for which llama_batch.logits[i] != 0 are stored contiguously +# // in the order they have appeared in the batch. +# // Rows: number of tokens for which llama_batch.logits[i] != 0 # // Cols: n_vocab # LLAMA_API float * llama_get_logits(struct llama_context * ctx); -def llama_get_logits( - ctx: llama_context_p, -): # type: (...) -> Array[float] # type: ignore - """Token logits obtained from the last call to llama_eval() - The logits for the last token are stored in the last row - Logits for which llama_batch.logits[i] == 0 are undefined - Rows: n_tokens provided with llama_batch - Cols: n_vocab""" - return _lib.llama_get_logits(ctx) - - -_lib.llama_get_logits.argtypes = [llama_context_p] -_lib.llama_get_logits.restype = c_float_p - - -# // Logits for the ith token. Equivalent to: -# // llama_get_logits(ctx) + i*n_vocab +@ctypes_function( + "llama_get_logits", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) +) +def llama_get_logits(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: + """Token logits obtained from the last call to llama_decode() + The logits for which llama_batch.logits[i] != 0 are stored contiguously + in the order they have appeared in the batch. + Rows: number of tokens for which llama_batch.logits[i] != 0 + Cols: n_vocab + + Returns: + Pointer to the logits buffer of shape (n_tokens, n_vocab)""" + ... + + +# // Logits for the ith token. For positive indices, Equivalent to: +# // llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab +# // Negative indicies can be used to access logits in reverse order, -1 is the last logit. +# // returns NULL for invalid ids. # LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); +@ctypes_function( + "llama_get_logits_ith", + [llama_context_p_ctypes, ctypes.c_int32], + ctypes.POINTER(ctypes.c_float), +) def llama_get_logits_ith( - ctx: llama_context_p, i: Union[c_int32, int] -): # type: (...) -> Array[float] # type: ignore + ctx: llama_context_p, i: Union[ctypes.c_int32, int], / +) -> CtypesArray[ctypes.c_float]: """Logits for the ith token. Equivalent to: llama_get_logits(ctx) + i*n_vocab""" - return _lib.llama_get_logits_ith(ctx, i) - + ... -_lib.llama_get_logits_ith.argtypes = [llama_context_p, c_int32] -_lib.llama_get_logits_ith.restype = c_float_p - -# Get the embeddings for the input -# shape: [n_embd] (1-dimensional) +# // Get all output token embeddings. +# // when pooling_type == LLAMA_POOLING_TYPE_NONE or when using a generative model, +# // the embeddings for which llama_batch.logits[i] != 0 are stored contiguously +# // in the order they have appeared in the batch. +# // shape: [n_outputs*n_embd] +# // Otherwise, returns NULL. # LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); -def llama_get_embeddings( - ctx: llama_context_p, -): # type: (...) -> Array[float] # type: ignore +@ctypes_function( + "llama_get_embeddings", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) +) +def llama_get_embeddings(ctx: llama_context_p, /) -> CtypesArray[ctypes.c_float]: """Get the embeddings for the input shape: [n_embd] (1-dimensional)""" - return _lib.llama_get_embeddings(ctx) - - -_lib.llama_get_embeddings.argtypes = [llama_context_p] -_lib.llama_get_embeddings.restype = c_float_p + ... + + +# // Get the embeddings for the ith token. For positive indices, Equivalent to: +# // llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd +# // Negative indicies can be used to access embeddings in reverse order, -1 is the last embedding. +# // shape: [n_embd] (1-dimensional) +# // returns NULL for invalid ids. +# LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i); +@ctypes_function( + "llama_get_embeddings_ith", + [llama_context_p_ctypes, ctypes.c_int32], + ctypes.POINTER(ctypes.c_float), +) +def llama_get_embeddings_ith( + ctx: llama_context_p, i: Union[ctypes.c_int32, int], / +) -> CtypesArray[ctypes.c_float]: + """Get the embeddings for the ith sequence + llama_get_embeddings(ctx) + i*n_embd""" + ... + + +# // Get the embeddings for a sequence id +# // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE +# // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[1] with the rank of the sequence +# // otherwise: float[n_embd] (1-dimensional) +# LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id); +@ctypes_function( + "llama_get_embeddings_seq", + [llama_context_p_ctypes, llama_seq_id], + ctypes.POINTER(ctypes.c_float), +) +def llama_get_embeddings_seq( + ctx: llama_context_p, seq_id: Union[llama_seq_id, int], / +) -> CtypesArray[ctypes.c_float]: + """Get the embeddings for a sequence id + Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE + shape: [n_embd] (1-dimensional)""" + ... # // @@ -1558,551 +2918,989 @@ def llama_get_embeddings( # // -# LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); -def llama_token_get_text(model: llama_model_p, token: Union[llama_token, int]) -> bytes: - return _lib.llama_token_get_text(model, token) - - -_lib.llama_token_get_text.argtypes = [llama_model_p, llama_token] -_lib.llama_token_get_text.restype = c_char_p +# LLAMA_API const char * llama_vocab_get_text(const struct llama_vocab * vocab, llama_token token); +@ctypes_function( + "llama_vocab_get_text", [llama_vocab_p_ctypes, llama_token], ctypes.c_char_p +) +def llama_vocab_get_text( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bytes: + ... -# LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); -def llama_token_get_score( - model: llama_model_p, token: Union[llama_token, int] +# LLAMA_API float llama_vocab_get_score(const struct llama_vocab * vocab, llama_token token); +@ctypes_function( + "llama_vocab_get_score", [llama_vocab_p_ctypes, llama_token], ctypes.c_float +) +def llama_vocab_get_score( + vocab: llama_vocab_p, token: Union[llama_token, int], / ) -> float: - return _lib.llama_token_get_score(model, token) + ... -_lib.llama_token_get_score.argtypes = [llama_model_p, llama_token] -_lib.llama_token_get_score.restype = c_float +# LLAMA_API enum llama_token_attr llama_vocab_get_attr(const struct llama_vocab * vocab, llama_token token); +@ctypes_function( + "llama_vocab_get_attr", [llama_vocab_p_ctypes, llama_token], ctypes.c_int +) +def llama_vocab_get_attr( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> int: + ... -# LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token); -def llama_token_get_type(model: llama_model_p, token: Union[llama_token, int]) -> int: - return _lib.llama_token_get_type(model, token) +# // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.) +# LLAMA_API bool llama_vocab_is_eog(const struct llama_vocab * vocab, llama_token token); +@ctypes_function( + "llama_vocab_is_eog", [llama_vocab_p_ctypes, llama_token], ctypes.c_bool +) +def llama_vocab_is_eog(vocab: llama_vocab_p, token: Union[llama_token, int], /) -> bool: + """Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)""" + ... -_lib.llama_token_get_type.argtypes = [llama_model_p, llama_token] -_lib.llama_token_get_type.restype = ctypes.c_int +# // Identify if Token Id is a control token or a render-able token +# LLAMA_API bool llama_vocab_is_control(const struct llama_vocab * vocab, llama_token token); +@ctypes_function( + "llama_vocab_is_control", [llama_vocab_p_ctypes, llama_token], ctypes.c_bool +) +def llama_vocab_is_control( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bool: + """Identify if Token Id is a control token or a render-able token""" + ... # // Special tokens -# LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence -def llama_token_bos(model: llama_model_p) -> int: +# LLAMA_API llama_token llama_vocab_bos(const struct llama_vocab * vocab); // beginning-of-sentence +@ctypes_function("llama_vocab_bos", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_bos(vocab: llama_vocab_p, /) -> llama_token: """beginning-of-sentence""" - return _lib.llama_token_bos(model) + ... -_lib.llama_token_bos.argtypes = [llama_model_p] -_lib.llama_token_bos.restype = llama_token +# LLAMA_API llama_token llama_vocab_eos(const struct llama_vocab * vocab); // end-of-sentence +@ctypes_function("llama_vocab_eos", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_eos(vocab: llama_vocab_p, /) -> llama_token: + """end-of-sentence""" + ... -# LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence -def llama_token_eos(model: llama_model_p) -> int: - """end-of-sentence""" - return _lib.llama_token_eos(model) +# LLAMA_API llama_token llama_vocab_eot(const struct llama_vocab * vocab); // end-of-turn +@ctypes_function("llama_vocab_eot", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_eot(vocab: llama_vocab_p, /) -> llama_token: + """end-of-turn""" + ... -_lib.llama_token_eos.argtypes = [llama_model_p] -_lib.llama_token_eos.restype = llama_token +# LLAMA_API llama_token llama_vocab_sep(const struct llama_vocab * vocab); // sentence separator +@ctypes_function("llama_vocab_sep", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_sep(vocab: llama_vocab_p, /) -> llama_token: + """sentence separator""" + ... -# LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line -def llama_token_nl(model: llama_model_p) -> int: +# LLAMA_API llama_token llama_vocab_nl (const struct llama_vocab * vocab); // next-line +@ctypes_function("llama_vocab_nl", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_nl(vocab: llama_vocab_p, /) -> llama_token: """next-line""" - return _lib.llama_token_nl(model) + ... + + +# LLAMA_API llama_token llama_vocab_pad(const struct llama_vocab * vocab); // padding +@ctypes_function("llama_vocab_pad", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_pad(vocab: llama_vocab_p, /) -> llama_token: + """padding""" + ... + +# LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_get_add_bos", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_vocab_get_add_bos(vocab: llama_vocab_p, /) -> bool: + ... + + +# LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_get_add_eos", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_vocab_get_add_eos(vocab: llama_vocab_p, /) -> bool: + ... + + +# LLAMA_API llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_pre", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_pre(vocab: llama_vocab_p, /) -> llama_token: + ... + + +# LLAMA_API llama_token llama_vocab_fim_suf(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_suf", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_suf(vocab: llama_vocab_p, /) -> llama_token: + ... + + +# LLAMA_API llama_token llama_vocab_fim_mid(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_mid", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_mid(vocab: llama_vocab_p, /) -> llama_token: + ... + + +# LLAMA_API llama_token llama_vocab_fim_pad(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_pad", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_pad(vocab: llama_vocab_p, /) -> llama_token: + ... -_lib.llama_token_nl.argtypes = [llama_model_p] -_lib.llama_token_nl.restype = llama_token +# LLAMA_API llama_token llama_vocab_fim_rep(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_rep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_rep(vocab: llama_vocab_p, /) -> llama_token: + ... -# // Returns -1 if unknown, 1 for true or 0 for false. -# LLAMA_API int llama_add_bos_token(const struct llama_model * model); -def llama_add_bos_token(model: llama_model_p) -> int: - """Returns -1 if unknown, 1 for true or 0 for false.""" - return _lib.llama_add_bos_token(model) +# LLAMA_API llama_token llama_vocab_fim_sep(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_sep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_sep(vocab: llama_vocab_p, /) -> llama_token: + ... -_lib.llama_add_bos_token.argtypes = [llama_model_p] -_lib.llama_add_bos_token.restype = c_int +# DEPRECATED(LLAMA_API const char * llama_token_get_text(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_text instead"); +@ctypes_function( + "llama_token_get_text", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_char_p, +) +def llama_token_get_text( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bytes: + ... -# // Returns -1 if unknown, 1 for true or 0 for false. -# LLAMA_API int llama_add_eos_token(const struct llama_model * model); -def llama_add_eos_token(model: llama_model_p) -> int: - """Returns -1 if unknown, 1 for true or 0 for false.""" - return _lib.llama_add_eos_token(model) +# DEPRECATED(LLAMA_API float llama_token_get_score(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_score instead"); +@ctypes_function( + "llama_token_get_score", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_float, +) +def llama_token_get_score( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> float: + ... + +# DEPRECATED(LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_attr instead"); +@ctypes_function( + "llama_token_get_attr", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_int, +) +def llama_token_get_attr( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> int: + ... + +# DEPRECATED(LLAMA_API bool llama_token_is_eog(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_eog instead"); +@ctypes_function( + "llama_token_is_eog", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_bool, +) +def llama_token_is_eog( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bool: + ... + +# DEPRECATED(LLAMA_API bool llama_token_is_control(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_control instead"); +@ctypes_function( + "llama_token_is_control", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_bool, +) +def llama_token_is_control( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bool: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_bos(const struct llama_vocab * vocab), "use llama_vocab_bos instead"); +@ctypes_function( + "llama_token_bos", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_bos(vocab: llama_vocab_p, /) -> int: + ... -_lib.llama_add_eos_token.argtypes = [llama_model_p] -_lib.llama_add_eos_token.restype = c_int +# DEPRECATED(LLAMA_API llama_token llama_token_eos(const struct llama_vocab * vocab), "use llama_vocab_eos instead"); +@ctypes_function( + "llama_token_eos", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_eos(vocab: llama_vocab_p, /) -> int: + ... +# DEPRECATED(LLAMA_API llama_token llama_token_eot(const struct llama_vocab * vocab), "use llama_vocab_eot instead"); +@ctypes_function( + "llama_token_eot", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_eot(vocab: llama_vocab_p, /) -> int: + ... -# // codellama infill tokens -# LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix -def llama_token_prefix(model: llama_model_p) -> int: - """codellama infill tokens""" - return _lib.llama_token_prefix(model) +# DEPRECATED(LLAMA_API llama_token llama_token_cls(const struct llama_vocab * vocab), "use llama_vocab_cls instead"); +@ctypes_function( + "llama_token_cls", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_cls(vocab: llama_vocab_p, /) -> int: + ... +# DEPRECATED(LLAMA_API llama_token llama_token_sep(const struct llama_vocab * vocab), "use llama_vocab_sep instead"); +@ctypes_function( + "llama_token_sep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_sep(vocab: llama_vocab_p, /) -> int: + ... -_lib.llama_token_prefix.argtypes = [llama_model_p] -_lib.llama_token_prefix.restype = llama_token +# DEPRECATED(LLAMA_API llama_token llama_token_nl (const struct llama_vocab * vocab), "use llama_vocab_nl instead"); +@ctypes_function( + "llama_token_nl", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_nl(vocab: llama_vocab_p, /) -> int: + ... -# LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle -def llama_token_middle(model: llama_model_p) -> int: - return _lib.llama_token_middle(model) +# DEPRECATED(LLAMA_API llama_token llama_token_pad(const struct llama_vocab * vocab), "use llama_vocab_pad instead"); +@ctypes_function( + "llama_token_pad", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_pad(vocab: llama_vocab_p, /) -> int: + ... -_lib.llama_token_middle.argtypes = [llama_model_p] -_lib.llama_token_middle.restype = llama_token +# DEPRECATED(LLAMA_API bool llama_add_bos_token(const struct llama_vocab * vocab), "use llama_vocab_get_add_bos instead"); +@ctypes_function( + "llama_add_bos_token", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_add_bos_token(vocab: llama_vocab_p, /) -> bool: + ... + +# DEPRECATED(LLAMA_API bool llama_add_eos_token(const struct llama_vocab * vocab), "use llama_vocab_get_add_eos instead"); +@ctypes_function( + "llama_add_eos_token", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_add_eos_token(vocab: llama_vocab_p, /) -> bool: + ... -# LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix -def llama_token_suffix(model: llama_model_p) -> int: - return _lib.llama_token_suffix(model) +# DEPRECATED(LLAMA_API llama_token llama_token_fim_pre(const struct llama_vocab * vocab), "use llama_vocab_fim_pre instead"); +@ctypes_function( + "llama_token_fim_pre", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_pre(vocab: llama_vocab_p, /) -> llama_token: + ... -_lib.llama_token_suffix.argtypes = [llama_model_p] -_lib.llama_token_suffix.restype = llama_token +# DEPRECATED(LLAMA_API llama_token llama_token_fim_suf(const struct llama_vocab * vocab), "use llama_vocab_fim_suf instead"); +@ctypes_function( + "llama_token_fim_suf", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_suf(vocab: llama_vocab_p, /) -> llama_token: + ... +# DEPRECATED(LLAMA_API llama_token llama_token_fim_mid(const struct llama_vocab * vocab), "use llama_vocab_fim_mid instead"); +@ctypes_function( + "llama_token_fim_mid", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_mid(vocab: llama_vocab_p, /) -> llama_token: + ... -# LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle -def llama_token_eot(model: llama_model_p) -> int: - return _lib.llama_token_eot(model) +# DEPRECATED(LLAMA_API llama_token llama_token_fim_pad(const struct llama_vocab * vocab), "use llama_vocab_fim_pad instead"); +@ctypes_function( + "llama_token_fim_pad", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_pad(vocab: llama_vocab_p, /) -> llama_token: + ... +# DEPRECATED(LLAMA_API llama_token llama_token_fim_rep(const struct llama_vocab * vocab), "use llama_vocab_fim_rep instead"); +@ctypes_function( + "llama_token_fim_rep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_rep(vocab: llama_vocab_p, /) -> llama_token: + ... -_lib.llama_token_eot.argtypes = [llama_model_p] -_lib.llama_token_eot.restype = llama_token +# DEPRECATED(LLAMA_API llama_token llama_token_fim_sep(const struct llama_vocab * vocab), "use llama_vocab_fim_sep instead"); +@ctypes_function( + "llama_token_fim_sep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_sep(vocab: llama_vocab_p, /) -> llama_token: + ... + +# // CLS is equivalent to BOS +# DEPRECATED(LLAMA_API llama_token llama_vocab_cls(const struct llama_vocab * vocab), // classification +# "use llama_vocab_bos instead"); +@ctypes_function( + "llama_vocab_cls", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_cls(vocab: llama_vocab_p, /) -> llama_token: + ... # // # // Tokenization # // +# // The API is thread-safe. +# // # /// @details Convert the provided text into tokens. # /// @param tokens The tokens pointer must be large enough to hold the resulting tokens. -# /// @return Returns the number of tokens on success, no more than n_max_tokens +# /// @return Returns the number of tokens on success, no more than n_tokens_max # /// @return Returns a negative number on failure - the number of tokens that would have been returned -# /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext. -# /// Does not insert a leading space. -# LLAMA_API int llama_tokenize( -# const struct llama_model * model, +# /// @param add_special Allow to add BOS and EOS tokens if model is configured to do so. +# /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated +# /// as plaintext. Does not insert a leading space. +# LLAMA_API int32_t llama_tokenize( +# const struct llama_vocab * vocab, # const char * text, -# int text_len, +# int32_t text_len, # llama_token * tokens, -# int n_max_tokens, -# bool add_bos, -# bool special); +# int32_t n_tokens_max, +# bool add_special, +# bool parse_special); +@ctypes_function( + "llama_tokenize", + [ + llama_vocab_p_ctypes, + ctypes.c_char_p, + ctypes.c_int32, + llama_token_p, + ctypes.c_int32, + ctypes.c_bool, + ctypes.c_bool, + ], + ctypes.c_int32, +) def llama_tokenize( - model: llama_model_p, + vocab: llama_vocab_p, text: bytes, - text_len: Union[c_int, int], - tokens, # type: Array[llama_token] - n_max_tokens: Union[c_int, int], - add_bos: Union[c_bool, bool], - special: Union[c_bool, bool], + text_len: Union[ctypes.c_int, int], + tokens: CtypesArray[llama_token], + n_tokens_max: Union[ctypes.c_int, int], + add_special: Union[ctypes.c_bool, bool], + parse_special: Union[ctypes.c_bool, bool], + /, ) -> int: - """Convert the provided text into tokens.""" - return _lib.llama_tokenize( - model, text, text_len, tokens, n_max_tokens, add_bos, special - ) - - -_lib.llama_tokenize.argtypes = [ - llama_model_p, - c_char_p, - c_int, - llama_token_p, - c_int, - c_bool, - c_bool, -] -_lib.llama_tokenize.restype = c_int + """Convert the provided text into tokens. + + Args: + vocab: The vocabulary to use for tokenization. + text: The text to tokenize. + text_len: The length of the text. + tokens: The tokens pointer must be large enough to hold the resulting tokens. + n_max_tokens: The maximum number of tokens to return. + add_special: Allow adding special tokenns if the model is configured to do so. + parse_special: Allow parsing special tokens. + + Returns: + Returns the number of tokens on success, no more than n_tokens_max + Returns a negative number on failure - the number of tokens that would have been returned + """ + ... # // Token Id -> Piece. # // Uses the vocabulary in the provided context. # // Does not write null terminator to the buffer. -# // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. -# LLAMA_API int llama_token_to_piece( -# const struct llama_model * model, +# // User can skip up to 'lstrip' leading spaces before copying (useful when encoding/decoding multiple tokens with 'add_space_prefix') +# // @param special If true, special tokens are rendered in the output. +# LLAMA_API int32_t llama_token_to_piece( +# const struct llama_vocab * vocab, # llama_token token, # char * buf, -# int length); +# int32_t length, +# int32_t lstrip, +# bool special); +@ctypes_function( + "llama_token_to_piece", + [ + llama_vocab_p_ctypes, + llama_token, + ctypes.c_char_p, + ctypes.c_int32, + ctypes.c_int32, + ctypes.c_bool, + ], + ctypes.c_int32, +) def llama_token_to_piece( - model: llama_model_p, + vocab: llama_vocab_p, token: Union[llama_token, int], - buf: Union[c_char_p, bytes], - length: Union[c_int, int], + buf: Union[ctypes.c_char_p, bytes, CtypesArray[ctypes.c_char]], + length: Union[ctypes.c_int, int], + lstrip: Union[ctypes.c_int, int], + special: Union[ctypes.c_bool, bool], + /, ) -> int: """Token Id -> Piece. Uses the vocabulary in the provided context. Does not write null terminator to the buffer. User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. - """ - return _lib.llama_token_to_piece(model, token, buf, length) + Args: + vocab: The vocabulary to use for tokenization. + token: The token to convert. + buf: The buffer to write the token to. + length: The length of the buffer. + lstrip: The number of leading spaces to skip. + special: If true, special tokens are rendered in the output.""" + ... + + +# # // check if token0 is contained as a prefix in token1 +# # LLAMA_API bool llama_token_is_prefix( +# # const struct llama_model * model, +# # llama_token token0, +# # llama_token token1); +# @ctypes_function( +# "llama_token_is_prefix", +# [llama_model_p_ctypes, llama_token, llama_token], +# ctypes.c_bool, +# ) +# def llama_token_is_prefix( +# model: llama_model_p, token0: Union[llama_token, int], token1: Union[llama_token, int], / +# ) -> bool: +# """Check if token0 is contained as a prefix in token1""" +# ... + + +# /// @details Convert the provided tokens into text (inverse of llama_tokenize()). +# /// @param text The char pointer must be large enough to hold the resulting text. +# /// @return Returns the number of chars/bytes on success, no more than text_len_max. +# /// @return Returns a negative number on failure - the number of chars/bytes that would have been returned. +# /// @param remove_special Allow to remove BOS and EOS tokens if model is configured to do so. +# /// @param unparse_special If true, special tokens are rendered in the output. +# LLAMA_API int32_t llama_detokenize( +# const struct llama_model * model, +# const llama_token * tokens, +# int32_t n_tokens, +# char * text, +# int32_t text_len_max, +# bool remove_special, +# bool unparse_special); +@ctypes_function( + "llama_detokenize", + [ + llama_model_p_ctypes, + ctypes.POINTER(llama_token), + ctypes.c_int32, + ctypes.c_char_p, + ctypes.c_int32, + ctypes.c_bool, + ctypes.c_bool, + ], + ctypes.c_int32, +) +def llama_detokenize( + model: llama_model_p, + tokens: CtypesArray[llama_token], + n_tokens: Union[ctypes.c_int, int], + text: bytes, + text_len_max: Union[ctypes.c_int, int], + remove_special: Union[ctypes.c_bool, bool], + unparse_special: Union[ctypes.c_bool, bool], + /, +) -> int: + """Convert the provided tokens into text (inverse of llama_tokenize()). -_lib.llama_token_to_piece.argtypes = [llama_model_p, llama_token, c_char_p, c_int] -_lib.llama_token_to_piece.restype = c_int + Args: + model: The model to use for tokenization. + tokens: The tokens to convert. + n_tokens: The number of tokens. + text: The buffer to write the text to. + text_len_max: The length of the buffer. + remove_special: Allow to remove BOS and EOS tokens if model is configured to do so. + unparse_special: If true, special tokens are rendered in the output.""" + ... # // -# // Grammar +# // Chat templates # // -# LLAMA_API struct llama_grammar * llama_grammar_init( -# const llama_grammar_element ** rules, -# size_t n_rules, -# size_t start_rule_index); -def llama_grammar_init( - rules, # type: Array[llama_grammar_element_p] # type: ignore - n_rules: Union[c_size_t, int], - start_rule_index: Union[c_size_t, int], -) -> llama_grammar_p: - """Initialize a grammar from a set of rules.""" - return _lib.llama_grammar_init(rules, n_rules, start_rule_index) +# /// Apply chat template. Inspired by hf apply_chat_template() on python. +# /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" +# /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template +# /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead. +# /// @param chat Pointer to a list of multiple llama_chat_message +# /// @param n_msg Number of llama_chat_message in this chat +# /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message. +# /// @param buf A buffer to hold the output formatted prompt. The recommended alloc size is 2 * (total number of characters of all messages) +# /// @param length The size of the allocated buffer +# /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template. +# LLAMA_API int32_t llama_chat_apply_template( +# const char * tmpl, +# const struct llama_chat_message * chat, +# size_t n_msg, +# bool add_ass, +# char * buf, +# int32_t length); +@ctypes_function( + "llama_chat_apply_template", + [ + ctypes.c_char_p, # tmpl + ctypes.POINTER(llama_chat_message), # chat + ctypes.c_size_t, # n_msg + ctypes.c_bool, # add_ass (added) + ctypes.c_char_p, # buf + ctypes.c_int32, # length + ], + ctypes.c_int32, +) +def llama_chat_apply_template( + tmpl: bytes, + chat: CtypesArray[llama_chat_message], + n_msg: int, + add_ass: bool, # Added parameter + buf: bytes, + length: int, + /, +) -> int: + """Apply chat template. + + Args: + tmpl: Template to use. If None, uses model's default + chat: Array of chat messages + n_msg: Number of messages + add_ass: Whether to end prompt with assistant token + buf: Output buffer + length: Buffer length + + Returns: + Number of bytes written, or needed if buffer too small + """ + ... + + +# // Get list of built-in chat templates +# LLAMA_API int32_t llama_chat_builtin_templates(const char ** output, size_t len); +@ctypes_function( + "llama_chat_builtin_templates", + [ + ctypes.POINTER(ctypes.c_char_p), + ctypes.c_size_t, + ], + ctypes.c_int32, +) +def llama_chat_builtin_templates( + output: CtypesArray[bytes], + len: Union[ctypes.c_size_t, int], + /, +) -> int: + """Get list of built-in chat templates. + Args: + output: Output buffer to store template names. + len: Length of the output buffer. -_lib.llama_grammar_init.argtypes = [ - POINTER(llama_grammar_element_p), - c_size_t, - c_size_t, -] -_lib.llama_grammar_init.restype = llama_grammar_p + Returns: + Number of templates available. + Returns a negative number on error. + """ + ... -# LLAMA_API void llama_grammar_free(struct llama_grammar * grammar); -def llama_grammar_free(grammar: llama_grammar_p): - """Free a grammar.""" - return _lib.llama_grammar_free(grammar) +# // +# // Sampling API +# // +# // Sample usage: +# // +# // // prepare the sampling chain at the start +# // auto sparams = llama_sampler_chain_default_params(); +# // +# // llama_sampler * smpl = llama_sampler_chain_init(sparams); +# // +# // llama_sampler_chain_add(smpl, llama_sampler_init_top_k(50)); +# // llama_sampler_chain_add(smpl, llama_sampler_init_top_p(0.9, 1)); +# // llama_sampler_chain_add(smpl, llama_sampler_init_temp (0.8)); +# // +# // // typically, the chain should end with a sampler such as "greedy", "dist" or "mirostat" +# // // this sampler will be responsible to select the actual token +# // llama_sampler_chain_add(smpl, llama_sampler_init_dist(seed)); +# // +# // ... +# // +# // // decoding loop: +# // while (...) { +# // ... +# // +# // llama_decode(ctx, batch); +# // +# // // sample from the logits of the last token in the batch +# // const llama_token id = llama_sampler_sample(smpl, ctx, -1); +# // +# // // accepting the token updates the internal state of certain samplers (e.g. grammar, repetition, etc.) +# // llama_sampler_accept(smpl, id); +# // ... +# // } +# // +# // llama_sampler_free(smpl); +# // +# // TODO: In the future, llama_sampler will be utilized to offload the sampling to the backends (e.g. GPU). +# // +# typedef void * llama_sampler_context_t; +llama_sampler_context_t = ctypes.c_void_p + + +# // user code can implement the interface below in order to create custom llama_sampler +# struct llama_sampler_i { +# const char * (*name) (const struct llama_sampler * smpl); // can be NULL +# void (*accept)( struct llama_sampler * smpl, llama_token token); // can be NULL +# void (*apply) ( struct llama_sampler * smpl, llama_token_data_array * cur_p); // required +# void (*reset) ( struct llama_sampler * smpl); // can be NULL +# struct llama_sampler * (*clone) (const struct llama_sampler * smpl); // can be NULL if ctx is NULL +# void (*free) ( struct llama_sampler * smpl); // can be NULL if ctx is NULL +# +# // TODO: API for internal libllama usage for appending the sampling to an existing ggml_cgraph +# //void (*apply_ggml) (struct llama_sampler * smpl, ...); +# }; +class llama_sampler_i(ctypes.Structure): + ... -_lib.llama_grammar_free.argtypes = [llama_grammar_p] -_lib.llama_grammar_free.restype = None +# struct llama_sampler { +# const struct llama_sampler_i * iface; +# llama_sampler_context_t ctx; +# }; +class llama_sampler(ctypes.Structure): + _fields_ = [ + ("iface", ctypes.POINTER(llama_sampler_i)), + ("ctx", llama_sampler_context_t), + ] -# LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar); -def llama_grammar_copy(grammar: llama_grammar_p) -> llama_grammar_p: - """Copy a grammar.""" - return _lib.llama_grammar_copy(grammar) +if TYPE_CHECKING: + llama_sampler_p = CtypesPointer[llama_sampler] -_lib.llama_grammar_copy.argtypes = [llama_grammar_p] -_lib.llama_grammar_copy.restype = llama_grammar_p +llama_sampler_p_ctypes = ctypes.POINTER(llama_sampler) -# // -# // Sampling functions -# // +llama_sampler_i_name = ctypes.CFUNCTYPE(ctypes.c_char_p, llama_sampler_p_ctypes) +llama_sampler_i_accept = ctypes.CFUNCTYPE(None, llama_sampler_p_ctypes, llama_token) +llama_sampler_i_apply = ctypes.CFUNCTYPE( + None, llama_sampler_p_ctypes, llama_token_data_array_p +) +llama_sampler_i_reset = ctypes.CFUNCTYPE(None, llama_sampler_p_ctypes) +llama_sampler_i_clone = ctypes.CFUNCTYPE(llama_sampler_p_ctypes, llama_sampler_p_ctypes) +llama_sampler_i_free = ctypes.CFUNCTYPE(None, llama_sampler_p_ctypes) + +llama_sampler_i._fields_ = [ + ("name", llama_sampler_i_name), + ("accept", llama_sampler_i_accept), + ("apply", llama_sampler_i_apply), + ("reset", llama_sampler_i_reset), + ("clone", llama_sampler_i_clone), + ("free", llama_sampler_i_free), +] -# // Sets the current rng seed. -# LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed); -def llama_set_rng_seed(ctx: llama_context_p, seed: Union[c_uint32, int]): - """Sets the current rng seed.""" - return _lib.llama_set_rng_seed(ctx, seed) +# // mirror of llama_sampler_i: +# LLAMA_API struct llama_sampler * llama_sampler_init (const struct llama_sampler_i * iface, llama_sampler_context_t ctx); +@ctypes_function( + "llama_sampler_init", + [ctypes.POINTER(llama_sampler_i), llama_sampler_context_t], + llama_sampler_p_ctypes, +) +def llama_sampler_init( + iface: ctypes.POINTER(llama_sampler_i), ctx: llama_sampler_context_t, / +) -> llama_sampler_p: + ... -_lib.llama_set_rng_seed.argtypes = [llama_context_p, c_uint32] -_lib.llama_set_rng_seed.restype = None +# LLAMA_API const char * llama_sampler_name (const struct llama_sampler * smpl); +@ctypes_function( + "llama_sampler_name", + [llama_sampler_p_ctypes], + ctypes.c_char_p, +) +def llama_sampler_name(smpl: llama_sampler_p, /) -> bytes: + ... -# /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. -# /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. -# LLAMA_API void llama_sample_repetition_penalties( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# const llama_token * last_tokens, -# size_t penalty_last_n, -# float penalty_repeat, -# float penalty_freq, -# float penalty_present); -def llama_sample_repetition_penalties( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - last_tokens_data, # type: Array[llama_token] - penalty_last_n: Union[c_size_t, int], - penalty_repeat: Union[c_float, float], - penalty_freq: Union[c_float, float], - penalty_present: Union[c_float, float], -): - """Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. - Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. - """ - return _lib.llama_sample_repetition_penalties( - ctx, - candidates, - last_tokens_data, - penalty_last_n, - penalty_repeat, - penalty_freq, - penalty_present, - ) +# LLAMA_API void llama_sampler_accept( struct llama_sampler * smpl, llama_token token); +@ctypes_function( + "llama_sampler_accept", + [llama_sampler_p_ctypes, llama_token], + None, +) +def llama_sampler_accept(smpl: llama_sampler_p, token: Union[llama_token, int], /): + ... -_lib.llama_sample_repetition_penalties.argtypes = [ - llama_context_p, - llama_token_data_array_p, - llama_token_p, - c_size_t, - c_float, - c_float, - c_float, -] -_lib.llama_sample_repetition_penalties.restype = None - - -# /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806 -# /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted. -# /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context. -# /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance. -# LLAMA_API void llama_sample_classifier_free_guidance( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# struct llama_context * guidance_ctx, -# float scale); -def llama_sample_classifier_free_guidance( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - guidance_ctx: llama_context_p, - scale: Union[c_float, float], +# LLAMA_API void llama_sampler_apply ( struct llama_sampler * smpl, llama_token_data_array * cur_p); +@ctypes_function( + "llama_sampler_apply", + [llama_sampler_p_ctypes, llama_token_data_array_p], + None, +) +def llama_sampler_apply( + smpl: llama_sampler_p, cur_p: CtypesArray[llama_token_data_array], / ): - """Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806""" - return _lib.llama_sample_classifier_free_guidance( - ctx, candidates, guidance_ctx, scale - ) - - -_lib.llama_sample_classifier_free_guidance.argtypes = [ - llama_context_p, - llama_token_data_array_p, - llama_context_p, - c_float, -] -_lib.llama_sample_classifier_free_guidance.restype = None + ... -# /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. -# LLAMA_API void llama_sample_softmax( -# struct llama_context * ctx, -# llama_token_data_array * candidates); -def llama_sample_softmax( - ctx: llama_context_p, candidates # type: _Pointer[llama_token_data] -): - """Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.""" - return _lib.llama_sample_softmax(ctx, candidates) +# LLAMA_API void llama_sampler_reset ( struct llama_sampler * smpl); +@ctypes_function( + "llama_sampler_reset", + [llama_sampler_p_ctypes], + None, +) +def llama_sampler_reset(smpl: llama_sampler_p, /): + ... -_lib.llama_sample_softmax.argtypes = [ - llama_context_p, - llama_token_data_array_p, -] -_lib.llama_sample_softmax.restype = None +# LLAMA_API struct llama_sampler * llama_sampler_clone (const struct llama_sampler * smpl); +@ctypes_function( + "llama_sampler_clone", + [llama_sampler_p_ctypes], + llama_sampler_p_ctypes, +) +def llama_sampler_clone(smpl: llama_sampler_p, /) -> llama_sampler_p: + ... -# /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 -# LLAMA_API void llama_sample_top_k( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# int k, -# size_t min_keep); -def llama_sample_top_k( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - k: Union[c_int, int], - min_keep: Union[c_size_t, int], -): - """Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751""" - return _lib.llama_sample_top_k(ctx, candidates, k, min_keep) +# // important: do not free if the sampler has been added to a llama_sampler_chain (via llama_sampler_chain_add) +# LLAMA_API void llama_sampler_free ( struct llama_sampler * smpl); +@ctypes_function( + "llama_sampler_free", + [llama_sampler_p_ctypes], + None, +) +def llama_sampler_free(smpl: llama_sampler_p, /): + ... + + +# // llama_sampler_chain +# // a type of llama_sampler that can chain multiple samplers one after another +# +# LLAMA_API struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_params params); +@ctypes_function( + "llama_sampler_chain_init", + [llama_sampler_chain_params], + llama_sampler_p_ctypes, +) +def llama_sampler_chain_init(params: llama_sampler_chain_params, /) -> llama_sampler_p: + ... -_lib.llama_sample_top_k.argtypes = [ - llama_context_p, - llama_token_data_array_p, - c_int, - c_size_t, -] -_lib.llama_sample_top_k.restype = None +# // important: takes ownership of the sampler object and will free it when llama_sampler_free is called +# LLAMA_API void llama_sampler_chain_add( struct llama_sampler * chain, struct llama_sampler * smpl); +@ctypes_function( + "llama_sampler_chain_add", + [llama_sampler_p_ctypes, llama_sampler_p_ctypes], + None, +) +def llama_sampler_chain_add(chain: llama_sampler_p, smpl: llama_sampler_p, /): + ... -# /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 -# LLAMA_API void llama_sample_top_p( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float p, -# size_t min_keep); -def llama_sample_top_p( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - p: Union[c_float, float], - min_keep: Union[c_size_t, int], -): - """Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751""" - return _lib.llama_sample_top_p(ctx, candidates, p, min_keep) +# LLAMA_API struct llama_sampler * llama_sampler_chain_get(const struct llama_sampler * chain, int32_t i); +@ctypes_function( + "llama_sampler_chain_get", + [llama_sampler_p_ctypes, ctypes.c_int32], + llama_sampler_p_ctypes, +) +def llama_sampler_chain_get( + chain: llama_sampler_p, i: Union[ctypes.c_int32, int], / +) -> llama_sampler_p: + ... -_lib.llama_sample_top_p.argtypes = [ - llama_context_p, - llama_token_data_array_p, - c_float, - c_size_t, -] -_lib.llama_sample_top_p.restype = None +# LLAMA_API int llama_sampler_chain_n (const struct llama_sampler * chain); +@ctypes_function( + "llama_sampler_chain_n", + [llama_sampler_p_ctypes], + ctypes.c_int, +) +def llama_sampler_chain_n(chain: llama_sampler_p, /) -> int: + ... -# /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 -# LLAMA_API void llama_sample_min_p( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float p, -# size_t min_keep); -def llama_sample_min_p( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - p: Union[c_float, float], - min_keep: Union[c_size_t, int], -): - """Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841""" - return _lib.llama_sample_min_p(ctx, candidates, p, min_keep) +# // after removing a sampler, the chain will no longer own it, and it will not be freed when the chain is freed +# LLAMA_API struct llama_sampler * llama_sampler_chain_remove( struct llama_sampler * chain, int32_t i); +@ctypes_function( + "llama_sampler_chain_remove", + [llama_sampler_p_ctypes, ctypes.c_int32], + llama_sampler_p_ctypes, +) +def llama_sampler_chain_remove( + chain: llama_sampler_p, i: Union[ctypes.c_int32, int], / +) -> llama_sampler_p: + ... -_lib.llama_sample_min_p.argtypes = [ - llama_context_p, - llama_token_data_array_p, - c_float, - c_size_t, -] -_lib.llama_sample_min_p.restype = None +# // available samplers: +# +# LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void); +@ctypes_function("llama_sampler_init_greedy", [], llama_sampler_p_ctypes) +def llama_sampler_init_greedy() -> llama_sampler_p: + ... -# /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. -# LLAMA_API void llama_sample_tail_free( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float z, -# size_t min_keep); -def llama_sample_tail_free( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - z: Union[c_float, float], - min_keep: Union[c_size_t, int], -): - """Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.""" - return _lib.llama_sample_tail_free(ctx, candidates, z, min_keep) +# LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed); +@ctypes_function("llama_sampler_init_dist", [ctypes.c_uint32], llama_sampler_p_ctypes) +def llama_sampler_init_dist(seed: int) -> llama_sampler_p: + ... -_lib.llama_sample_tail_free.argtypes = [ - llama_context_p, - llama_token_data_array_p, - c_float, - c_size_t, -] -_lib.llama_sample_tail_free.restype = None +# /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. +# /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first. +# DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void), +# "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)"); +@ctypes_function("llama_sampler_init_softmax", [], llama_sampler_p_ctypes) +def llama_sampler_init_softmax() -> llama_sampler_p: + ... -# /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. -# LLAMA_API void llama_sample_typical( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float p, -# size_t min_keep); -def llama_sample_typical( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - p: Union[c_float, float], - min_keep: Union[c_size_t, int], -): - """Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.""" - return _lib.llama_sample_typical(ctx, candidates, p, min_keep) +# /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 +# /// Setting k <= 0 makes this a noop +# LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k); +@ctypes_function("llama_sampler_init_top_k", [ctypes.c_int32], llama_sampler_p_ctypes) +def llama_sampler_init_top_k(k: int) -> llama_sampler_p: + ... -_lib.llama_sample_typical.argtypes = [ - llama_context_p, - llama_token_data_array_p, - c_float, - c_size_t, -] -_lib.llama_sample_typical.restype = None +# /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 +# LLAMA_API struct llama_sampler * llama_sampler_init_top_p (float p, size_t min_keep); +@ctypes_function( + "llama_sampler_init_top_p", + [ctypes.c_float, ctypes.c_size_t], + llama_sampler_p_ctypes, +) +def llama_sampler_init_top_p(p: float, min_keep: int) -> llama_sampler_p: + ... -# LLAMA_API void llama_sample_temp( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float temp); -def llama_sample_temp( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - temp: Union[c_float, float], -): - """Temperature sampling described in academic paper "Generating Long Sequences with Sparse Transformers" https://arxiv.org/abs/1904.10509 - - Parameters: - candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - temp: The temperature value to use for the sampling. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.""" - return _lib.llama_sample_temp(ctx, candidates, temp) - - -_lib.llama_sample_temp.argtypes = [ - llama_context_p, - llama_token_data_array_p, - c_float, -] -_lib.llama_sample_temp.restype = None +# /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 +# LLAMA_API struct llama_sampler * llama_sampler_init_min_p (float p, size_t min_keep); +@ctypes_function( + "llama_sampler_init_min_p", + [ctypes.c_float, ctypes.c_size_t], + llama_sampler_p_ctypes, +) +def llama_sampler_init_min_p(p: float, min_keep: int) -> llama_sampler_p: + ... -# LLAMA_API DEPRECATED(void llama_sample_temperature( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float temp), -# "use llama_sample_temp instead"); -def llama_sample_temperature( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - temp: Union[c_float, float], -): - """use llama_sample_temp instead""" - return _lib.llama_sample_temperature(ctx, candidates, temp) +# /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. +# LLAMA_API struct llama_sampler * llama_sampler_init_typical (float p, size_t min_keep); +@ctypes_function( + "llama_sampler_init_typical", + [ctypes.c_float, ctypes.c_size_t], + llama_sampler_p_ctypes, +) +def llama_sampler_init_typical(p: float, min_keep: int) -> llama_sampler_p: + ... -_lib.llama_sample_temperature.argtypes = [ - llama_context_p, - llama_token_data_array_p, - c_float, -] -_lib.llama_sample_temperature.restype = None +# LLAMA_API struct llama_sampler * llama_sampler_init_temp (float t); +@ctypes_function("llama_sampler_init_temp", [ctypes.c_float], llama_sampler_p_ctypes) +def llama_sampler_init_temp(t: float) -> llama_sampler_p: + ... -# /// @details Apply constraints from grammar -# LLAMA_API void llama_sample_grammar( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# const struct llama_grammar * grammar); -def llama_sample_grammar( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - grammar, # type: llama_grammar_p -): - """Apply constraints from grammar - - Parameters: - candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - grammar: A grammar object containing the rules and constraints to apply to the generated text.""" - return _lib.llama_sample_grammar(ctx, candidates, grammar) - - -_lib.llama_sample_grammar.argtypes = [ - llama_context_p, - llama_token_data_array_p, - llama_grammar_p, -] -_lib.llama_sample_grammar.restype = None +# /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772. +# LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext (float t, float delta, float exponent); +@ctypes_function( + "llama_sampler_init_temp_ext", + [ctypes.c_float, ctypes.c_float, ctypes.c_float], + llama_sampler_p_ctypes, +) +def llama_sampler_init_temp_ext( + t: float, delta: float, exponent: float +) -> llama_sampler_p: + ... + + +# /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335 +# LLAMA_API struct llama_sampler * llama_sampler_init_xtc (float p, float t, size_t min_keep, uint32_t seed); +@ctypes_function( + "llama_sampler_init_xtc", + [ctypes.c_float, ctypes.c_float, ctypes.c_size_t, ctypes.c_uint32], + llama_sampler_p_ctypes, +) +def llama_sampler_init_xtc( + p: float, t: float, min_keep: int, seed: int, / +) -> llama_sampler_p: + ... + + +# /// @details Top n sigma sampling as described in academic paper "Top-nσ: Not All Logits Are You Need" https://arxiv.org/pdf/2411.07641 +# LLAMA_API struct llama_sampler * llama_sampler_init_top_n_sigma(float n); +@ctypes_function( + "llama_sampler_init_top_n_sigma", + [ctypes.c_float], + llama_sampler_p_ctypes, +) +def llama_sampler_init_top_n_sigma(n: float, /) -> llama_sampler_p: + ... # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -2111,41 +3909,21 @@ def llama_sample_grammar( # /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. # /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. # /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. -# LLAMA_API llama_token llama_sample_token_mirostat( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float tau, -# float eta, -# int m, -# float * mu); -def llama_sample_token_mirostat( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - tau: Union[c_float, float], - eta: Union[c_float, float], - m: Union[c_int, int], - mu, # type: _Pointer[c_float] -) -> int: - """Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. - - Parameters: - candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. - eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. - m: The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. - mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.""" - return _lib.llama_sample_token_mirostat(ctx, candidates, tau, eta, m, mu) - - -_lib.llama_sample_token_mirostat.argtypes = [ - llama_context_p, - llama_token_data_array_p, - c_float, - c_float, - c_int, - c_float_p, -] -_lib.llama_sample_token_mirostat.restype = llama_token +# LLAMA_API struct llama_sampler * llama_sampler_init_mirostat( +# int32_t n_vocab, +# uint32_t seed, +# float tau, +# float eta, +# int32_t m); +@ctypes_function( + "llama_sampler_init_mirostat", + [ctypes.c_int32, ctypes.c_uint32, ctypes.c_float, ctypes.c_float, ctypes.c_int32], + llama_sampler_p_ctypes, +) +def llama_sampler_init_mirostat( + n_vocab: int, seed: int, tau: float, eta: float, m: int, / +) -> llama_sampler_p: + ... # /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. @@ -2153,251 +3931,385 @@ def llama_sample_token_mirostat( # /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. # /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. # /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. -# LLAMA_API llama_token llama_sample_token_mirostat_v2( -# struct llama_context * ctx, -# llama_token_data_array * candidates, -# float tau, -# float eta, -# float * mu); -def llama_sample_token_mirostat_v2( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] - tau: Union[c_float, float], - eta: Union[c_float, float], - mu, # type: _Pointer[c_float] -) -> int: - """Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. - - Parameters: - candidates: A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. - tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. - eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. - mu: Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.""" - return _lib.llama_sample_token_mirostat_v2(ctx, candidates, tau, eta, mu) - - -_lib.llama_sample_token_mirostat_v2.argtypes = [ - llama_context_p, - llama_token_data_array_p, - c_float, - c_float, - c_float_p, -] -_lib.llama_sample_token_mirostat_v2.restype = llama_token +# LLAMA_API struct llama_sampler * llama_sampler_init_mirostat_v2( +# uint32_t seed, +# float tau, +# float eta); +@ctypes_function( + "llama_sampler_init_mirostat_v2", + [ctypes.c_uint32, ctypes.c_float, ctypes.c_float], + llama_sampler_p_ctypes, +) +def llama_sampler_init_mirostat_v2( + seed: int, tau: float, eta: float, / +) -> llama_sampler_p: + ... + + +# /// @details Intializes a GBNF grammar, see grammars/README.md for details. +# /// @param vocab The vocabulary that this grammar will be used with. +# /// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails. +# /// @param grammar_root The name of the start symbol for the grammar. +# LLAMA_API struct llama_sampler * llama_sampler_init_grammar( +# const struct llama_vocab * vocab, +# const char * grammar_str, +# const char * grammar_root); +@ctypes_function( + "llama_sampler_init_grammar", + [llama_vocab_p_ctypes, ctypes.c_char_p, ctypes.c_char_p], + llama_sampler_p_ctypes, +) +def llama_sampler_init_grammar( + vocab: llama_vocab_p, grammar_str: bytes, grammar_root: bytes, / +) -> llama_sampler_p: + ... + + +# /// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639 +# /// @param trigger_patterns A list of patterns that will trigger the grammar sampler. Pattern will be matched from the start of the generation output, and grammar sampler will be fed content starting from its first match group. +# /// @param trigger_tokens A list of tokens that will trigger the grammar sampler. Grammar sampler will be fed content starting from the trigger token included. +# LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy_patterns( +# const struct llama_vocab * vocab, +# const char * grammar_str, +# const char * grammar_root, +# const char ** trigger_patterns, +# size_t num_trigger_patterns, +# const llama_token * trigger_tokens, +# size_t num_trigger_tokens); +@ctypes_function( + "llama_sampler_init_grammar_lazy_patterns", + [ + llama_vocab_p_ctypes, + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.POINTER(ctypes.c_char_p), + ctypes.c_size_t, + ctypes.POINTER(llama_token), + ctypes.c_size_t, + ], + llama_sampler_p_ctypes, +) +def llama_sampler_init_grammar_lazy_patterns( + vocab: llama_vocab_p, + grammar_str: bytes, + grammar_root: bytes, + trigger_patterns: CtypesArray[bytes], + num_trigger_patterns: int, + trigger_tokens: CtypesArray[llama_token], + num_trigger_tokens: int, + /, +) -> llama_sampler_p: + ... + + +# /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first. +# LLAMA_API struct llama_sampler * llama_sampler_init_penalties( +# int32_t penalty_last_n, // last n tokens to penalize (0 = disable penalty, -1 = context size) +# float penalty_repeat, // 1.0 = disabled +# float penalty_freq, // 0.0 = disabled +# float penalty_present); // 0.0 = disabled +@ctypes_function( + "llama_sampler_init_penalties", + [ctypes.c_int32, ctypes.c_float, ctypes.c_float, ctypes.c_float], + llama_sampler_p_ctypes, +) +def llama_sampler_init_penalties( + penalty_last_n: int, + penalty_repeat: float, + penalty_freq: float, + penalty_present: float, + /, +) -> llama_sampler_p: + ... + + +# /// @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982 +# LLAMA_API struct llama_sampler * llama_sampler_init_dry( +# const struct llama_vocab * vocab, +# int32_t n_ctx_train, +# float dry_multiplier, +# float dry_base, +# int32_t dry_allowed_length, +# int32_t dry_penalty_last_n, +# const char ** seq_breakers, +# size_t num_breakers); +@ctypes_function( + "llama_sampler_init_dry", + [ + llama_vocab_p_ctypes, + ctypes.c_int32, + ctypes.c_float, + ctypes.c_float, + ctypes.c_int32, + ctypes.c_int32, + ctypes.POINTER(ctypes.c_char_p), + ctypes.c_size_t, + ], + llama_sampler_p_ctypes, +) +def llama_sampler_init_dry( + vocab: llama_vocab_p, + n_ctx_train: int, + dry_multiplier: float, + dry_base: float, + dry_allowed_length: int, + dry_penalty_last_n: int, + seq_breakers, + num_breakers: int, + /, +) -> llama_sampler_p: + ... + + +# LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias( +# int32_t n_vocab, +# int32_t n_logit_bias, +# const llama_logit_bias * logit_bias); +@ctypes_function( + "llama_sampler_init_logit_bias", + [ctypes.c_int32, ctypes.c_int32, llama_logit_bias_p], + llama_sampler_p_ctypes, +) +def llama_sampler_init_logit_bias( + n_vocab: int, n_logit_bias: int, logit_bias: CtypesArray[llama_logit_bias], / +) -> llama_sampler_p: + ... -# /// @details Selects the token with the highest probability. -# /// Does not compute the token probabilities. Use llama_sample_softmax() instead. -# LLAMA_API llama_token llama_sample_token_greedy( -# struct llama_context * ctx, -# llama_token_data_array * candidates); -def llama_sample_token_greedy( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] -) -> int: - """Selects the token with the highest probability.""" - return _lib.llama_sample_token_greedy(ctx, candidates) +# // this sampler is meant to be used for fill-in-the-middle infilling +# // it's supposed to be used after top_k + top_p sampling +# // +# // 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG +# // 2. combine probs of tokens that have the same prefix +# // +# // example: +# // +# // - before: +# // "hel": 0.5 +# // "hell": 0.2 +# // "hello": 0.1 +# // "dummy": 0.1 +# // +# // - after: +# // "hel": 0.8 +# // "dummy": 0.1 +# // +# // 3. discard non-EOG tokens with low prob +# // 4. if no tokens are left -> pick EOT +# // +# LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_vocab * vocab); +@ctypes_function( + "llama_sampler_init_infill", + [llama_vocab_p_ctypes], + llama_sampler_p_ctypes, +) +def llama_sampler_init_infill(vocab: llama_vocab_p, /) -> llama_sampler_p: + ... -_lib.llama_sample_token_greedy.argtypes = [ - llama_context_p, - llama_token_data_array_p, -] -_lib.llama_sample_token_greedy.restype = llama_token +# // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise +# LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl); +@ctypes_function( + "llama_sampler_get_seed", + [llama_sampler_p_ctypes], + ctypes.c_uint32, +) +def llama_sampler_get_seed(smpl: llama_sampler_p, /) -> int: + ... -# /// @details Randomly selects a token from the candidates based on their probabilities. -# LLAMA_API llama_token llama_sample_token( -# struct llama_context * ctx, -# llama_token_data_array * candidates); -def llama_sample_token( - ctx: llama_context_p, - candidates, # type: _Pointer[llama_token_data_array] +# /// @details Sample and accept a token from the idx-th output of the last evaluation +# // +# // Shorthand for: +# // const auto * logits = llama_get_logits_ith(ctx, idx); +# // llama_token_data_array cur_p = { ... init from logits ... }; +# // llama_sampler_apply(smpl, &cur_p); +# // auto token = cur_p.data[cur_p.selected].id; +# // llama_sampler_accept(smpl, token); +# // return token; +# // Returns the sampled token +# LLAMA_API llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_context * ctx, int32_t idx); +@ctypes_function( + "llama_sampler_sample", + [llama_sampler_p_ctypes, llama_context_p_ctypes, ctypes.c_int32], + llama_token, +) +def llama_sampler_sample( + smpl: llama_sampler_p, ctx: llama_context_p, idx: int, / ) -> int: - """Randomly selects a token from the candidates based on their probabilities.""" - return _lib.llama_sample_token(ctx, candidates) + ... -_lib.llama_sample_token.argtypes = [ - llama_context_p, - llama_token_data_array_p, -] -_lib.llama_sample_token.restype = llama_token +# // +# // Model split +# // -# /// @details Accepts the sampled token into the grammar -# LLAMA_API void llama_grammar_accept_token( -# struct llama_context * ctx, -# struct llama_grammar * grammar, -# llama_token token); -def llama_grammar_accept_token( - ctx: llama_context_p, - grammar: llama_grammar_p, - token: Union[llama_token, int], -) -> None: - """Accepts the sampled token into the grammar""" - _lib.llama_grammar_accept_token(ctx, grammar, token) +# /// @details Build a split GGUF final path for this chunk. +# /// llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf" +# // Returns the split_path length. +# LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count); +@ctypes_function( + "llama_split_path", + [ctypes.c_char_p, ctypes.c_size_t, ctypes.c_char_p, ctypes.c_int, ctypes.c_int], + ctypes.c_int, +) +def llama_split_path( + split_path: bytes, + maxlen: Union[ctypes.c_size_t, int], + path_prefix: bytes, + split_no: Union[ctypes.c_int, int], + split_count: Union[ctypes.c_int, int], + /, +) -> int: + """Build a split GGUF final path for this chunk.""" + ... + + +# /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match. +# /// llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0" +# // Returns the split_prefix length. +# LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count); +@ctypes_function( + "llama_split_prefix", + [ctypes.c_char_p, ctypes.c_size_t, ctypes.c_char_p, ctypes.c_int, ctypes.c_int], + ctypes.c_int, +) +def llama_split_prefix( + split_prefix: bytes, + maxlen: Union[ctypes.c_size_t, int], + split_path: bytes, + split_no: Union[ctypes.c_int, int], + split_count: Union[ctypes.c_int, int], + /, +) -> int: + """Extract the path prefix from the split_path if and only if the split_no and split_count match.""" + ... -_lib.llama_grammar_accept_token.argtypes = [ - llama_context_p, - llama_grammar_p, - llama_token, -] -_lib.llama_grammar_accept_token.restype = None +# // Print system information +# LLAMA_API const char * llama_print_system_info(void); +@ctypes_function("llama_print_system_info", [], ctypes.c_char_p) +def llama_print_system_info() -> bytes: + ... + + +# // Set callback for all future logging events. +# // If this is not called, or NULL is supplied, everything is output on stderr. +# LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); +@ctypes_function( + "llama_log_set", + [ctypes.c_void_p, ctypes.c_void_p], + None, +) +def llama_log_set( + log_callback: Optional[CtypesFuncPointer], + user_data: ctypes.c_void_p, + /, +): + """Set callback for all future logging events. + + If this is not called, or NULL is supplied, everything is output on stderr.""" + ... # // -# // Beam search +# // Performance utils +# // +# // NOTE: Used by llama.cpp examples, avoid using in third-party apps. Instead, do your own performance measurements. # // - -# struct llama_beam_view { -# const llama_token * tokens; -# size_t n_tokens; -# float p; // Cumulative beam probability (renormalized relative to all beams) -# bool eob; // Callback should set this to true when a beam is at end-of-beam. +# struct llama_perf_context_data { +# double t_start_ms; +# double t_load_ms; +# double t_p_eval_ms; +# double t_eval_ms; +# +# int32_t n_p_eval; +# int32_t n_eval; # }; -class llama_beam_view(ctypes.Structure): +class llama_perf_context_data(ctypes.Structure): _fields_ = [ - ("tokens", llama_token_p), - ("n_tokens", c_size_t), - ("p", c_float), - ("eob", c_bool), + ("t_start_ms", ctypes.c_double), + ("t_load_ms", ctypes.c_double), + ("t_p_eval_ms", ctypes.c_double), + ("t_eval_ms", ctypes.c_double), + ("n_p_eval", ctypes.c_int32), + ("n_eval", ctypes.c_int32), ] -# // Passed to beam_search_callback function. -# // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams -# // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks. -# // These pointers are valid only during the synchronous callback, so should not be saved. -# struct llama_beams_state { -# struct llama_beam_view * beam_views; -# size_t n_beams; // Number of elements in beam_views[]. -# size_t common_prefix_length; // Current max length of prefix tokens shared by all beams. -# bool last_call; // True iff this is the last callback invocation. +# struct llama_perf_sampler_data { +# double t_sample_ms; +# +# int32_t n_sample; # }; -class llama_beams_state(ctypes.Structure): +class llama_perf_sampler_data(ctypes.Structure): _fields_ = [ - ("beam_views", POINTER(llama_beam_view)), - ("n_beams", c_size_t), - ("common_prefix_length", c_size_t), - ("last_call", c_bool), + ("t_sample_ms", ctypes.c_double), + ("n_sample", ctypes.c_int32), ] -# // Type of pointer to the beam_search_callback function. -# // void* callback_data is any custom data passed to llama_beam_search, that is subsequently -# // passed back to beam_search_callback. This avoids having to use global variables in the callback. -# typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state); -llama_beam_search_callback_fn_t = ctypes.CFUNCTYPE(None, c_void_p, llama_beams_state) - - -# /// @details Deterministically returns entire sentence constructed by a beam search. -# /// @param ctx Pointer to the llama_context. -# /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state. -# /// @param callback_data A pointer that is simply passed back to callback. -# /// @param n_beams Number of beams to use. -# /// @param n_past Number of tokens already evaluated. -# /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier. -# /// @param n_threads Number of threads as passed to llama_eval(). -# LLAMA_API void llama_beam_search( -# struct llama_context * ctx, -# llama_beam_search_callback_fn_t callback, -# void * callback_data, -# size_t n_beams, -# int n_past, -# int n_predict); -def llama_beam_search( - ctx: llama_context_p, - callback: "ctypes._CFuncPtr[None, c_void_p, llama_beams_state]", # type: ignore - callback_data: c_void_p, - n_beams: Union[c_size_t, int], - n_past: Union[c_int, int], - n_predict: Union[c_int, int], -): - return _lib.llama_beam_search( - ctx, callback, callback_data, n_beams, n_past, n_predict - ) - - -_lib.llama_beam_search.argtypes = [ - llama_context_p, - llama_beam_search_callback_fn_t, - c_void_p, - c_size_t, - c_int, - c_int, -] -_lib.llama_beam_search.restype = None - - -# Performance information - - -# LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx); -def llama_get_timings(ctx: llama_context_p) -> llama_timings: - """Get performance information""" - return _lib.llama_get_timings(ctx) - - -_lib.llama_get_timings.argtypes = [llama_context_p] -_lib.llama_get_timings.restype = llama_timings - - -# LLAMA_API void llama_print_timings(struct llama_context * ctx); -def llama_print_timings(ctx: llama_context_p): - """Print performance information""" - _lib.llama_print_timings(ctx) - - -_lib.llama_print_timings.argtypes = [llama_context_p] -_lib.llama_print_timings.restype = None - - -# LLAMA_API void llama_reset_timings(struct llama_context * ctx); -def llama_reset_timings(ctx: llama_context_p): - """Reset performance information""" - _lib.llama_reset_timings(ctx) - - -_lib.llama_reset_timings.argtypes = [llama_context_p] -_lib.llama_reset_timings.restype = None - +# LLAMA_API struct llama_perf_context_data llama_perf_context (const struct llama_context * ctx); +@ctypes_function( + "llama_perf_context", + [llama_context_p_ctypes], + llama_perf_context_data, +) +def llama_perf_context(ctx: llama_context_p, /) -> llama_perf_context_data: + ... -# Print system information -# LLAMA_API const char * llama_print_system_info(void); -def llama_print_system_info() -> bytes: - """Print system information""" - return _lib.llama_print_system_info() +# LLAMA_API void llama_perf_context_print(const struct llama_context * ctx); +@ctypes_function( + "llama_perf_context_print", + [llama_context_p_ctypes], + None, +) +def llama_perf_context_print(ctx: llama_context_p, /): + ... -_lib.llama_print_system_info.argtypes = [] -_lib.llama_print_system_info.restype = c_char_p +# LLAMA_API void llama_perf_context_reset( struct llama_context * ctx); +@ctypes_function( + "llama_perf_context_reset", + [llama_context_p_ctypes], + None, +) +def llama_perf_context_reset(ctx: llama_context_p, /): + ... -# NOTE: THIS IS CURRENTLY BROKEN AS ggml_log_callback IS NOT EXPOSED IN LLAMA.H -# // Set callback for all future logging events. -# // If this is not called, or NULL is supplied, everything is output on stderr. -# LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data); -def llama_log_set( - log_callback: "ctypes._FuncPointer", user_data: c_void_p # type: ignore -): - """Set callback for all future logging events. - If this is not called, or NULL is supplied, everything is output on stderr.""" - return _lib.llama_log_set(log_callback, user_data) +# // NOTE: the following work only with samplers constructed via llama_sampler_chain_init +# LLAMA_API struct llama_perf_sampler_data llama_perf_sampler (const struct llama_sampler * chain); +@ctypes_function( + "llama_perf_sampler", + [llama_sampler_p_ctypes], + llama_perf_sampler_data, +) +def llama_perf_sampler(chain: llama_sampler_p, /) -> llama_perf_sampler_data: + ... -_lib.llama_log_set.argtypes = [llama_log_callback, c_void_p] -_lib.llama_log_set.restype = None +# LLAMA_API void llama_perf_sampler_print(const struct llama_sampler * chain); +@ctypes_function( + "llama_perf_sampler_print", + [llama_sampler_p_ctypes], + None, +) +def llama_perf_sampler_print(chain: llama_sampler_p, /): + ... -# LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx); -def llama_dump_timing_info_yaml(stream: ctypes.c_void_p, ctx: llama_context_p): - return _lib.llama_dump_timing_info_yaml(stream, ctx) +# LLAMA_API void llama_perf_sampler_reset( struct llama_sampler * chain); +@ctypes_function( + "llama_perf_sampler_reset", + [llama_sampler_p_ctypes], + None, +) +def llama_perf_sampler_reset(chain: llama_sampler_p, /): + ... -_lib.llama_dump_timing_info_yaml.argtypes = [ctypes.c_void_p, llama_context_p] -_lib.llama_dump_timing_info_yaml.restype = None diff --git a/llama_cpp/llama_grammar.py b/llama_cpp/llama_grammar.py index aec023cd6..b95c77ab5 100644 --- a/llama_cpp/llama_grammar.py +++ b/llama_cpp/llama_grammar.py @@ -2,88 +2,28 @@ # flake8: noqa from pathlib import Path -import sys -from ctypes import * # type: ignore -from enum import Enum -from itertools import islice + +from itertools import groupby from typing import ( Any, - Callable, - Dict, - Generic, + Set, List, Optional, - OrderedDict, - TextIO, Tuple, - TypeVar, Union, - overload, ) -import llama_cpp.llama_cpp as llama_cpp - -# Type aliases -llama_grammar_element = llama_cpp.llama_grammar_element -llama_grammar_element_p = llama_cpp.llama_grammar_element_p -llama_grammar_p = llama_cpp.llama_grammar_p - -# Type variables -Ptr = TypeVar("Ptr", bound="const_char_p") -T = TypeVar("T") -U = TypeVar("U") -V = TypeVar("V") -W = TypeVar("W") - - -class Sentinel: - """Used to mark the end of a iterator of std::vector & std::map.""" +LLAMA_GRAMMAR_DEFAULT_ROOT = "root" class LlamaGrammar: - """Keeps reference counts of all the arguments, so that they are not - garbage collected by Python.""" - - def __del__(self) -> None: - """Free the grammar pointer when the object is deleted.""" - if self.grammar is not None: - llama_cpp.llama_grammar_free(self.grammar) - self.grammar = None - - def __init__( - self, - parsed_grammar: "parse_state", - ) -> None: - """Initialize the grammar pointer from the parsed state.""" - self._grammar_rules = ( - parsed_grammar.c_rules() - ) # type: std.vector[std.vector[LlamaGrammarElement]] - self._n_rules = self._grammar_rules.size() # type: int - self._start_rule_index = parsed_grammar.symbol_ids.at("root") # type: int - self.init() + def __init__(self, *args, _grammar: str, **kwargs): + self._grammar = _grammar + self._root = LLAMA_GRAMMAR_DEFAULT_ROOT @classmethod def from_string(cls, grammar: str, verbose: bool = True) -> "LlamaGrammar": - """Convert a GBNF grammar to a Llama grammar.""" - parsed_grammar = parse(const_char_p(grammar)) # type: parse_state - if parsed_grammar.rules.empty(): - raise ValueError( - f"{cls.from_string.__name__}: error parsing grammar file: parsed_grammar.rules is empty" - ) - if verbose: - print(f"{cls.from_string.__name__} grammar:", file=sys.stderr) - print_grammar(sys.stdout, parsed_grammar) - print(file=sys.stderr) - return cls(parsed_grammar) - - @classmethod - def from_json_schema( - cls, - json_schema: str, - verbose: bool = True, - ) -> "LlamaGrammar": - """Convert a JSON schema to a Llama grammar.""" - return cls.from_string(json_schema_to_gbnf(json_schema), verbose=verbose) + return cls(_grammar=grammar) @classmethod def from_file(cls, file: Union[str, Path], verbose: bool = True) -> "LlamaGrammar": @@ -102,1100 +42,9 @@ def from_file(cls, file: Union[str, Path], verbose: bool = True) -> "LlamaGramma f"{cls.from_file.__name__}: error parsing grammar file: params_grammer is empty" ) - def init(self) -> None: - # Step 1: Convert LlamaGrammarElement to llama_grammar_element - self._element_lists = [ - [ - llama_grammar_element(c_int(elem.type.value), c_uint32(elem.value)) - for elem in subvector - ] - for subvector in self._grammar_rules - ] # type: List[List[llama_grammar_element]] - - # Step 2: Convert each list to llama_grammar_element array and get pointer - self._element_arrays = [ - (llama_grammar_element * len(sublist))(*sublist) - for sublist in self._element_lists - ] # type: List[Array[llama_grammar_element]] - - # Step 3: Get pointer of each array - self._element_array_pointers = [ - cast(subarray, llama_grammar_element_p) for subarray in self._element_arrays - ] # type: List[llama_grammar_element_p] - - # Step 4: Make array of these pointers and get its pointer - self._rules = (llama_grammar_element_p * len(self._element_array_pointers))( - *self._element_array_pointers - ) - self.grammar = llama_cpp.llama_grammar_init( - self._rules, c_size_t(self._n_rules), c_size_t(self._start_rule_index) - ) - - def reset(self) -> None: - if self.grammar is not None: - llama_cpp.llama_grammar_free(self.grammar) - self.init() - - -class LlamaGrammarElement: - def __init__(self, type: "llama_gretype", value: int): - self.type = type - self.value = value # Unicode code point or rule ID - - -class const_char_p: - """C++ implementation of const char *.""" - - def __init__(self, value: Union[str, Ptr], move: Optional[int] = None): - if isinstance(value, const_char_p): - # We're copying an existing const_char_p - self.value = value.value - self.pos = value.pos + (move or 0) - return - - # We're creating a new const_char_p - self.value = value - self.pos = move or 0 - - def __str__(self) -> str: - assert self.value is not None, "null pointer" - return self.value[self.pos :] - - def __getitem__(self, index: int) -> str: - value = str(self) - return value[index] if index < len(value) else "" - - @overload - def __add__(self: Ptr, other: int) -> Ptr: - ... - - @overload - def __add__(self: Ptr, other: Ptr) -> int: - ... - - def __add__(self: Ptr, other: Union[int, Ptr]) -> Union[int, Ptr]: - return ( - self.__class__(self.value, self.pos + other) - if isinstance(other, int) - else self.pos + other.pos - ) - - @overload - def __sub__(self: Ptr, other: int) -> Ptr: - ... - - @overload - def __sub__(self: Ptr, other: Ptr) -> int: - ... - - def __sub__(self: Ptr, other: Union[int, Ptr]) -> Union[int, Ptr]: - return ( - self.__class__(self.value, self.pos - other) - if isinstance(other, int) - else self.pos - other.pos - ) - - def __eq__(self: Ptr, other: Ptr) -> bool: - assert self.value == other.value, "comparing pointers from different strings" - return self.pos == other.pos - - def __lt__(self: Ptr, other: Ptr) -> bool: - assert self.value == other.value, "comparing pointers from different strings" - return self.pos < other.pos - - def __gt__(self: Ptr, other: Ptr) -> bool: - assert self.value == other.value, "comparing pointers from different strings" - return self.pos > other.pos - - -class std: - @staticmethod - def string(ptr: const_char_p, length: Optional[int] = None) -> str: - """C++ implementation of std::string constructor.""" - value = str(ptr) - if length is not None: - value = value[:length] - return value - - class vector(Generic[T], List[T]): - """C++ implementation of std::vector.""" - - class iterator: - def __init__(self, vector: "std.vector[T]", index: int): - self._vector = vector - self._index = index - self._version = vector._version - - def _check_version(self): - if self._version != self._vector._version: - raise RuntimeError("Iterator used after vector was modified.") - - def __iter__(self): - return self - - def __next__(self) -> T: - self._check_version() - if self._index >= self._vector.size(): - raise StopIteration - value = self._vector[self._index] - self._index += 1 - return value - - def __add__(self, value: int) -> "std.vector[T].iterator": - return self.__class__(self._vector, self._index + value) - - def __sub__(self, value: int) -> "std.vector[T].iterator": - return self.__class__(self._vector, self._index - value) - - def __init__(self): - self._version = 0 - - def modify(self): - # This is a bit of a hack to make sure iterators are invalidated - self._version += 1 - - def push_back(self, value: T) -> None: - self.modify() - self.append(value) - - def pop_back(self) -> None: - self.modify() - if not self.empty(): - self.pop() - - def back(self) -> T: - return self[-1] - - def size(self) -> int: - return len(self) - - def clear(self) -> None: - self.modify() - super().clear() - - def empty(self) -> bool: - return self.size() == 0 - - def data(self) -> "std.vector[T]": - return self - - def resize( - self, - new_size: int, - fill_value_factory: Optional[Callable[[], T]] = None, - ) -> None: - if new_size > self.size(): - if fill_value_factory is None: - raise ValueError("A fill value factory function must be provided.") - self.reserve(new_size, fill_value_factory) - elif new_size < self.size(): - self[:] = self[:new_size] - - def reserve(self, capacity: int, fill_value_factory: Callable[[], T]) -> None: - if capacity > self.size(): - fill_value = fill_value_factory() - self.extend([fill_value] * (capacity - self.size())) - - def front(self) -> T: - if not self.empty(): - return self[0] - else: - raise IndexError("Vector is empty.") - - def assign(self, count: int, value: T) -> None: - self.clear() - self.extend([value] * count) - - def insert( - self, - pos: "std.vector[T].iterator", - first: "std.vector[T].iterator", - last: "std.vector[T].iterator", - ) -> None: - self[pos._index : pos._index] = list( - islice(first._vector, first._index, last._index) - ) - - def begin(self) -> "std.vector[T].iterator": - return self.iterator(self, 0) - - def end(self) -> "std.vector[T].iterator": - return self.iterator(self, self.size()) - - class map(Generic[T, U], OrderedDict[T, U]): - """C++ implementation of std::map.""" - - class iterator(Generic[V, W]): - def __init__(self, _map: "std.map[T, U]", key: Union[T, Sentinel]): - self._map = _map - self.iter = iter(_map) - self.key = key - self._advance() - - def _sanitize_key(self) -> T: - if isinstance(self.key, Sentinel): - raise StopIteration - return self.key - - def _advance(self) -> None: - try: - while next(self.iter) != self.key: - pass - except StopIteration: - self.key = Sentinel() - - def __next__(self) -> Tuple[T, U]: - key = self._sanitize_key() - if key in self._map: - value = self._map[key] - self._advance() - return key, value - else: - raise StopIteration - - def get(self) -> Tuple[T, U]: - key = self._sanitize_key() - return key, self._map[key] - - @property - def first(self) -> T: - return self._sanitize_key() - - @property - def second(self) -> U: - return self._map[self._sanitize_key()] - - def insert( - self, key: T, value: U - ) -> Tuple["std.map[T, U].iterator[T, U]", bool]: - if key in self: - return self.iterator(self, key), False - else: - self[key] = value - return self.iterator(self, key), True - - def find(self, key: T) -> "std.map[T, U].iterator[T, U]": - if key in self: - return self.iterator(self, key) - else: - return self.end() - - def at(self, key: T) -> U: - if key in self: - return self[key] - else: - raise KeyError("The provided key is not found in the map.") - - def erase(self, iterator: "std.map[T, U].iterator[T, U]") -> None: - key = iterator.first - if key in self: - del self[key] - - def size(self) -> int: - return len(self) - - def empty(self) -> bool: - return self.size() == 0 - - def lower_bound(self, key: T) -> "std.map[T, U].iterator[T, U]": - try: - keys = sorted(list(self.keys())) # type: ignore - for k in keys: - if k >= key: - return self.iterator(self, k) - raise ValueError("No key found that is not less than the input key") - except TypeError: - raise TypeError("Keys of type T cannot be sorted.") - - def begin(self) -> "std.map[T, U].iterator[T, U]": - return self.iterator(self, next(iter(self))) - - def end(self) -> "std.map[T, U].iterator[T, U]": - return self.iterator(self, Sentinel()) - - -# // grammar element type -# enum llama_gretype { -# // end of rule definition -# LLAMA_GRETYPE_END = 0, - -# // start of alternate definition for rule -# LLAMA_GRETYPE_ALT = 1, - -# // non-terminal element: reference to rule -# LLAMA_GRETYPE_RULE_REF = 2, - -# // terminal element: character (code point) -# LLAMA_GRETYPE_CHAR = 3, - -# // inverse char(s) ([^a], [^a-b] [^abc]) -# LLAMA_GRETYPE_CHAR_NOT = 4, - -# // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to -# // be an inclusive range ([a-z]) -# LLAMA_GRETYPE_CHAR_RNG_UPPER = 5, - - -# // modifies a preceding LLAMA_GRETYPE_CHAR or -# // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA]) -# LLAMA_GRETYPE_CHAR_ALT = 6, -# }; -class llama_gretype(Enum): - """grammar element type""" - - LLAMA_GRETYPE_END = 0 # end of rule definition - LLAMA_GRETYPE_ALT = 1 # start of alternate definition for rule - LLAMA_GRETYPE_RULE_REF = 2 # non-terminal element: reference to rule - LLAMA_GRETYPE_CHAR = 3 # terminal element: character (code point) - LLAMA_GRETYPE_CHAR_NOT = 4 # inverse char(s) ([^a], [^a-b] [^abc]) - LLAMA_GRETYPE_CHAR_RNG_UPPER = 5 # modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to be an inclusive range ([a-z]) - LLAMA_GRETYPE_CHAR_ALT = 6 # modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA]) - - -# struct parse_state { -# std::map symbol_ids; -# std::vector> rules; -# std::vector c_rules(); -# }; -class parse_state: - def __init__(self): - self.symbol_ids: std.map[str, int] = std.map() - self.rules: std.vector[std.vector[LlamaGrammarElement]] = std.vector() - - # std::vector parse_state::c_rules() { - # std::vector ret; - # for (const auto & rule : rules) { - # ret.push_back(rule.data()); - # } - # return ret; - # } - def c_rules(self) -> std.vector[std.vector[LlamaGrammarElement]]: - ret = std.vector() # type: std.vector[std.vector[LlamaGrammarElement]] - for rule in self.rules: - ret.push_back(rule.data()) - return ret - - def __repr__(self) -> str: - return ( - f"parse_state(symbol_ids={len(self.symbol_ids)}, rules={len(self.rules)})" - ) - - -# struct llama_grammar { -# const std::vector> rules; -# std::vector> stacks; -# }; -# class llama_grammar: -# def __init__( -# self, -# rules: std.vector[std.vector[llama_grammar_element]], -# stacks: std.vector[std.vector[llama_grammar_element]], -# ): -# self.rules = rules -# self.stacks = stacks - - -# uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) { -# uint32_t next_id = static_cast(state.symbol_ids.size()); -# auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id)); -# return result.first->second; -# } -def get_symbol_id(state: parse_state, src: const_char_p, len: int) -> int: - next_id = state.symbol_ids.size() # type: int - result = state.symbol_ids.insert(std.string(src, len), next_id) - return result[0].second # type: ignore - - -# uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) { -# uint32_t next_id = static_cast(state.symbol_ids.size()); -# state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id; -# return next_id; -# } -def generate_symbol_id(state: parse_state, base_name: str) -> int: - next_id = state.symbol_ids.size() # type: int - state.symbol_ids[base_name + "_" + str(next_id)] = next_id - return next_id - - -# void add_rule( -# parse_state & state, -# uint32_t rule_id, -# const std::vector & rule) { -# if (state.rules.size() <= rule_id) { -# state.rules.resize(rule_id + 1); -# } -# state.rules[rule_id] = rule; -# } -def add_rule( - state: parse_state, - rule_id: int, - rule: std.vector[LlamaGrammarElement], -) -> None: - if state.rules.size() <= rule_id: - state.rules.resize( - rule_id + 1, - fill_value_factory=std.vector[LlamaGrammarElement], - ) - state.rules[rule_id] = rule - - -# std::pair decode_utf8(const char * src) { -# static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 }; -# uint8_t first_byte = static_cast(*src); -# uint8_t highbits = first_byte >> 4; -# int len = lookup[highbits]; -# uint8_t mask = (1 << (8 - len)) - 1; -# uint32_t value = first_byte & mask; -# const char * end = src + len; // may overrun! -# const char * pos = src + 1; -# for ( ; pos < end && *pos; pos++) { -# value = (value << 6) + (static_cast(*pos) & 0x3F); -# } -# return std::make_pair(value, pos); -# } -def decode_utf8(src: const_char_p) -> Tuple[int, const_char_p]: - """Decodes a UTF-8 character from the source string.""" - lookup = (1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4) - first_byte = ord(src[0]) # type: int - highbits = first_byte >> 4 # type: int - len = lookup[highbits] # type: int - mask = (1 << (8 - len)) - 1 # type: int - value = first_byte & mask # type: int - end = src + len # type: const_char_p # may overrun! - pos = src + 1 # type: const_char_p - while pos < end and pos[0]: - value = (value << 6) + (ord(pos[0]) & 0x3F) - pos += 1 - return value, pos - - -# bool is_word_char(char c) { -# return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9'); -# } -def is_word_char(c: str) -> bool: - return ("a" <= c <= "z") or ("A" <= c <= "Z") or c == "-" or ("0" <= c <= "9") - - -# std::pair parse_hex(const char * src, int size) { -# const char * pos = src; -# const char * end = src + size; -# uint32_t value = 0; -# for ( ; pos < end && *pos; pos++) { -# value <<= 4; -# char c = *pos; -# if ('a' <= c && c <= 'f') { -# value += c - 'a' + 10; -# } else if ('A' <= c && c <= 'F') { -# value += c - 'A' + 10; -# } else if ('0' <= c && c <= '9') { -# value += c - '0'; -# } else { -# break; -# } -# } -# if (pos != end) { -# throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src); -# } -# return std::make_pair(value, pos); -# } -def parse_hex(src: const_char_p, size: int) -> Tuple[int, const_char_p]: - pos = const_char_p(src) # type: const_char_p - end = src + size # type: const_char_p - value = 0 # type: int - while pos < end and pos[0]: - value <<= 4 - c = pos[0] # type: str - if "a" <= c <= "f": - value += ord(c) - ord("a") + 10 - elif "A" <= c <= "F": - value += ord(c) - ord("A") + 10 - elif "0" <= c <= "9": - value += ord(c) - ord("0") - else: - break - pos += 1 - if pos != end: - raise RuntimeError("expecting " + str(size) + " hex chars at " + str(src)) - return (value, pos) - - -# std::pair parse_char(const char * src) { -# if (*src == '\\') { -# switch (src[1]) { -# case 'x': return parse_hex(src + 2, 2); -# case 'u': return parse_hex(src + 2, 4); -# case 'U': return parse_hex(src + 2, 8); -# case 't': return std::make_pair('\t', src + 2); -# case 'r': return std::make_pair('\r', src + 2); -# case 'n': return std::make_pair('\n', src + 2); -# case '\\': -# case '"': -# case '[': -# case ']': -# return std::make_pair(src[1], src + 2); -# default: -# throw std::runtime_error(std::string("unknown escape at ") + src); -# } -# } else if (*src) { -# return decode_utf8(src); -# } -# throw std::runtime_error("unexpected end of input"); -# } -def parse_char(src: const_char_p) -> Tuple[int, const_char_p]: - if src[0] == "\\": - case = src[1] # type: str - if case == "x": - return parse_hex(src + 2, 2) - elif case == "u": - return parse_hex(src + 2, 4) - elif case == "U": - return parse_hex(src + 2, 8) - elif case == "t": - return (ord("\t"), src + 2) # implicit cast - elif case == "r": - return (ord("\r"), src + 2) # implicit cast - elif case == "n": - return (ord("\n"), src + 2) # implicit cast - elif case in ("\\", '"', "[", "]"): - return (ord(case), src + 2) # implicit cast - else: - raise RuntimeError("unknown escape at " + str(src)) - elif src[0]: - return decode_utf8(src) - else: - raise RuntimeError("unexpected end of input") - - -# const char * parse_name(const char * src) { -# const char * pos = src; -# while (is_word_char(*pos)) { -# pos++; -# } -# if (pos == src) { -# throw std::runtime_error(std::string("expecting name at ") + src); -# } -# return pos; -# } -def parse_name(src: const_char_p) -> const_char_p: - pos = const_char_p(src) # type: const_char_p - while is_word_char(pos[0]): - pos += 1 - if pos == src: - raise RuntimeError("expecting name at " + str(src)) - return pos - - -# const char * parse_space(const char * src, bool newline_ok) { -# const char * pos = src; -# while (*pos == ' ' || *pos == '\t' || *pos == '#' || -# (newline_ok && (*pos == '\r' || *pos == '\n'))) { -# if (*pos == '#') { -# while (*pos && *pos != '\r' && *pos != '\n') { -# pos++; -# } -# } else { -# pos++; -# } -# } -# return pos; -# } -def parse_space(src: const_char_p, newline_ok: bool) -> const_char_p: - pos = const_char_p(src) # type: const_char_p - while pos[0] in (" ", "\t", "#") or (newline_ok and pos[0] in ("\r", "\n")): - if pos[0] == "#": - while pos[0] is not None and pos[0] not in ("\r", "\n"): - pos += 1 - else: - pos += 1 - return pos - - -# const char * parse_sequence( -# parse_state & state, -# const char * src, -# const std::string & rule_name, -# std::vector & out_elements, -# bool is_nested) { -def parse_sequence( - state: parse_state, - src: const_char_p, - rule_name: str, - out_elements: std.vector[LlamaGrammarElement], - is_nested: bool, -) -> const_char_p: - # size_t last_sym_start = out_elements.size(); - # const char * pos = src; - last_sym_start = out_elements.size() # type: int - pos = const_char_p(src) # type: const_char_p - # while (*pos) { - while pos[0]: - # if (*pos == '"') { // literal string - # pos++; - # last_sym_start = out_elements.size(); - # while (*pos != '"') { - # auto char_pair = parse_char(pos); - # pos = char_pair.second; - # out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first}); - # } - # pos = parse_space(pos + 1, is_nested); - if pos[0] == '"': # literal string - pos += 1 - last_sym_start = out_elements.size() - while pos[0] != '"': - char_pair = parse_char(pos) # type: Tuple[int, const_char_p] - pos = char_pair[1] - out_elements.push_back( - LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_CHAR, char_pair[0]) - ) - pos = parse_space(pos + 1, is_nested) - # } else if (*pos == '[') { // char range(s) - # pos++; - # enum llama_gretype start_type = LLAMA_GRETYPE_CHAR; - elif pos[0] == "[": # char range(s) - pos += 1 - start_type = llama_gretype.LLAMA_GRETYPE_CHAR # type: llama_gretype - # if (*pos == '^') { - # pos++; - # start_type = LLAMA_GRETYPE_CHAR_NOT; - # } - # last_sym_start = out_elements.size(); - if pos[0] == "^": - pos += 1 - start_type = llama_gretype.LLAMA_GRETYPE_CHAR_NOT - last_sym_start = out_elements.size() - # while (*pos != ']') { - # auto char_pair = parse_char(pos); - # pos = char_pair.second; - # enum llama_gretype type = last_sym_start < out_elements.size() - # ? LLAMA_GRETYPE_CHAR_ALT - # : start_type; - # out_elements.push_back({type, char_pair.first}); - while pos[0] != "]": - char_pair = parse_char(pos) # type: Tuple[int, const_char_p] - pos = char_pair[1] - type = ( - llama_gretype.LLAMA_GRETYPE_CHAR_ALT - if last_sym_start < out_elements.size() - else start_type - ) # type: llama_gretype - out_elements.push_back(LlamaGrammarElement(type, char_pair[0])) - # if (pos[0] == '-' && pos[1] != ']') { - # auto endchar_pair = parse_char(pos + 1); - # pos = endchar_pair.second; - # out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first}); - # } - # } - if pos[0] == "-" and pos[1] != "]": - endchar_pair = parse_char(pos + 1) # type: Tuple[int, const_char_p] - pos = endchar_pair[1] - out_elements.push_back( - LlamaGrammarElement( - llama_gretype.LLAMA_GRETYPE_CHAR_RNG_UPPER, - endchar_pair[0], - ) - ) - # pos = parse_space(pos + 1, is_nested); - pos = parse_space(pos + 1, is_nested) - # } else if (is_word_char(*pos)) { // rule reference - # const char * name_end = parse_name(pos); - # uint32_t ref_rule_id = get_symbol_id(state, pos, name_end - pos); - # pos = parse_space(name_end, is_nested); - # last_sym_start = out_elements.size(); - # out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id}); - elif is_word_char(pos[0]): # rule reference - name_end = parse_name(pos) # type: const_char_p - ref_rule_id = get_symbol_id(state, pos, name_end - pos) # type: int - pos = parse_space(name_end, is_nested) - last_sym_start = out_elements.size() - out_elements.push_back( - LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_RULE_REF, ref_rule_id) - ) - # } else if (*pos == '(') { // grouping - # // parse nested alternates into synthesized rule - # pos = parse_space(pos + 1, true); - # uint32_t sub_rule_id = generate_symbol_id(state, rule_name); - # pos = parse_alternates(state, pos, rule_name, sub_rule_id, true); - # last_sym_start = out_elements.size(); - # // output reference to synthesized rule - # out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); - # if (*pos != ')') { - # throw std::runtime_error(std::string("expecting ')' at ") + pos); - # } - # pos = parse_space(pos + 1, is_nested); - elif pos[0] == "(": # grouping - # parse nested alternates into synthesized rule - pos = parse_space(pos + 1, True) - sub_rule_id = generate_symbol_id(state, rule_name) # type: int - pos = parse_alternates(state, pos, rule_name, sub_rule_id, True) - last_sym_start = out_elements.size() - # output reference to synthesized rule - out_elements.push_back( - LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_RULE_REF, sub_rule_id) - ) - if pos[0] != ")": - raise RuntimeError("expecting ')' at " + str(pos)) - pos = parse_space(pos + 1, is_nested) - # } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator - # if (last_sym_start == out_elements.size()) { - # throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos); - # } - elif pos[0] in ("*", "+", "?"): # repetition operator - if last_sym_start == out_elements.size(): - raise RuntimeError("expecting preceding item to */+/? at " + str(pos)) - # // apply transformation to previous symbol (last_sym_start to end) according to - # // rewrite rules: - # // S* --> S' ::= S S' | - # // S+ --> S' ::= S S' | S - # // S? --> S' ::= S | - # uint32_t sub_rule_id = generate_symbol_id(state, rule_name); - # std::vector sub_rule; - # // add preceding symbol to generated rule - # sub_rule.insert( - # sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end()); - sub_rule_id = generate_symbol_id(state, rule_name) # type: int - sub_rule = std.vector[ - LlamaGrammarElement - ]() # type: std.vector[LlamaGrammarElement] - sub_rule.insert( - sub_rule.end(), - out_elements.begin() + last_sym_start, - out_elements.end(), - ) - # if (*pos == '*' || *pos == '+') { - # // cause generated rule to recurse - # sub_rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); - # } - # // mark start of alternate def - # sub_rule.push_back({LLAMA_GRETYPE_ALT, 0}); - if pos[0] in ("*", "+"): - sub_rule.push_back( - LlamaGrammarElement( - llama_gretype.LLAMA_GRETYPE_RULE_REF, sub_rule_id - ) - ) - sub_rule.push_back(LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_ALT, 0)) - # if (*pos == '+') { - # // add preceding symbol as alternate only for '+' (otherwise empty) - # sub_rule.insert( - # sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end()); - # } - # sub_rule.push_back({LLAMA_GRETYPE_END, 0}); - # add_rule(state, sub_rule_id, sub_rule); - # // in original rule, replace previous symbol with reference to generated rule - # out_elements.resize(last_sym_start); - # out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); - # pos = parse_space(pos + 1, is_nested); - if pos[0] == "+": - # add preceding symbol as alternate only for '+' (otherwise empty) - sub_rule.insert( - sub_rule.end(), - out_elements.begin() + last_sym_start, - out_elements.end(), - ) - sub_rule.push_back(LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_END, 0)) - add_rule(state, sub_rule_id, sub_rule) - # in original rule, replace previous symbol with reference to generated rule - out_elements.resize(last_sym_start) - out_elements.push_back( - LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_RULE_REF, sub_rule_id) - ) - pos = parse_space(pos + 1, is_nested) - # } else { - # break; - # } - else: - break - # } - # return pos; - # } - return pos - - -# const char * parse_alternates( -# parse_state & state, -# const char * src, -# const std::string & rule_name, -# uint32_t rule_id, -# bool is_nested) { -# std::vector rule; -# const char * pos = parse_sequence(state, src, rule_name, rule, is_nested); -# while (*pos == '|') { -# rule.push_back({LLAMA_GRETYPE_ALT, 0}); -# pos = parse_space(pos + 1, true); -# pos = parse_sequence(state, pos, rule_name, rule, is_nested); -# } -# rule.push_back({LLAMA_GRETYPE_END, 0}); -# add_rule(state, rule_id, rule); -# return pos; -# } -def parse_alternates( - state: parse_state, - src: const_char_p, - rule_name: str, - rule_id: int, - is_nested: bool, -) -> const_char_p: - rule = std.vector() # type: std.vector[LlamaGrammarElement] - pos = parse_sequence(state, src, rule_name, rule, is_nested) # type: const_char_p - while pos[0] == "|": - rule.push_back(LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_ALT, 0)) - pos = parse_space(pos + 1, True) - pos = parse_sequence(state, pos, rule_name, rule, is_nested) - rule.push_back(LlamaGrammarElement(llama_gretype.LLAMA_GRETYPE_END, 0)) - add_rule(state, rule_id, rule) - return pos - - -# const char * parse_rule(parse_state & state, const char * src) { -# const char * name_end = parse_name(src); -# const char * pos = parse_space(name_end, false); -# size_t name_len = name_end - src; -# uint32_t rule_id = get_symbol_id(state, src, name_len); -# const std::string name(src, name_len); - -# if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) { -# throw std::runtime_error(std::string("expecting ::= at ") + pos); -# } -# pos = parse_space(pos + 3, true); - -# pos = parse_alternates(state, pos, name, rule_id, false); - - -# if (*pos == '\r') { -# pos += pos[1] == '\n' ? 2 : 1; -# } else if (*pos == '\n') { -# pos++; -# } else if (*pos) { -# throw std::runtime_error(std::string("expecting newline or end at ") + pos); -# } -# return parse_space(pos, true); -# } -def parse_rule(state: parse_state, src: const_char_p) -> const_char_p: - name_end = parse_name(src) # type: const_char_p - pos = parse_space(name_end, False) # type: const_char_p - name_len = name_end - src # type: int - rule_id = get_symbol_id(state, src, name_len) # type: int - name = std.string(src, name_len) # type: str - - if not (pos[0] == ":" and pos[1] == ":" and pos[2] == "="): - raise RuntimeError("expecting ::= at " + str(pos)) - - pos = parse_space(pos + 3, True) # type: const_char_p - pos = parse_alternates(state, pos, name, rule_id, False) # type: const_char_p - - if pos[0] == "\r": - pos += 2 if pos[1] == "\n" else 1 - elif pos[0] == "\n": - pos += 1 - elif pos[0]: - raise RuntimeError("expecting newline or end at " + str(pos)) - return parse_space(pos, True) - - -# parse_state parse(const char * src) { -# try { -# parse_state state; -# const char * pos = parse_space(src, true); -# while (*pos) { -# pos = parse_rule(state, pos); -# } -# return state; -# } catch (const std::exception & err) { -# fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what()); -# return parse_state(); -# } -# } -def parse(src: const_char_p) -> parse_state: - try: - state = parse_state() # type: parse_state - pos = parse_space(src, True) # type: const_char_p - while pos[0]: - pos = parse_rule(state, pos) - return state - except Exception as err: - print(f"{parse.__name__}: error parsing grammar: {err}") - return parse_state() - - -# void print_grammar_char(FILE * file, uint32_t c) { -# if (0x20 <= c && c <= 0x7f) { -# fprintf(file, "%c", static_cast(c)); -# } else { -# // cop out of encoding UTF-8 -# fprintf(file, "", c); -# } -# } -def print_grammar_char(file: TextIO, c: int) -> None: - if 0x20 <= c and c <= 0x7F: - file.write(chr(c)) - else: - # cop out of encoding UTF-8 - file.write(f"") - - -# bool is_char_element(llama_grammar_element elem) { -# switch (elem.type) { -# case LLAMA_GRETYPE_CHAR: return true; -# case LLAMA_GRETYPE_CHAR_NOT: return true; -# case LLAMA_GRETYPE_CHAR_ALT: return true; -# case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true; -# default: return false; -# } -# } -def is_char_element(elem: LlamaGrammarElement) -> bool: - return elem.type in ( - llama_gretype.LLAMA_GRETYPE_CHAR, - llama_gretype.LLAMA_GRETYPE_CHAR_NOT, - llama_gretype.LLAMA_GRETYPE_CHAR_ALT, - llama_gretype.LLAMA_GRETYPE_CHAR_RNG_UPPER, - ) - - -# void print_rule( -# FILE * file, -# uint32_t rule_id, -# const std::vector & rule, -# const std::map & symbol_id_names) { -def print_rule( - file: TextIO, - rule_id: int, - rule: std.vector[LlamaGrammarElement], - symbol_id_names: std.map[int, str], -) -> None: - # if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) { - # throw std::runtime_error( - # "malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id)); - # } - # fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str()); - if rule.empty() or rule.back().type != llama_gretype.LLAMA_GRETYPE_END: - raise RuntimeError( - "malformed rule, does not end with LLAMA_GRETYPE_END: " + str(rule_id) - ) - print(f"{symbol_id_names.at(rule_id)} ::=", file=file, end=" ") - # for (size_t i = 0, end = rule.size() - 1; i < end; i++) { - # llama_grammar_element elem = rule[i]; - # switch (elem.type) { - # case LLAMA_GRETYPE_END: - # throw std::runtime_error( - # "unexpected end of rule: " + std::to_string(rule_id) + "," + - # std::to_string(i)); - # case LLAMA_GRETYPE_ALT: - # fprintf(file, "| "); - # break; - # case LLAMA_GRETYPE_RULE_REF: - # fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str()); - # break; - # case LLAMA_GRETYPE_CHAR: - # fprintf(file, "["); - # print_grammar_char(file, elem.value); - # break; - # case LLAMA_GRETYPE_CHAR_NOT: - # fprintf(file, "[^"); - # print_grammar_char(file, elem.value); - # break; - # case LLAMA_GRETYPE_CHAR_RNG_UPPER: - # if (i == 0 || !is_char_element(rule[i - 1])) { - # throw std::runtime_error( - # "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " + - # std::to_string(rule_id) + "," + std::to_string(i)); - # } - # fprintf(file, "-"); - # print_grammar_char(file, elem.value); - # break; - # case LLAMA_GRETYPE_CHAR_ALT: - # if (i == 0 || !is_char_element(rule[i - 1])) { - # throw std::runtime_error( - # "LLAMA_GRETYPE_CHAR_ALT without preceding char: " + - # std::to_string(rule_id) + "," + std::to_string(i)); - # } - # print_grammar_char(file, elem.value); - # break; - # } - for i, elem in enumerate(rule[:-1]): - case = elem.type # type: llama_gretype - if case is llama_gretype.LLAMA_GRETYPE_END: - raise RuntimeError("unexpected end of rule: " + str(rule_id) + "," + str(i)) - elif case is llama_gretype.LLAMA_GRETYPE_ALT: - print("| ", file=file, end="") - elif case is llama_gretype.LLAMA_GRETYPE_RULE_REF: - print(f"{symbol_id_names.at(elem.value)} ", file=file, end="") - elif case is llama_gretype.LLAMA_GRETYPE_CHAR: - print("[", file=file, end="") - print_grammar_char(file, elem.value) - elif case is llama_gretype.LLAMA_GRETYPE_CHAR_NOT: - print("[^", file=file, end="") - print_grammar_char(file, elem.value) - elif case is llama_gretype.LLAMA_GRETYPE_CHAR_RNG_UPPER: - if i == 0 or not is_char_element(rule[i - 1]): - raise RuntimeError( - "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " - + str(rule_id) - + "," - + str(i) - ) - print("-", file=file, end="") - print_grammar_char(file, elem.value) - elif case is llama_gretype.LLAMA_GRETYPE_CHAR_ALT: - if i == 0 or not is_char_element(rule[i - 1]): - raise RuntimeError( - "LLAMA_GRETYPE_CHAR_ALT without preceding char: " - + str(rule_id) - + "," - + str(i) - ) - print_grammar_char(file, elem.value) - # if (is_char_element(elem)) { - # switch (rule[i + 1].type) { - # case LLAMA_GRETYPE_CHAR_ALT: - # case LLAMA_GRETYPE_CHAR_RNG_UPPER: - # break; - # default: - # fprintf(file, "] "); - if is_char_element(elem): - if rule[i + 1].type in ( - llama_gretype.LLAMA_GRETYPE_CHAR_ALT, - llama_gretype.LLAMA_GRETYPE_CHAR_RNG_UPPER, - ): - pass - else: - print("] ", file=file, end="") - # } - # } - # } - # fprintf(file, "\n"); - # } - print(file=file) - - -# void print_grammar(FILE * file, const parse_state & state) { -# try { -# std::map symbol_id_names; -# for (auto kv : state.symbol_ids) { -# symbol_id_names[kv.second] = kv.first; -# } -# for (size_t i = 0, end = state.rules.size(); i < end; i++) { -# // fprintf(file, "%zu: ", i); -# // print_rule_binary(file, state.rules[i]); -# print_rule(file, i, state.rules[i], symbol_id_names); -# // fprintf(file, "\n"); -# } -# } catch (const std::exception & err) { -# fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what()); -# } -# } -def print_grammar(file: TextIO, state: parse_state) -> None: - try: - symbol_id_names = std.map() # type: std.map[int, str] - for kv in state.symbol_ids.items(): - symbol_id_names[kv[1]] = kv[0] - - for i, rule in enumerate(state.rules): - print_rule(file, i, rule, symbol_id_names) - except Exception as err: - print( - f"{print_grammar.__name__}: error printing grammar: {err}", - file=sys.stderr, - ) + @classmethod + def from_json_schema(cls, json_schema: str, verbose: bool = True) -> "LlamaGrammar": + return cls.from_string(json_schema_to_gbnf(json_schema), verbose=verbose) """llama.cpp gbnf rules from vendor/llama.cpp/grammars""" @@ -1337,7 +186,7 @@ def print_grammar(file: TextIO, state: parse_state) -> None: string ::= "\"" ( - [^"\\] | + [^"\\\x7F\x00-\x1F] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes )* "\"" ws @@ -1366,13 +215,14 @@ def print_grammar(file: TextIO, state: parse_state) -> None: string ::= "\"" ( - [^"\\] | - "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes + [^"\\\x7F\x00-\x1F] | + "\\" (["\\bfnrt] | "u" [0-9a-fA-F]{4}) # escapes )* "\"" ws -number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws +number ::= ("-"? ([0-9] | [1-9] [0-9]{0,15})) ("." [0-9]+)? ([eE] [-+]? [0-9] [1-9]{0,15})? ws -ws ::= ([ \t\n] ws)? +# Optional space: by convention, applied in this grammar after literal chars when allowed +ws ::= | " " | "\n" [ \t]{0,20} """ LIST_GBNF = r""" @@ -1391,128 +241,713 @@ def print_grammar(file: TextIO, state: parse_state) -> None: # whitespace. Also maybe improves generation quality? SPACE_RULE = '" "?' -PRIMITIVE_RULES = { - "boolean": '("true" | "false") space', - "number": '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space', - "integer": '("-"? ([0-9] | [1-9] [0-9]*)) space', - "string": r""" "\"" ( - [^"\\] | - "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) - )* "\"" space """, - "null": '"null" space', -} INVALID_RULE_CHARS_RE = re.compile(r"[^a-zA-Z0-9-]+") GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]') GRAMMAR_LITERAL_ESCAPES = {"\r": "\\r", "\n": "\\n", '"': '\\"'} +# whitespace is constrained to a single space char to prevent model "running away" in +# whitespace. Also maybe improves generation quality? +SPACE_RULE = '" "?' + + +def _build_repetition( + item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False +): + if not separator_rule: + if min_items == 0 and max_items == 1: + return f"{item_rule}?" + elif min_items == 1 and max_items is None: + return f"{item_rule}+" + + result = "" + + if min_items > 0: + if item_rule_is_literal and separator_rule is None: + result = '"' + (item_rule[1:-1] * min_items) + '"' + else: + result = (f" {separator_rule} " if separator_rule else " ").join( + [item_rule] * min_items + ) + + def opt_repetitions(up_to_n, prefix_with_sep=False): + """ + - n=4, no sep: '(a (a (a (a)?)?)?)?' + - n=4, sep=',', prefix: '("," a ("," a ("," a ("," a)?)?)?)?' + - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?' + """ + + content = ( + f"{separator_rule} {item_rule}" + if prefix_with_sep and separator_rule + else item_rule + ) + if up_to_n == 0: + return "" + elif up_to_n == 1: + return f"({content})?" + elif separator_rule and not prefix_with_sep: + return f"({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?" + else: + return (f"({content} " * up_to_n).rstrip() + (")?" * up_to_n) + + if min_items > 0 and max_items != min_items: + result += " " + + if max_items is not None: + result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0) + else: + item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})' + + if min_items == 0 and separator_rule: + result = f"({item_rule} {item_operator}*)?" + else: + result += f"{item_operator}*" + + return result + + +class BuiltinRule: + def __init__(self, content: str, deps: list = None): + self.content = content + self.deps = deps or [] + + +_up_to_15_digits = _build_repetition("[0-9]", 0, 15) + +PRIMITIVE_RULES = { + "boolean": BuiltinRule('("true" | "false") space', []), + "decimal-part": BuiltinRule("[0-9] " + _up_to_15_digits, []), + "integral-part": BuiltinRule("[0-9] | [1-9] " + _up_to_15_digits, []), + "number": BuiltinRule( + '("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', + ["integral-part", "decimal-part"], + ), + "integer": BuiltinRule('("-"? integral-part) space', ["integral-part"]), + "value": BuiltinRule( + "object | array | string | number | boolean | null", + ["object", "array", "string", "number", "boolean", "null"], + ), + "object": BuiltinRule( + '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', + ["string", "value"], + ), + "array": BuiltinRule( + '"[" space ( value ("," space value)* )? "]" space', ["value"] + ), + "uuid": BuiltinRule( + r'"\"" ' + + ' "-" '.join("[0-9a-fA-F]" * n for n in [8, 4, 4, 4, 12]) + + r' "\"" space', + [], + ), + "char": BuiltinRule( + r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', + [], + ), + "string": BuiltinRule(r'"\"" char* "\"" space', ["char"]), + "null": BuiltinRule('"null" space', []), +} + +# TODO: support "uri", "email" string formats +STRING_FORMAT_RULES = { + "date": BuiltinRule( + '[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( "0" [1-9] | [1-2] [0-9] | "3" [0-1] )', + [], + ), + "time": BuiltinRule( + '([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', + [], + ), + "date-time": BuiltinRule('date "T" time', ["date", "time"]), + "date-string": BuiltinRule('"\\"" date "\\"" space', ["date"]), + "time-string": BuiltinRule('"\\"" time "\\"" space', ["time"]), + "date-time-string": BuiltinRule('"\\"" date-time "\\"" space', ["date-time"]), +} + +DOTALL = "[\\U00000000-\\U0010FFFF]" +DOT = "[^\\x0A\\x0D]" + +RESERVED_NAMES = set( + ["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()] +) + + +NON_LITERAL_SET = set("|.()[]{}*+?") +ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set("[]()|{}*+?") + class SchemaConverter: - def __init__(self, prop_order): + def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern): self._prop_order = prop_order - self._rules = {"space": SPACE_RULE} - self._defs: Dict[str, Any] = {} - - def _format_literal(self, literal: str): - escaped: str = GRAMMAR_LITERAL_ESCAPE_RE.sub( - lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal) + self._allow_fetch = allow_fetch + self._dotall = dotall + self._raw_pattern = raw_pattern + self._rules = { + "space": SPACE_RULE, + } + self._refs = {} + self._refs_being_resolved = set() + + def _format_literal(self, literal): + escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub( + lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal ) return f'"{escaped}"' - def _add_rule(self, name: str, rule: str): + def not_literal( + self, literal: str, dotall: bool = True, maybe_escaped_underscores=False + ) -> str: + """ + not_literal('a') -> '[^a]' + not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?' + """ + assert len(literal) > 0, "Empty literal not supported" + + def recurse(i: int): + c = literal[i] + if maybe_escaped_underscores and c == "_": + yield f"[^{c}\\\\]" + yield " | " + yield f'"\\\\"? "{c}"' + else: + yield f"[^{c}]" + if i < len(literal) - 1: + yield " | " + yield self._format_literal(c) + yield " (" + yield from recurse(i + 1) + yield ")?" + + return "".join(("(", *recurse(0), ")")) + + def _add_rule(self, name, rule): esc_name = INVALID_RULE_CHARS_RE.sub("-", name) if esc_name not in self._rules or self._rules[esc_name] == rule: key = esc_name else: i = 0 - while f"{esc_name}{i}" in self._rules: + while ( + f"{esc_name}{i}" in self._rules + and self._rules[f"{esc_name}{i}"] != rule + ): i += 1 key = f"{esc_name}{i}" self._rules[key] = rule return key - def visit(self, schema: Dict[str, Any], name: str) -> str: - schema_type: Optional[str] = schema.get("type") # type: ignore - assert isinstance(schema_type, str), f"Unrecognized schema: {schema}" - rule_name = name or "root" - - if "$defs" in schema: - # add defs to self._defs for later inlining - for def_name, def_schema in schema["$defs"].items(): - self._defs[def_name] = def_schema - - if "oneOf" in schema or "anyOf" in schema: - rule = " | ".join( - ( - self.visit(alt_schema, f'{name}{"-" if name else ""}{i}') - for i, alt_schema in enumerate( - schema.get("oneOf") or schema["anyOf"] + def resolve_refs(self, schema: dict, url: str): + """ + Resolves all $ref fields in the given schema, fetching any remote schemas, + replacing $ref with absolute reference URL and populating self._refs with the + respective referenced (sub)schema dictionaries. + """ + + def visit(n: dict): + if isinstance(n, list): + return [visit(x) for x in n] + elif isinstance(n, dict): + ref = n.get("$ref") + if ref is not None and ref not in self._refs: + if ref.startswith("https://"): + assert ( + self._allow_fetch + ), "Fetching remote schemas is not allowed (use --allow-fetch for force)" + import requests + + frag_split = ref.split("#") + base_url = frag_split[0] + + target = self._refs.get(base_url) + if target is None: + target = self.resolve_refs( + requests.get(ref).json(), base_url + ) + self._refs[base_url] = target + + if len(frag_split) == 1 or frag_split[-1] == "": + return target + elif ref.startswith("#/"): + target = schema + ref = f"{url}{ref}" + n["$ref"] = ref + else: + raise ValueError(f"Unsupported ref {ref}") + + for sel in ref.split("#")[-1].split("/")[1:]: + assert ( + target is not None and sel in target + ), f"Error resolving ref {ref}: {sel} not in {target}" + target = target[sel] + + self._refs[ref] = target + else: + for v in n.values(): + visit(v) + + return n + + return visit(schema) + + def _generate_union_rule(self, name, alt_schemas): + return " | ".join( + ( + self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}') + for i, alt_schema in enumerate(alt_schemas) + ) + ) + + def _visit_pattern(self, pattern, name): + """ + Transforms a regular expression pattern into a GBNF rule. + + Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions + Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md + + Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers. + + Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which + we define sub-rules to keep the output lean. + """ + + assert pattern.startswith("^") and pattern.endswith( + "$" + ), 'Pattern must start with "^" and end with "$"' + pattern = pattern[1:-1] + sub_rule_ids = {} + + i = 0 + length = len(pattern) + + def to_rule(s: Tuple[str, bool]) -> str: + (txt, is_literal) = s + return '"' + txt + '"' if is_literal else txt + + def transform() -> Tuple[str, bool]: + """ + Parse a unit at index i (advancing it), and return its string representation + whether it's a literal. + """ + nonlocal i + nonlocal pattern + nonlocal sub_rule_ids + + start = i + # For each component of this sequence, store its string representation and whether it's a literal. + # We only need a flat structure here to apply repetition operators to the last item, and + # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially + # (GBNF's syntax is luckily very close to regular expressions!) + seq: list[Tuple[str, bool]] = [] + + def get_dot(): + if self._dotall: + rule = DOTALL + else: + # Accept any character... except \n and \r line break chars (\x0A and \xOD) + rule = DOT + return self._add_rule(f"dot", rule) + + def join_seq(): + nonlocal seq + ret = [] + for is_literal, g in groupby(seq, lambda x: x[1]): + if is_literal: + ret.append(("".join(x[0] for x in g), True)) + else: + ret.extend(g) + if len(ret) == 1: + return ret[0] + return (" ".join(to_rule(x) for x in seq), False) + + while i < length: + c = pattern[i] + if c == ".": + seq.append((get_dot(), False)) + i += 1 + elif c == "(": + i += 1 + if i < length: + assert ( + pattern[i] != "?" + ), f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/' + seq.append((f"({to_rule(transform())})", False)) + elif c == ")": + i += 1 + assert ( + start > 0 and pattern[start - 1] == "(" + ), f"Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}" + return join_seq() + elif c == "[": + square_brackets = c + i += 1 + while i < length and pattern[i] != "]": + if pattern[i] == "\\": + square_brackets += pattern[i : i + 2] + i += 2 + else: + square_brackets += pattern[i] + i += 1 + assert ( + i < length + ), f"Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}" + square_brackets += "]" + i += 1 + seq.append((square_brackets, False)) + elif c == "|": + seq.append(("|", False)) + i += 1 + elif c in ("*", "+", "?"): + seq[-1] = (to_rule(seq[-1]) + c, False) + i += 1 + elif c == "{": + curly_brackets = c + i += 1 + while i < length and pattern[i] != "}": + curly_brackets += pattern[i] + i += 1 + assert ( + i < length + ), f"Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}" + curly_brackets += "}" + i += 1 + nums = [s.strip() for s in curly_brackets[1:-1].split(",")] + min_times = 0 + max_times = None + try: + if len(nums) == 1: + min_times = int(nums[0]) + max_times = min_times + else: + assert len(nums) == 2 + min_times = int(nums[0]) if nums[0] else 0 + max_times = int(nums[1]) if nums[1] else None + except ValueError: + raise ValueError( + f"Invalid quantifier {curly_brackets} in /{pattern}/" + ) + + (sub, sub_is_literal) = seq[-1] + + if not sub_is_literal: + id = sub_rule_ids.get(sub) + if id is None: + id = self._add_rule(f"{name}-{len(sub_rule_ids) + 1}", sub) + sub_rule_ids[sub] = id + sub = id + + seq[-1] = ( + _build_repetition( + f'"{sub}"' if sub_is_literal else sub, + min_times, + max_times, + item_rule_is_literal=sub_is_literal, + ), + False, ) - ) + else: + literal = "" + while i < length: + if pattern[i] == "\\" and i < length - 1: + next = pattern[i + 1] + if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS: + i += 1 + literal += pattern[i] + i += 1 + else: + literal += pattern[i : i + 2] + i += 2 + elif pattern[i] == '"' and not self._raw_pattern: + literal += '\\"' + i += 1 + elif pattern[i] not in NON_LITERAL_SET and ( + i == length - 1 + or literal == "" + or pattern[i + 1] == "." + or pattern[i + 1] not in NON_LITERAL_SET + ): + literal += pattern[i] + i += 1 + else: + break + if literal: + seq.append((literal, True)) + + return join_seq() + + return self._add_rule( + name, + ( + to_rule(transform()) + if self._raw_pattern + else '"\\"" ' + to_rule(transform()) + ' "\\"" space' + ), + ) + + def _resolve_ref(self, ref): + ref_name = ref.split("/")[-1] + if ref_name not in self._rules and ref not in self._refs_being_resolved: + self._refs_being_resolved.add(ref) + resolved = self._refs[ref] + ref_name = self.visit(resolved, ref_name) + self._refs_being_resolved.remove(ref) + return ref_name + + def _generate_constant_rule(self, value): + return self._format_literal(json.dumps(value)) + + def visit(self, schema, name): + schema_type = schema.get("type") + schema_format = schema.get("format") + rule_name = name + "-" if name in RESERVED_NAMES else name or "root" + + if (ref := schema.get("$ref")) is not None: + return self._add_rule(rule_name, self._resolve_ref(ref)) + + elif "oneOf" in schema or "anyOf" in schema: + return self._add_rule( + rule_name, + self._generate_union_rule(name, schema.get("oneOf") or schema["anyOf"]), + ) + + elif isinstance(schema_type, list): + return self._add_rule( + rule_name, + self._generate_union_rule(name, [{"type": t} for t in schema_type]), ) - return self._add_rule(rule_name, rule) elif "const" in schema: - return self._add_rule(rule_name, self._format_literal(schema["const"])) + return self._add_rule( + rule_name, self._generate_constant_rule(schema["const"]) + ) elif "enum" in schema: - rule = " | ".join((self._format_literal(v) for v in schema["enum"])) + rule = " | ".join((self._generate_constant_rule(v) for v in schema["enum"])) return self._add_rule(rule_name, rule) - elif schema_type == "object" and "properties" in schema: - # TODO: `required` keyword - prop_order = self._prop_order - prop_pairs = sorted( - schema["properties"].items(), - # sort by position in prop_order (if specified) then by key - key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]), + elif schema_type in (None, "object") and ( + "properties" in schema + or ( + "additionalProperties" in schema + and schema["additionalProperties"] is not True + ) + ): + required = set(schema.get("required", [])) + properties = list(schema.get("properties", {}).items()) + return self._add_rule( + rule_name, + self._build_object_rule( + properties, required, name, schema.get("additionalProperties") + ), + ) + + elif schema_type in (None, "object") and "allOf" in schema: + required = set() + properties = [] + hybrid_name = name + + def add_component(comp_schema, is_required): + if (ref := comp_schema.get("$ref")) is not None: + comp_schema = self._refs[ref] + + if "properties" in comp_schema: + for prop_name, prop_schema in comp_schema["properties"].items(): + properties.append((prop_name, prop_schema)) + if is_required: + required.add(prop_name) + + for t in schema["allOf"]: + if "anyOf" in t: + for tt in t["anyOf"]: + add_component(tt, is_required=False) + else: + add_component(t, is_required=True) + + return self._add_rule( + rule_name, + self._build_object_rule( + properties, required, hybrid_name, additional_properties=[] + ), ) - rule = '"{" space' - for i, (prop_name, prop_schema) in enumerate(prop_pairs): - prop_rule_name = self.visit( - prop_schema, f'{name}{"-" if name else ""}{prop_name}' + elif schema_type in (None, "array") and ( + "items" in schema or "prefixItems" in schema + ): + items = schema.get("items") or schema["prefixItems"] + if isinstance(items, list): + return self._add_rule( + rule_name, + '"[" space ' + + ' "," space '.join( + self.visit(item, f'{name}{"-" if name else ""}tuple-{i}') + for i, item in enumerate(items) + ) + + ' "]" space', + ) + else: + item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item') + min_items = schema.get("minItems", 0) + max_items = schema.get("maxItems") + return self._add_rule( + rule_name, + '"[" space ' + + _build_repetition( + item_rule_name, min_items, max_items, separator_rule='"," space' + ) + + ' "]" space', ) - if i > 0: - rule += ' "," space' - rule += rf' {self._format_literal(prop_name)} space ":" space {prop_rule_name}' - rule += ' "}" space' - return self._add_rule(rule_name, rule) + elif schema_type in (None, "string") and "pattern" in schema: + return self._visit_pattern(schema["pattern"], rule_name) - elif schema_type == "array" and "items" in schema: - # TODO `prefixItems` keyword - item_rule_name = self.visit( - schema["items"], f'{name}{"-" if name else ""}item' + elif schema_type in (None, "string") and re.match( + r"^uuid[1-5]?$", schema_format or "" + ): + return self._add_primitive( + "root" if rule_name == "root" else schema_format, + PRIMITIVE_RULES["uuid"], ) - rule = ( - f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space' + + elif ( + schema_type in (None, "string") + and f"{schema_format}-string" in STRING_FORMAT_RULES + ): + prim_name = f"{schema_format}-string" + return self._add_rule( + rule_name, + self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name]), ) - return self._add_rule(rule_name, rule) - elif "$ref" in schema: - ref = schema["$ref"] - assert ref.startswith("#/$defs/"), f"Unrecognized schema: {schema}" - # inline $defs - def_name = ref[len("#/$defs/") :] - def_schema = self._defs[def_name] - return self.visit(def_schema, f'{name}{"-" if name else ""}{def_name}') + elif schema_type == "string" and ( + "minLength" in schema or "maxLength" in schema + ): + char_rule = self._add_primitive("char", PRIMITIVE_RULES["char"]) + min_len = schema.get("minLength", 0) + max_len = schema.get("maxLength") + + return self._add_rule( + rule_name, + r'"\"" ' + + _build_repetition(char_rule, min_len, max_len) + + r' "\"" space', + ) + + elif (schema_type == "object") or (len(schema) == 0): + return self._add_rule( + rule_name, self._add_primitive("object", PRIMITIVE_RULES["object"]) + ) else: assert schema_type in PRIMITIVE_RULES, f"Unrecognized schema: {schema}" - return self._add_rule( + # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero + return self._add_primitive( "root" if rule_name == "root" else schema_type, PRIMITIVE_RULES[schema_type], ) + def _add_primitive(self, name: str, rule: BuiltinRule): + n = self._add_rule(name, rule.content) + + for dep in rule.deps: + dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep) + assert dep_rule, f"Rule {dep} not known" + if dep not in self._rules: + self._add_primitive(dep, dep_rule) + return n + + def _build_object_rule( + self, + properties: List[Tuple[str, Any]], + required: Set[str], + name: str, + additional_properties: Union[bool, Any], + ): + prop_order = self._prop_order + # sort by position in prop_order (if specified) then by original order + sorted_props = [ + kv[0] + for _, kv in sorted( + enumerate(properties), + key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]), + ) + ] + + prop_kv_rule_names = {} + for prop_name, prop_schema in properties: + prop_rule_name = self.visit( + prop_schema, f'{name}{"-" if name else ""}{prop_name}' + ) + prop_kv_rule_names[prop_name] = self._add_rule( + f'{name}{"-" if name else ""}{prop_name}-kv', + rf'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}', + ) + required_props = [k for k in sorted_props if k in required] + optional_props = [k for k in sorted_props if k not in required] + + if additional_properties == True or isinstance(additional_properties, dict): + sub_name = f'{name}{"-" if name else ""}additional' + value_rule = self.visit( + {} if additional_properties == True else additional_properties, + f"{sub_name}-value", + ) + prop_kv_rule_names["*"] = self._add_rule( + f"{sub_name}-kv", + self._add_primitive("string", PRIMITIVE_RULES["string"]) + + f' ":" space {value_rule}', + ) + optional_props.append("*") + + rule = '"{" space ' + rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props) + + if optional_props: + rule += " (" + if required_props: + rule += ' "," space ( ' + + def get_recursive_refs(ks, first_is_optional): + [k, *rest] = ks + kv_rule_name = prop_kv_rule_names[k] + if k == "*": + res = self._add_rule( + f'{name}{"-" if name else ""}additional-kvs', + f'{kv_rule_name} ( "," space ' + kv_rule_name + " )*", + ) + elif first_is_optional: + res = f'( "," space {kv_rule_name} )?' + else: + res = kv_rule_name + if len(rest) > 0: + res += " " + self._add_rule( + f'{name}{"-" if name else ""}{k}-rest', + get_recursive_refs(rest, first_is_optional=True), + ) + return res + + rule += " | ".join( + get_recursive_refs(optional_props[i:], first_is_optional=False) + for i in range(len(optional_props)) + ) + if required_props: + rule += " )" + rule += " )?" + + rule += ' "}" space' + + return rule + def format_grammar(self): - return "\n".join((f"{name} ::= {rule}" for name, rule in self._rules.items())) + return "\n".join( + f"{name} ::= {rule}" + for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0]) + ) def json_schema_to_gbnf(schema: str, prop_order: Optional[List[str]] = None): prop_order = prop_order or [] schema = json.loads(schema) prop_order = {name: idx for idx, name in enumerate(prop_order)} - converter = SchemaConverter(prop_order) + converter = SchemaConverter( + prop_order=prop_order, allow_fetch=False, dotall=False, raw_pattern=False + ) + schema = converter.resolve_refs(schema, "stdin") converter.visit(schema, "") return converter.format_grammar() diff --git a/llama_cpp/llama_speculative.py b/llama_cpp/llama_speculative.py new file mode 100644 index 000000000..39dfb903b --- /dev/null +++ b/llama_cpp/llama_speculative.py @@ -0,0 +1,64 @@ +import abc + +from typing import Any + +import numpy as np +import numpy.typing as npt + + +class LlamaDraftModel(abc.ABC): + @abc.abstractmethod + def __call__( + self, input_ids: npt.NDArray[np.intc], /, **kwargs: Any + ) -> npt.NDArray[np.intc]: + raise NotImplementedError() + + +class LlamaPromptLookupDecoding(LlamaDraftModel): + """Based on https://github.com/apoorvumang/prompt-lookup-decoding""" + + def __init__(self, max_ngram_size: int = 2, num_pred_tokens: int = 10): + self.max_ngram_size = max_ngram_size + self.num_pred_tokens = num_pred_tokens + + @staticmethod + def find_candidate_pred_tokens( + input_ids: npt.NDArray[np.intc], + max_ngram_size: int, + num_pred_tokens: int, + ): + input_length = input_ids.shape[0] + + for ngram_size in range(min(max_ngram_size, input_length - 1), 0, -1): + # Create sliding windows of size ngram_size + windows = np.lib.stride_tricks.sliding_window_view(input_ids, (ngram_size,)) + + # Convert ngram to an array for comparison + ngram_array = input_ids[-ngram_size:] + + # Find where the windows match the ngram + matches = np.all(windows == ngram_array, axis=1) + + # Get the indices of matches + match_indices = np.nonzero(matches)[0] + + # Iterate through match indices to find a valid continuation + for idx in match_indices: + start_idx = idx + ngram_size + end_idx = start_idx + num_pred_tokens + end_idx = min(end_idx, input_length) + + if start_idx < end_idx: + return input_ids[start_idx:end_idx] + + # If no match is found, return an empty array + return np.array([], dtype=np.intc) + + def __call__( + self, input_ids: npt.NDArray[np.intc], /, **kwargs: Any + ) -> npt.NDArray[np.intc]: + return self.find_candidate_pred_tokens( + input_ids=input_ids, + max_ngram_size=self.max_ngram_size, + num_pred_tokens=self.num_pred_tokens, + ) diff --git a/llama_cpp/llama_tokenizer.py b/llama_cpp/llama_tokenizer.py new file mode 100644 index 000000000..1375e1392 --- /dev/null +++ b/llama_cpp/llama_tokenizer.py @@ -0,0 +1,120 @@ +from __future__ import annotations + +import abc +from typing import ( + List, + Optional, + Any, +) + +import llama_cpp +from llama_cpp.llama_types import List + + +class BaseLlamaTokenizer(abc.ABC): + @abc.abstractmethod + def tokenize( + self, text: bytes, add_bos: bool = True, special: bool = True + ) -> List[int]: + """Tokenize the text into tokens. + + Args: + text: The utf-8 encoded string to tokenize. + add_bos: Whether to add a beginning of sequence token. + special: Whether to tokenize special tokens. + """ + raise NotImplementedError + + @abc.abstractmethod + def detokenize( + self, + tokens: List[int], + prev_tokens: Optional[List[int]] = None, + special: bool = False, + ) -> bytes: + """Detokenize the tokens into text. + + Args: + tokens: The list of tokens to detokenize. + prev_tokens: The list of previous tokens. Offset mapping will be performed if provided. + special: Whether to detokenize special tokens. + """ + raise NotImplementedError + + +class LlamaTokenizer(BaseLlamaTokenizer): + def __init__(self, llama: llama_cpp.Llama): + self._model = llama._model # type: ignore + + def tokenize( + self, text: bytes, add_bos: bool = True, special: bool = True + ) -> List[int]: + return self._model.tokenize(text, add_bos=add_bos, special=special) + + def detokenize( + self, + tokens: List[int], + prev_tokens: Optional[List[int]] = None, + special: bool = False, + ) -> bytes: + return self._model.detokenize(tokens, special=special) + + def encode( + self, text: str, add_bos: bool = True, special: bool = True + ) -> List[int]: + return self.tokenize( + text.encode("utf-8", errors="ignore"), add_bos=add_bos, special=special + ) + + def decode(self, tokens: List[int]) -> str: + return self.detokenize(tokens).decode("utf-8", errors="ignore") + + @classmethod + def from_ggml_file(cls, path: str) -> "LlamaTokenizer": + return cls(llama_cpp.Llama(model_path=path, vocab_only=True)) + + +class LlamaHFTokenizer(BaseLlamaTokenizer): + def __init__(self, hf_tokenizer: Any): + self.hf_tokenizer = hf_tokenizer + + def tokenize( + self, text: bytes, add_bos: bool = True, special: bool = True + ) -> List[int]: + return self.hf_tokenizer.encode( + text.decode("utf-8", errors="ignore"), add_special_tokens=special + ) + + def detokenize( + self, + tokens: List[int], + prev_tokens: Optional[List[int]] = None, + special: bool = False, + ) -> bytes: + skip_special_tokens = not special + if prev_tokens is not None: + text = self.hf_tokenizer.decode( + prev_tokens + tokens, skip_special_tokens=skip_special_tokens + ).encode("utf-8", errors="ignore") + prev_text = self.hf_tokenizer.decode( + prev_tokens, skip_special_tokens=skip_special_tokens + ).encode("utf-8", errors="ignore") + return text[len(prev_text) :] + else: + return self.hf_tokenizer.decode( + tokens, skip_special_tokens=skip_special_tokens + ).encode("utf-8", errors="ignore") + + @classmethod + def from_pretrained(cls, pretrained_model_name_or_path: str) -> "LlamaHFTokenizer": + try: + from transformers import AutoTokenizer + except ImportError: + raise ImportError( + "The `transformers` library is required to use the `HFTokenizer`." + "You can install it with `pip install transformers`." + ) + hf_tokenizer = AutoTokenizer.from_pretrained( + pretrained_model_name_or_path=pretrained_model_name_or_path + ) + return cls(hf_tokenizer) diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index 5b51e98ce..f647822ff 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -6,6 +6,7 @@ https://github.com/openai/openai-openapi/blob/master/openapi.yaml """ + from typing import Any, List, Optional, Dict, Union from typing_extensions import TypedDict, NotRequired, Literal @@ -24,7 +25,7 @@ class EmbeddingUsage(TypedDict): class Embedding(TypedDict): index: int object: str - embedding: List[float] + embedding: Union[List[float], List[List[float]]] class CreateEmbeddingResponse(TypedDict): @@ -81,9 +82,28 @@ class ChatCompletionFunction(TypedDict): parameters: Dict[str, JsonType] # TODO: make this more specific +class ChatCompletionTopLogprobToken(TypedDict): + token: str + logprob: float + bytes: Optional[List[int]] + + +class ChatCompletionLogprobToken(ChatCompletionTopLogprobToken): + token: str + logprob: float + bytes: Optional[List[int]] + top_logprobs: List[ChatCompletionTopLogprobToken] + + +class ChatCompletionLogprobs(TypedDict): + content: Optional[List[ChatCompletionLogprobToken]] + refusal: Optional[List[ChatCompletionLogprobToken]] + + class ChatCompletionResponseChoice(TypedDict): index: int message: "ChatCompletionResponseMessage" + logprobs: Optional[ChatCompletionLogprobs] finish_reason: Optional[str] @@ -97,7 +117,7 @@ class CreateChatCompletionResponse(TypedDict): class ChatCompletionMessageToolCallChunkFunction(TypedDict): - name: str + name: Optional[str] arguments: str @@ -118,12 +138,12 @@ class ChatCompletionStreamResponseDeltaFunctionCall(TypedDict): class ChatCompletionStreamResponseDelta(TypedDict): - content: NotRequired[str] + content: NotRequired[Optional[str]] function_call: NotRequired[ - ChatCompletionStreamResponseDeltaFunctionCall + Optional[ChatCompletionStreamResponseDeltaFunctionCall] ] # DEPRECATED - tool_calls: NotRequired[List[ChatCompletionMessageToolCallChunk]] - role: NotRequired[Literal["system", "user", "assistant", "tool"]] + tool_calls: NotRequired[Optional[List[ChatCompletionMessageToolCallChunk]]] + role: NotRequired[Optional[Literal["system", "user", "assistant", "tool"]]] class ChatCompletionStreamResponseChoice(TypedDict): @@ -132,6 +152,7 @@ class ChatCompletionStreamResponseChoice(TypedDict): ChatCompletionStreamResponseDelta, ChatCompletionStreamResponseDeltaEmpty ] finish_reason: Optional[Literal["stop", "length", "tool_calls", "function_call"]] + logprobs: NotRequired[Optional[ChatCompletionLogprobs]] class CreateChatCompletionStreamResponse(TypedDict): @@ -154,6 +175,9 @@ class ChatCompletionFunctionCallOption(TypedDict): class ChatCompletionRequestResponseFormat(TypedDict): type: Literal["text", "json_object"] + schema: NotRequired[ + JsonType + ] # https://docs.endpoints.anyscale.com/guides/json_mode/ class ChatCompletionRequestMessageContentPartText(TypedDict): @@ -208,7 +232,7 @@ class ChatCompletionRequestAssistantMessageFunctionCall(TypedDict): class ChatCompletionRequestAssistantMessage(TypedDict): role: Literal["assistant"] - content: Optional[str] + content: NotRequired[str] tool_calls: NotRequired[ChatCompletionMessageToolCalls] function_call: NotRequired[ ChatCompletionRequestAssistantMessageFunctionCall @@ -269,7 +293,7 @@ class ChatCompletionNamedToolChoice(TypedDict): ChatCompletionToolChoiceOption = Union[ - Literal["none", "auto"], ChatCompletionNamedToolChoice + Literal["none", "auto", "required"], ChatCompletionNamedToolChoice ] diff --git a/llama_cpp/llava_cpp.py b/llama_cpp/llava_cpp.py index b1f90b915..d9dfaf5fd 100644 --- a/llama_cpp/llava_cpp.py +++ b/llama_cpp/llava_cpp.py @@ -1,86 +1,47 @@ -import sys +from __future__ import annotations + import os -import ctypes from ctypes import ( c_bool, c_char_p, c_int, - c_int8, - c_int32, c_uint8, - c_uint32, - c_size_t, c_float, - c_double, c_void_p, POINTER, _Pointer, # type: ignore Structure, - Array, ) import pathlib -from typing import List, Union +from typing import ( + Union, + NewType, + Optional, + TYPE_CHECKING, +) import llama_cpp.llama_cpp as llama_cpp -# Load the library -def _load_shared_library(lib_base_name: str): - # Construct the paths to the possible shared library names - _base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) - # Searching for the library in the current directory under the name "libllama" (default name - # for llamacpp) and "llama" (default name for this repo) - _lib_paths: List[pathlib.Path] = [] - # Determine the file extension based on the platform - if sys.platform.startswith("linux"): - _lib_paths += [ - _base_path / f"lib{lib_base_name}.so", - ] - elif sys.platform == "darwin": - _lib_paths += [ - _base_path / f"lib{lib_base_name}.so", - _base_path / f"lib{lib_base_name}.dylib", - ] - elif sys.platform == "win32": - _lib_paths += [ - _base_path / f"{lib_base_name}.dll", - _base_path / f"lib{lib_base_name}.dll", - ] - else: - raise RuntimeError("Unsupported platform") - - if "LLAVA_CPP_LIB" in os.environ: - lib_base_name = os.environ["LLAVA_CPP_LIB"] - _lib = pathlib.Path(lib_base_name) - _base_path = _lib.parent.resolve() - _lib_paths = [_lib.resolve()] - - cdll_args = dict() # type: ignore - # Add the library directory to the DLL search path on Windows (if needed) - if sys.platform == "win32" and sys.version_info >= (3, 8): - os.add_dll_directory(str(_base_path)) - if "CUDA_PATH" in os.environ: - os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin")) - os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib")) - cdll_args["winmode"] = ctypes.RTLD_GLOBAL - - # Try to load the shared library, handling potential errors - for _lib_path in _lib_paths: - if _lib_path.exists(): - try: - return ctypes.CDLL(str(_lib_path), **cdll_args) - except Exception as e: - raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}") - - raise FileNotFoundError( - f"Shared library with base name '{lib_base_name}' not found" +from llama_cpp._ctypes_extensions import ( + load_shared_library, + ctypes_function_for_shared_library, +) + +if TYPE_CHECKING: + from llama_cpp._ctypes_extensions import ( + CtypesArray, ) # Specify the base name of the shared library to load _libllava_base_name = "llava" +_libllava_override_path = os.environ.get("LLAVA_CPP_LIB") +_libllava_base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" if _libllava_override_path is None else pathlib.Path() # Load the library -_libllava = _load_shared_library(_libllava_base_name) +_libllava = load_shared_library(_libllava_base_name, _libllava_base_path) + +ctypes_function = ctypes_function_for_shared_library(_libllava) ################################################ @@ -88,7 +49,9 @@ def _load_shared_library(lib_base_name: str): ################################################ # struct clip_ctx; -clip_ctx_p = c_void_p +clip_ctx_p = NewType("clip_ctx_p", int) +clip_ctx_p_ctypes = c_void_p + # struct llava_image_embed { # float * embed; @@ -100,45 +63,77 @@ class llava_image_embed(Structure): ("n_image_pos", c_int), ] + # /** sanity check for clip <-> llava embed size match */ # LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip); -def llava_validate_embed_size(ctx_llama: llama_cpp.llama_context_p, ctx_clip: clip_ctx_p) -> bool: - return _libllava.llava_validate_embed_size(ctx_llama, ctx_clip) +@ctypes_function( + "llava_validate_embed_size", + [llama_cpp.llama_context_p_ctypes, clip_ctx_p_ctypes], + c_bool, +) +def llava_validate_embed_size( + ctx_llama: llama_cpp.llama_context_p, ctx_clip: clip_ctx_p, / +) -> bool: + ... -_libllava.llava_validate_embed_size.argtypes = [llama_cpp.llama_context_p, clip_ctx_p] -_libllava.llava_validate_embed_size.restype = c_bool # /** build an image embed from image file bytes */ # LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length); -def llava_image_embed_make_with_bytes(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_bytes: bytes, image_bytes_length: Union[c_int, int]) -> "_Pointer[llava_image_embed]": - return _libllava.llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length) +@ctypes_function( + "llava_image_embed_make_with_bytes", + [clip_ctx_p_ctypes, c_int, POINTER(c_uint8), c_int], + POINTER(llava_image_embed), +) +def llava_image_embed_make_with_bytes( + ctx_clip: clip_ctx_p, + n_threads: Union[c_int, int], + image_bytes: CtypesArray[c_uint8], + image_bytes_length: Union[c_int, int], + /, +) -> "_Pointer[llava_image_embed]": + ... -_libllava.llava_image_embed_make_with_bytes.argtypes = [clip_ctx_p, c_int, POINTER(c_uint8), c_int] -_libllava.llava_image_embed_make_with_bytes.restype = POINTER(llava_image_embed) # /** build an image embed from a path to an image filename */ # LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path); -def llava_image_embed_make_with_filename(ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_path: bytes) -> "_Pointer[llava_image_embed]": - return _libllava.llava_image_embed_make_with_filename(ctx_clip, n_threads, image_path) +@ctypes_function( + "llava_image_embed_make_with_filename", + [clip_ctx_p_ctypes, c_int, c_char_p], + POINTER(llava_image_embed), +) +def llava_image_embed_make_with_filename( + ctx_clip: clip_ctx_p, n_threads: Union[c_int, int], image_path: bytes, / +) -> "_Pointer[llava_image_embed]": + ... -_libllava.llava_image_embed_make_with_filename.argtypes = [clip_ctx_p, c_int, c_char_p] -_libllava.llava_image_embed_make_with_filename.restype = POINTER(llava_image_embed) # LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); # /** free an embedding made with llava_image_embed_make_* */ -def llava_image_embed_free(embed: "_Pointer[llava_image_embed]"): - return _libllava.llava_image_embed_free(embed) +@ctypes_function("llava_image_embed_free", [POINTER(llava_image_embed)], None) +def llava_image_embed_free(embed: "_Pointer[llava_image_embed]", /): + ... -_libllava.llava_image_embed_free.argtypes = [POINTER(llava_image_embed)] -_libllava.llava_image_embed_free.restype = None # /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */ # LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past); -def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointer[llava_image_embed]", n_batch: Union[c_int, int], n_past: "_Pointer[c_int]") -> bool: - return _libllava.llava_eval_image_embed(ctx_llama, embed, n_batch, n_past) - -_libllava.llava_eval_image_embed.argtypes = [llama_cpp.llama_context_p, POINTER(llava_image_embed), c_int, POINTER(c_int)] -_libllava.llava_eval_image_embed.restype = c_bool +@ctypes_function( + "llava_eval_image_embed", + [ + llama_cpp.llama_context_p_ctypes, + POINTER(llava_image_embed), + c_int, + POINTER(c_int), + ], + c_bool, +) +def llava_eval_image_embed( + ctx_llama: llama_cpp.llama_context_p, + embed: "_Pointer[llava_image_embed]", + n_batch: Union[c_int, int], + n_past: "_Pointer[c_int]", + /, +) -> bool: + ... ################################################ @@ -146,87 +141,18 @@ def llava_eval_image_embed(ctx_llama: llama_cpp.llama_context_p, embed: "_Pointe ################################################ -# struct clip_vision_hparams { -# int32_t image_size; -# int32_t patch_size; -# int32_t hidden_size; -# int32_t n_intermediate; -# int32_t projection_dim; -# int32_t n_head; -# int32_t n_layer; -# float eps; -# }; -class clip_vision_hparams(Structure): - _fields_ = [ - ("image_size", c_int32), - ("patch_size", c_int32), - ("hidden_size", c_int32), - ("n_intermediate", c_int32), - ("projection_dim", c_int32), - ("n_head", c_int32), - ("n_layer", c_int32), - ("eps", c_float), - ] - # /** load mmproj model */ -# CLIP_API struct clip_ctx * clip_model_load(const char * fname, const int verbosity); -def clip_model_load(fname: bytes, verbosity: Union[c_int, int]) -> clip_ctx_p: - return _libllava.clip_model_load(fname, verbosity) +# CLIP_API struct clip_ctx * clip_model_load (const char * fname, int verbosity); +@ctypes_function("clip_model_load", [c_char_p, c_int], clip_ctx_p_ctypes) +def clip_model_load( + fname: bytes, verbosity: Union[c_int, int], / +) -> Optional[clip_ctx_p]: + ... -_libllava.clip_model_load.argtypes = [c_char_p, c_int] -_libllava.clip_model_load.restype = clip_ctx_p # /** free mmproj model */ # CLIP_API void clip_free(struct clip_ctx * ctx); -def clip_free(ctx: clip_ctx_p): - return _libllava.clip_free(ctx) - -_libllava.clip_free.argtypes = [clip_ctx_p] -_libllava.clip_free.restype = None - -# size_t clip_embd_nbytes(const struct clip_ctx * ctx); -# int clip_n_patches(const struct clip_ctx * ctx); -# int clip_n_mmproj_embd(const struct clip_ctx * ctx); - -# // RGB uint8 image -# struct clip_image_u8 { -# int nx; -# int ny; -# uint8_t * data = NULL; -# size_t size; -# }; - -# // RGB float32 image (NHWC) -# // Memory layout: RGBRGBRGB... -# struct clip_image_f32 { -# int nx; -# int ny; -# float * data = NULL; -# size_t size; -# }; - -# struct clip_image_u8_batch { -# struct clip_image_u8 * data; -# size_t size; -# }; - -# struct clip_image_f32_batch { -# struct clip_image_f32 * data; -# size_t size; -# }; - -# struct clip_image_u8 * make_clip_image_u8(); -# struct clip_image_f32 * make_clip_image_f32(); -# CLIP_API void clip_image_u8_free(clip_image_u8 * img); -# CLIP_API void clip_image_f32_free(clip_image_f32 * img); -# CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img); -# /** interpret bytes as an image file with length bytes_length, and use the result to populate img */ -# CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img); - -# bool clip_image_preprocess(const struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, const bool pad2square); -# bool clip_image_encode(const struct clip_ctx * ctx, const int n_threads, struct clip_image_f32 * img, float * vec); - -# bool clip_image_batch_encode(const struct clip_ctx * ctx, const int n_threads, const struct clip_image_f32_batch * imgs, -# float * vec); +@ctypes_function("clip_free", [clip_ctx_p_ctypes], None) +def clip_free(ctx: clip_ctx_p, /): + ... -# bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype); \ No newline at end of file diff --git a/llama_cpp/server/__main__.py b/llama_cpp/server/__main__.py index 45fc5a857..bbac4957e 100644 --- a/llama_cpp/server/__main__.py +++ b/llama_cpp/server/__main__.py @@ -9,7 +9,7 @@ Then run: ``` -uvicorn llama_cpp.server.app:app --reload +uvicorn llama_cpp.server.app:create_app --reload ``` or @@ -21,81 +21,80 @@ Then visit http://localhost:8000/docs to see the interactive API docs. """ + +from __future__ import annotations + import os +import sys import argparse -from typing import List, Literal, Union import uvicorn -from llama_cpp.server.app import create_app, Settings - -def get_base_type(annotation): - if getattr(annotation, '__origin__', None) is Literal: - return type(annotation.__args__[0]) - elif getattr(annotation, '__origin__', None) is Union: - non_optional_args = [arg for arg in annotation.__args__ if arg is not type(None)] - if non_optional_args: - return get_base_type(non_optional_args[0]) - elif getattr(annotation, '__origin__', None) is list or getattr(annotation, '__origin__', None) is List: - return get_base_type(annotation.__args__[0]) - else: - return annotation - -def contains_list_type(annotation) -> bool: - origin = getattr(annotation, '__origin__', None) - - if origin is list or origin is List: - return True - elif origin in (Literal, Union): - return any(contains_list_type(arg) for arg in annotation.__args__) - else: - return False - -def parse_bool_arg(arg): - if isinstance(arg, bytes): - arg = arg.decode('utf-8') - - true_values = {'1', 'on', 't', 'true', 'y', 'yes'} - false_values = {'0', 'off', 'f', 'false', 'n', 'no'} - - arg_str = str(arg).lower().strip() - - if arg_str in true_values: - return True - elif arg_str in false_values: - return False - else: - raise ValueError(f'Invalid boolean argument: {arg}') - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - for name, field in Settings.model_fields.items(): - description = field.description - if field.default is not None and description is not None: - description += f" (default: {field.default})" - base_type = get_base_type(field.annotation) if field.annotation is not None else str - list_type = contains_list_type(field.annotation) - if base_type is not bool: - parser.add_argument( - f"--{name}", - dest=name, - nargs="*" if list_type else None, - type=base_type, - help=description, - ) - if base_type is bool: - parser.add_argument( - f"--{name}", - dest=name, - type=parse_bool_arg, - help=f"{description}", - ) - +from llama_cpp.server.app import create_app +from llama_cpp.server.settings import ( + Settings, + ServerSettings, + ModelSettings, + ConfigFileSettings, +) +from llama_cpp.server.cli import add_args_from_model, parse_model_from_args + + +def main(): + description = "🦙 Llama.cpp python server. Host your own LLMs!🚀" + parser = argparse.ArgumentParser(description=description) + + add_args_from_model(parser, Settings) + parser.add_argument( + "--config_file", + type=str, + help="Path to a config file to load.", + ) + server_settings: ServerSettings | None = None + model_settings: list[ModelSettings] = [] args = parser.parse_args() - settings = Settings(**{k: v for k, v in vars(args).items() if v is not None}) - app = create_app(settings=settings) - + try: + # Load server settings from config_file if provided + config_file = os.environ.get("CONFIG_FILE", args.config_file) + if config_file: + if not os.path.exists(config_file): + raise ValueError(f"Config file {config_file} not found!") + with open(config_file, "rb") as f: + # Check if yaml file + if config_file.endswith(".yaml") or config_file.endswith(".yml"): + import yaml + import json + + config_file_settings = ConfigFileSettings.model_validate_json( + json.dumps(yaml.safe_load(f)) + ) + else: + config_file_settings = ConfigFileSettings.model_validate_json( + f.read() + ) + server_settings = ServerSettings.model_validate(config_file_settings) + model_settings = config_file_settings.models + else: + server_settings = parse_model_from_args(ServerSettings, args) + model_settings = [parse_model_from_args(ModelSettings, args)] + except Exception as e: + print(e, file=sys.stderr) + parser.print_help() + sys.exit(1) + assert server_settings is not None + assert model_settings is not None + app = create_app( + server_settings=server_settings, + model_settings=model_settings, + ) uvicorn.run( - app, host=os.getenv("HOST", settings.host), port=int(os.getenv("PORT", settings.port)), - ssl_keyfile=settings.ssl_keyfile, ssl_certfile=settings.ssl_certfile + app, + host=os.getenv("HOST", server_settings.host), + port=int(os.getenv("PORT", server_settings.port)), + ssl_keyfile=server_settings.ssl_keyfile, + ssl_certfile=server_settings.ssl_certfile, ) + + +if __name__ == "__main__": + main() diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index fa39047f8..5120f2416 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -1,371 +1,141 @@ -import sys +from __future__ import annotations + +import os import json -import traceback -import multiprocessing -import time -from re import compile, Match, Pattern -from threading import Lock +import typing +import contextlib + +from anyio import Lock from functools import partial -from typing import Callable, Coroutine, Iterator, List, Optional, Tuple, Union, Dict -from typing_extensions import TypedDict, Literal +from typing import List, Optional, Union, Dict import llama_cpp import anyio from anyio.streams.memory import MemoryObjectSendStream from starlette.concurrency import run_in_threadpool, iterate_in_threadpool -from fastapi import Depends, FastAPI, APIRouter, Request, Response +from fastapi import Depends, FastAPI, APIRouter, Request, HTTPException, status, Body from fastapi.middleware import Middleware from fastapi.middleware.cors import CORSMiddleware -from fastapi.responses import JSONResponse -from fastapi.routing import APIRoute -from pydantic import BaseModel, Field -from pydantic_settings import BaseSettings +from fastapi.security import HTTPBearer from sse_starlette.sse import EventSourceResponse -from starlette_context import plugins +from starlette_context.plugins import RequestIdPlugin # type: ignore from starlette_context.middleware import RawContextMiddleware -import numpy as np -import numpy.typing as npt +from llama_cpp.server.model import ( + LlamaProxy, +) +from llama_cpp.server.settings import ( + ConfigFileSettings, + Settings, + ModelSettings, + ServerSettings, +) +from llama_cpp.server.types import ( + CreateCompletionRequest, + CreateEmbeddingRequest, + CreateChatCompletionRequest, + ModelList, + TokenizeInputRequest, + TokenizeInputResponse, + TokenizeInputCountResponse, + DetokenizeInputRequest, + DetokenizeInputResponse, +) +from llama_cpp.server.errors import RouteErrorHandler -# Disable warning for model and model_alias settings -BaseSettings.model_config["protected_namespaces"] = () +router = APIRouter(route_class=RouteErrorHandler) - -class Settings(BaseSettings): - model: str = Field( - description="The path to the model to use for generating completions." - ) - model_alias: Optional[str] = Field( - default=None, - description="The alias of the model to use for generating completions.", - ) - # Model Params - n_gpu_layers: int = Field( - default=0, - ge=-1, - description="The number of layers to put on the GPU. The rest will be on the CPU. Set -1 to move all to GPU.", - ) - main_gpu: int = Field( - default=0, - ge=0, - description="Main GPU to use.", - ) - tensor_split: Optional[List[float]] = Field( - default=None, - description="Split layers across multiple GPUs in proportion.", - ) - vocab_only: bool = Field( - default=False, description="Whether to only return the vocabulary." - ) - use_mmap: bool = Field( - default=llama_cpp.llama_mmap_supported(), - description="Use mmap.", - ) - use_mlock: bool = Field( - default=llama_cpp.llama_mlock_supported(), - description="Use mlock.", - ) - # Context Params - seed: int = Field( - default=llama_cpp.LLAMA_DEFAULT_SEED, description="Random seed. -1 for random." - ) - n_ctx: int = Field(default=2048, ge=1, description="The context size.") - n_batch: int = Field( - default=512, ge=1, description="The batch size to use per eval." - ) - n_threads: int = Field( - default=max(multiprocessing.cpu_count() // 2, 1), - ge=1, - description="The number of threads to use.", - ) - n_threads_batch: int = Field( - default=max(multiprocessing.cpu_count() // 2, 1), - ge=0, - description="The number of threads to use when batch processing.", - ) - rope_scaling_type: int = Field(default=llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED) - rope_freq_base: float = Field(default=0.0, description="RoPE base frequency") - rope_freq_scale: float = Field( - default=0.0, description="RoPE frequency scaling factor" - ) - yarn_ext_factor: float = Field(default=-1.0) - yarn_attn_factor: float = Field(default=1.0) - yarn_beta_fast: float = Field(default=32.0) - yarn_beta_slow: float = Field(default=1.0) - yarn_orig_ctx: int = Field(default=0) - mul_mat_q: bool = Field( - default=True, description="if true, use experimental mul_mat_q kernels" - ) - logits_all: bool = Field(default=True, description="Whether to return logits.") - embedding: bool = Field(default=True, description="Whether to use embeddings.") - offload_kqv: bool = Field( - default=False, description="Whether to offload kqv to the GPU." - ) - # Sampling Params - last_n_tokens_size: int = Field( - default=64, - ge=0, - description="Last n tokens to keep for repeat penalty calculation.", - ) - # LoRA Params - lora_base: Optional[str] = Field( - default=None, - description="Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model.", - ) - lora_path: Optional[str] = Field( - default=None, - description="Path to a LoRA file to apply to the model.", - ) - # Backend Params - numa: bool = Field( - default=False, - description="Enable NUMA support.", - ) - # Chat Format Params - chat_format: str = Field( - default="llama-2", - description="Chat format to use.", - ) - clip_model_path: Optional[str] = Field( - default=None, - description="Path to a CLIP model to use for multi-modal chat completion.", - ) - # Cache Params - cache: bool = Field( - default=False, - description="Use a cache to reduce processing times for evaluated prompts.", - ) - cache_type: Literal["ram", "disk"] = Field( - default="ram", - description="The type of cache to use. Only used if cache is True.", - ) - cache_size: int = Field( - default=2 << 30, - description="The size of the cache in bytes. Only used if cache is True.", - ) - # Misc - verbose: bool = Field( - default=True, description="Whether to print debug information." - ) - # Server Params - host: str = Field(default="localhost", description="Listen address") - port: int = Field(default=8000, description="Listen port") - # SSL Params - ssl_keyfile: Optional[str] = Field( - default=None, description="SSL key file for HTTPS" - ) - ssl_certfile: Optional[str] = Field( - default=None, description="SSL certificate file for HTTPS" - ) - interrupt_requests: bool = Field( - default=True, - description="Whether to interrupt requests when a new request is received.", - ) +_server_settings: Optional[ServerSettings] = None -class ErrorResponse(TypedDict): - """OpenAI style error response""" - - message: str - type: str - param: Optional[str] - code: Optional[str] - - -class ErrorResponseFormatters: - """Collection of formatters for error responses. - - Args: - request (Union[CreateCompletionRequest, CreateChatCompletionRequest]): - Request body - match (Match[str]): Match object from regex pattern - - Returns: - Tuple[int, ErrorResponse]: Status code and error response - """ - - @staticmethod - def context_length_exceeded( - request: Union["CreateCompletionRequest", "CreateChatCompletionRequest"], - match, # type: Match[str] # type: ignore - ) -> Tuple[int, ErrorResponse]: - """Formatter for context length exceeded error""" - - context_window = int(match.group(2)) - prompt_tokens = int(match.group(1)) - completion_tokens = request.max_tokens - if hasattr(request, "messages"): - # Chat completion - message = ( - "This model's maximum context length is {} tokens. " - "However, you requested {} tokens " - "({} in the messages, {} in the completion). " - "Please reduce the length of the messages or completion." - ) - else: - # Text completion - message = ( - "This model's maximum context length is {} tokens, " - "however you requested {} tokens " - "({} in your prompt; {} for the completion). " - "Please reduce your prompt; or completion length." - ) - return 400, ErrorResponse( - message=message.format( - context_window, - completion_tokens + prompt_tokens, - prompt_tokens, - completion_tokens, - ), - type="invalid_request_error", - param="messages", - code="context_length_exceeded", - ) +def set_server_settings(server_settings: ServerSettings): + global _server_settings + _server_settings = server_settings - @staticmethod - def model_not_found( - request: Union["CreateCompletionRequest", "CreateChatCompletionRequest"], - match, # type: Match[str] # type: ignore - ) -> Tuple[int, ErrorResponse]: - """Formatter for model_not_found error""" - - model_path = str(match.group(1)) - message = f"The model `{model_path}` does not exist" - return 400, ErrorResponse( - message=message, - type="invalid_request_error", - param=None, - code="model_not_found", - ) +def get_server_settings(): + yield _server_settings -class RouteErrorHandler(APIRoute): - """Custom APIRoute that handles application errors and exceptions""" - # key: regex pattern for original error message from llama_cpp - # value: formatter function - pattern_and_formatters: Dict[ - "Pattern", - Callable[ - [ - Union["CreateCompletionRequest", "CreateChatCompletionRequest"], - "Match[str]", - ], - Tuple[int, ErrorResponse], - ], - ] = { - compile( - r"Requested tokens \((\d+)\) exceed context window of (\d+)" - ): ErrorResponseFormatters.context_length_exceeded, - compile( - r"Model path does not exist: (.+)" - ): ErrorResponseFormatters.model_not_found, - } +_llama_proxy: Optional[LlamaProxy] = None - def error_message_wrapper( - self, - error: Exception, - body: Optional[ - Union[ - "CreateChatCompletionRequest", - "CreateCompletionRequest", - "CreateEmbeddingRequest", - ] - ] = None, - ) -> Tuple[int, ErrorResponse]: - """Wraps error message in OpenAI style error response""" - print(f"Exception: {str(error)}", file=sys.stderr) - traceback.print_exc(file=sys.stderr) - if body is not None and isinstance( - body, - ( - CreateCompletionRequest, - CreateChatCompletionRequest, - ), - ): - # When text completion or chat completion - for pattern, callback in self.pattern_and_formatters.items(): - match = pattern.search(str(error)) - if match is not None: - return callback(body, match) - - # Wrap other errors as internal server error - return 500, ErrorResponse( - message=str(error), - type="internal_server_error", - param=None, - code=None, - ) +llama_outer_lock = Lock() +llama_inner_lock = Lock() - def get_route_handler( - self, - ) -> Callable[[Request], Coroutine[None, None, Response]]: - """Defines custom route handler that catches exceptions and formats - in OpenAI style error response""" - original_route_handler = super().get_route_handler() +def set_llama_proxy(model_settings: List[ModelSettings]): + global _llama_proxy + _llama_proxy = LlamaProxy(models=model_settings) - async def custom_route_handler(request: Request) -> Response: - try: - start_sec = time.perf_counter() - response = await original_route_handler(request) - elapsed_time_ms = int((time.perf_counter() - start_sec) * 1000) - response.headers["openai-processing-ms"] = f"{elapsed_time_ms}" - return response - except Exception as exc: - json_body = await request.json() - try: - if "messages" in json_body: - # Chat completion - body: Optional[ - Union[ - CreateChatCompletionRequest, - CreateCompletionRequest, - CreateEmbeddingRequest, - ] - ] = CreateChatCompletionRequest(**json_body) - elif "prompt" in json_body: - # Text completion - body = CreateCompletionRequest(**json_body) - else: - # Embedding - body = CreateEmbeddingRequest(**json_body) - except Exception: - # Invalid request body - body = None - - # Get proper error message from the exception - ( - status_code, - error_message, - ) = self.error_message_wrapper(error=exc, body=body) - return JSONResponse( - {"error": error_message}, - status_code=status_code, - ) - return custom_route_handler +async def get_llama_proxy(): + # NOTE: This double lock allows the currently streaming llama model to + # check if any other requests are pending in the same thread and cancel + # the stream if so. + await llama_outer_lock.acquire() + release_outer_lock = True + try: + await llama_inner_lock.acquire() + try: + llama_outer_lock.release() + release_outer_lock = False + yield _llama_proxy + finally: + llama_inner_lock.release() + finally: + if release_outer_lock: + llama_outer_lock.release() -router = APIRouter(route_class=RouteErrorHandler) +_ping_message_factory: typing.Optional[typing.Callable[[], bytes]] = None -settings: Optional[Settings] = None -llama: Optional[llama_cpp.Llama] = None +def set_ping_message_factory(factory: typing.Callable[[], bytes]): + global _ping_message_factory + _ping_message_factory = factory -def create_app(settings: Optional[Settings] = None): - if settings is None: - settings = Settings() - middleware = [ - Middleware(RawContextMiddleware, plugins=(plugins.RequestIdPlugin(),)) - ] +def create_app( + settings: Settings | None = None, + server_settings: ServerSettings | None = None, + model_settings: List[ModelSettings] | None = None, +): + config_file = os.environ.get("CONFIG_FILE", None) + if config_file is not None: + if not os.path.exists(config_file): + raise ValueError(f"Config file {config_file} not found!") + with open(config_file, "rb") as f: + # Check if yaml file + if config_file.endswith(".yaml") or config_file.endswith(".yml"): + import yaml + + config_file_settings = ConfigFileSettings.model_validate_json( + json.dumps(yaml.safe_load(f)) + ) + else: + config_file_settings = ConfigFileSettings.model_validate_json(f.read()) + server_settings = ServerSettings.model_validate(config_file_settings) + model_settings = config_file_settings.models + + if server_settings is None and model_settings is None: + if settings is None: + settings = Settings() + server_settings = ServerSettings.model_validate(settings) + model_settings = [ModelSettings.model_validate(settings)] + + assert ( + server_settings is not None and model_settings is not None + ), "server_settings and model_settings must be provided together" + + set_server_settings(server_settings) + middleware = [Middleware(RawContextMiddleware, plugins=(RequestIdPlugin(),))] app = FastAPI( middleware=middleware, title="🦙 llama.cpp Python API", - version="0.0.1", + version=llama_cpp.__version__, + root_path=server_settings.root_path, ) app.add_middleware( CORSMiddleware, @@ -375,275 +145,81 @@ def create_app(settings: Optional[Settings] = None): allow_headers=["*"], ) app.include_router(router) - global llama - - ## - chat_handler = None - if settings.chat_format == "llava-1-5": - assert settings.clip_model_path is not None - chat_handler = llama_cpp.llama_chat_format.Llava15ChatHandler( - clip_model_path=settings.clip_model_path, verbose=settings.verbose - ) - ## - - llama = llama_cpp.Llama( - model_path=settings.model, - # Model Params - n_gpu_layers=settings.n_gpu_layers, - main_gpu=settings.main_gpu, - tensor_split=settings.tensor_split, - vocab_only=settings.vocab_only, - use_mmap=settings.use_mmap, - use_mlock=settings.use_mlock, - # Context Params - seed=settings.seed, - n_ctx=settings.n_ctx, - n_batch=settings.n_batch, - n_threads=settings.n_threads, - n_threads_batch=settings.n_threads_batch, - rope_scaling_type=settings.rope_scaling_type, - rope_freq_base=settings.rope_freq_base, - rope_freq_scale=settings.rope_freq_scale, - yarn_ext_factor=settings.yarn_ext_factor, - yarn_attn_factor=settings.yarn_attn_factor, - yarn_beta_fast=settings.yarn_beta_fast, - yarn_beta_slow=settings.yarn_beta_slow, - yarn_orig_ctx=settings.yarn_orig_ctx, - mul_mat_q=settings.mul_mat_q, - logits_all=settings.logits_all, - embedding=settings.embedding, - offload_kqv=settings.offload_kqv, - # Sampling Params - last_n_tokens_size=settings.last_n_tokens_size, - # LoRA Params - lora_base=settings.lora_base, - lora_path=settings.lora_path, - # Backend Params - numa=settings.numa, - # Chat Format Params - chat_format=settings.chat_format, - chat_handler=chat_handler, - # Misc - verbose=settings.verbose, - ) - if settings.cache: - if settings.cache_type == "disk": - if settings.verbose: - print(f"Using disk cache with size {settings.cache_size}") - cache = llama_cpp.LlamaDiskCache(capacity_bytes=settings.cache_size) - else: - if settings.verbose: - print(f"Using ram cache with size {settings.cache_size}") - cache = llama_cpp.LlamaRAMCache(capacity_bytes=settings.cache_size) - cache = llama_cpp.LlamaCache(capacity_bytes=settings.cache_size) - llama.set_cache(cache) + assert model_settings is not None + set_llama_proxy(model_settings=model_settings) - def set_settings(_settings: Settings): - global settings - settings = _settings + if server_settings.disable_ping_events: + set_ping_message_factory(lambda: bytes()) - set_settings(settings) return app -llama_outer_lock = Lock() -llama_inner_lock = Lock() - - -def get_llama(): - # NOTE: This double lock allows the currently streaming llama model to - # check if any other requests are pending in the same thread and cancel - # the stream if so. - llama_outer_lock.acquire() - release_outer_lock = True - try: - llama_inner_lock.acquire() - try: - llama_outer_lock.release() - release_outer_lock = False - yield llama - finally: - llama_inner_lock.release() - finally: - if release_outer_lock: - llama_outer_lock.release() +def prepare_request_resources( + body: CreateCompletionRequest | CreateChatCompletionRequest, + llama_proxy: LlamaProxy, + body_model: str | None, + kwargs, +) -> llama_cpp.Llama: + if llama_proxy is None: + raise HTTPException( + status_code=status.HTTP_503_SERVICE_UNAVAILABLE, + detail="Service is not available", + ) + llama = llama_proxy(body_model) + if body.logit_bias is not None: + kwargs["logit_bias"] = ( + _logit_bias_tokens_to_input_ids(llama, body.logit_bias) + if body.logit_bias_type == "tokens" + else body.logit_bias + ) + if body.grammar is not None: + kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar) -def get_settings(): - yield settings + if body.min_tokens > 0: + _min_tokens_logits_processor = llama_cpp.LogitsProcessorList( + [llama_cpp.MinTokensLogitsProcessor(body.min_tokens, llama.token_eos())] + ) + if "logits_processor" not in kwargs: + kwargs["logits_processor"] = _min_tokens_logits_processor + else: + kwargs["logits_processor"].extend(_min_tokens_logits_processor) + return llama async def get_event_publisher( request: Request, - inner_send_chan: MemoryObjectSendStream, - iterator: Iterator, + inner_send_chan: MemoryObjectSendStream[typing.Any], + body: CreateCompletionRequest | CreateChatCompletionRequest, + body_model: str | None, + llama_call, + kwargs, ): - async with inner_send_chan: - try: - async for chunk in iterate_in_threadpool(iterator): - await inner_send_chan.send(dict(data=json.dumps(chunk))) - if await request.is_disconnected(): - raise anyio.get_cancelled_exc_class()() - if settings.interrupt_requests and llama_outer_lock.locked(): - await inner_send_chan.send(dict(data="[DONE]")) - raise anyio.get_cancelled_exc_class()() - await inner_send_chan.send(dict(data="[DONE]")) - except anyio.get_cancelled_exc_class() as e: - print("disconnected") - with anyio.move_on_after(1, shield=True): - print(f"Disconnected from client (via refresh/close) {request.client}") - raise e - - -model_field = Field( - description="The model to use for generating completions.", default=None -) - -max_tokens_field = Field( - default=16, ge=1, description="The maximum number of tokens to generate." -) - -temperature_field = Field( - default=0.8, - ge=0.0, - le=2.0, - description="Adjust the randomness of the generated text.\n\n" - + "Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.", -) - -top_p_field = Field( - default=0.95, - ge=0.0, - le=1.0, - description="Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P.\n\n" - + "Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top_p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text.", -) - -min_p_field = Field( - default=0.05, - ge=0.0, - le=1.0, - description="Sets a minimum base probability threshold for token selection.\n\n" - + "The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter min_p represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with min_p=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out.", -) - -stop_field = Field( - default=None, - description="A list of tokens at which to stop generation. If None, no stop tokens are used.", -) - -stream_field = Field( - default=False, - description="Whether to stream the results as they are generated. Useful for chatbots.", -) - -top_k_field = Field( - default=40, - ge=0, - description="Limit the next token selection to the K most probable tokens.\n\n" - + "Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top_k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text.", -) - -repeat_penalty_field = Field( - default=1.1, - ge=0.0, - description="A penalty applied to each token that is already generated. This helps prevent the model from repeating itself.\n\n" - + "Repeat penalty is a hyperparameter used to penalize the repetition of token sequences during text generation. It helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient.", -) - -presence_penalty_field = Field( - default=0.0, - ge=-2.0, - le=2.0, - description="Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics.", -) - -frequency_penalty_field = Field( - default=0.0, - ge=-2.0, - le=2.0, - description="Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim.", -) - -mirostat_mode_field = Field( - default=0, - ge=0, - le=2, - description="Enable Mirostat constant-perplexity algorithm of the specified version (1 or 2; 0 = disabled)", -) - -mirostat_tau_field = Field( - default=5.0, - ge=0.0, - le=10.0, - description="Mirostat target entropy, i.e. the target perplexity - lower values produce focused and coherent text, larger values produce more diverse and less coherent text", -) - -mirostat_eta_field = Field( - default=0.1, ge=0.001, le=1.0, description="Mirostat learning rate" -) - -grammar = Field( - default=None, - description="A CBNF grammar (as string) to be used for formatting the model's output.", -) - - -class CreateCompletionRequest(BaseModel): - prompt: Union[str, List[str]] = Field( - default="", description="The prompt to generate completions for." - ) - suffix: Optional[str] = Field( - default=None, - description="A suffix to append to the generated text. If None, no suffix is appended. Useful for chatbots.", + server_settings = next(get_server_settings()) + interrupt_requests = ( + server_settings.interrupt_requests if server_settings else False ) - max_tokens: int = max_tokens_field - temperature: float = temperature_field - top_p: float = top_p_field - min_p: float = min_p_field - echo: bool = Field( - default=False, - description="Whether to echo the prompt in the generated text. Useful for chatbots.", - ) - stop: Optional[Union[str, List[str]]] = stop_field - stream: bool = stream_field - logprobs: Optional[int] = Field( - default=None, - ge=0, - description="The number of logprobs to generate. If None, no logprobs are generated.", - ) - presence_penalty: Optional[float] = presence_penalty_field - frequency_penalty: Optional[float] = frequency_penalty_field - logit_bias: Optional[Dict[str, float]] = Field(None) - logprobs: Optional[int] = Field(None) - seed: Optional[int] = Field(None) - - # ignored or currently unsupported - model: Optional[str] = model_field - n: Optional[int] = 1 - best_of: Optional[int] = 1 - user: Optional[str] = Field(default=None) - - # llama.cpp specific parameters - top_k: int = top_k_field - repeat_penalty: float = repeat_penalty_field - logit_bias_type: Optional[Literal["input_ids", "tokens"]] = Field(None) - mirostat_mode: int = mirostat_mode_field - mirostat_tau: float = mirostat_tau_field - mirostat_eta: float = mirostat_eta_field - grammar: Optional[str] = None - - model_config = { - "json_schema_extra": { - "examples": [ - { - "prompt": "\n\n### Instructions:\nWhat is the capital of France?\n\n### Response:\n", - "stop": ["\n", "###"], - } - ] - } - } + async with contextlib.asynccontextmanager(get_llama_proxy)() as llama_proxy: + llama = prepare_request_resources(body, llama_proxy, body_model, kwargs) + async with inner_send_chan: + try: + iterator = await run_in_threadpool(llama_call, llama, **kwargs) + async for chunk in iterate_in_threadpool(iterator): + await inner_send_chan.send(dict(data=json.dumps(chunk))) + if await request.is_disconnected(): + raise anyio.get_cancelled_exc_class()() + if interrupt_requests and llama_outer_lock.locked(): + await inner_send_chan.send(dict(data="[DONE]")) + raise anyio.get_cancelled_exc_class()() + await inner_send_chan.send(dict(data="[DONE]")) + except anyio.get_cancelled_exc_class() as e: + print("disconnected") + with anyio.move_on_after(1, shield=True): + print( + f"Disconnected from client (via refresh/close) {request.client}" + ) + raise e def _logit_bias_tokens_to_input_ids( @@ -658,53 +234,97 @@ def _logit_bias_tokens_to_input_ids( return to_bias +# Setup Bearer authentication scheme +bearer_scheme = HTTPBearer(auto_error=False) + + +async def authenticate( + settings: Settings = Depends(get_server_settings), + authorization: Optional[str] = Depends(bearer_scheme), +): + # Skip API key check if it's not set in settings + if settings.api_key is None: + return True + + # check bearer credentials against the api_key + if authorization and authorization.credentials == settings.api_key: + # api key is valid + return authorization.credentials + + # raise http error 401 + raise HTTPException( + status_code=status.HTTP_401_UNAUTHORIZED, + detail="Invalid API key", + ) + + +openai_v1_tag = "OpenAI V1" + + @router.post( "/v1/completions", - summary="Completion" + summary="Completion", + dependencies=[Depends(authenticate)], + response_model=Union[ + llama_cpp.CreateCompletionResponse, + str, + ], + responses={ + "200": { + "description": "Successful Response", + "content": { + "application/json": { + "schema": { + "anyOf": [ + {"$ref": "#/components/schemas/CreateCompletionResponse"} + ], + "title": "Completion response, when stream=False", + } + }, + "text/event-stream": { + "schema": { + "type": "string", + "title": "Server Side Streaming response, when stream=True. " + + "See SSE format: https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#Event_stream_format", # noqa: E501 + "example": """data: {... see CreateCompletionResponse ...} \\n\\n data: ... \\n\\n ... data: [DONE]""", + } + }, + }, + } + }, + tags=[openai_v1_tag], +) +@router.post( + "/v1/engines/copilot-codex/completions", + include_in_schema=False, + dependencies=[Depends(authenticate)], + tags=[openai_v1_tag], ) -@router.post("/v1/engines/copilot-codex/completions", include_in_schema=False) async def create_completion( request: Request, body: CreateCompletionRequest, - llama: llama_cpp.Llama = Depends(get_llama), ) -> llama_cpp.Completion: if isinstance(body.prompt, list): assert len(body.prompt) <= 1 body.prompt = body.prompt[0] if len(body.prompt) > 0 else "" + body_model = ( + body.model + if request.url.path != "/v1/engines/copilot-codex/completions" + else "copilot-codex" + ) + exclude = { "n", "best_of", "logit_bias_type", "user", + "min_tokens", } kwargs = body.model_dump(exclude=exclude) - if body.logit_bias is not None: - kwargs["logit_bias"] = ( - _logit_bias_tokens_to_input_ids(llama, body.logit_bias) - if body.logit_bias_type == "tokens" - else body.logit_bias - ) - - if body.grammar is not None: - kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar) - - iterator_or_completion: Union[ - llama_cpp.CreateCompletionResponse, - Iterator[llama_cpp.CreateCompletionStreamResponse], - ] = await run_in_threadpool(llama, **kwargs) - - if isinstance(iterator_or_completion, Iterator): - # EAFP: It's easier to ask for forgiveness than permission - first_response = await run_in_threadpool(next, iterator_or_completion) - - # If no exception was raised from first_response, we can assume that - # the iterator is valid and we can use it to stream the response. - def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: - yield first_response - yield from iterator_or_completion - + # handle streaming request + if kwargs.get("stream", False): send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( recv_chan, @@ -712,160 +332,168 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: get_event_publisher, request=request, inner_send_chan=send_chan, - iterator=iterator(), + body=body, + body_model=body_model, + llama_call=llama_cpp.Llama.__call__, + kwargs=kwargs, ), + sep="\n", + ping_message_factory=_ping_message_factory, ) - else: - return iterator_or_completion - - -class CreateEmbeddingRequest(BaseModel): - model: Optional[str] = model_field - input: Union[str, List[str]] = Field(description="The input to embed.") - user: Optional[str] = Field(default=None) - - model_config = { - "json_schema_extra": { - "examples": [ - { - "input": "The food was delicious and the waiter...", - } - ] - } - } + + # handle regular request + async with contextlib.asynccontextmanager(get_llama_proxy)() as llama_proxy: + llama = prepare_request_resources(body, llama_proxy, body_model, kwargs) + + if await request.is_disconnected(): + print( + f"Disconnected from client (via refresh/close) before llm invoked {request.client}" + ) + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail="Client closed request", + ) + + return await run_in_threadpool(llama, **kwargs) @router.post( "/v1/embeddings", - summary="Embedding" + summary="Embedding", + dependencies=[Depends(authenticate)], + tags=[openai_v1_tag], ) async def create_embedding( - request: CreateEmbeddingRequest, llama: llama_cpp.Llama = Depends(get_llama) + request: CreateEmbeddingRequest, + llama_proxy: LlamaProxy = Depends(get_llama_proxy), ): return await run_in_threadpool( - llama.create_embedding, **request.model_dump(exclude={"user"}) + llama_proxy(request.model).create_embedding, + **request.model_dump(exclude={"user"}), ) -class ChatCompletionRequestMessage(BaseModel): - role: Literal["system", "user", "assistant", "function"] = Field( - default="user", description="The role of the message." - ) - content: Optional[str] = Field( - default="", description="The content of the message." - ) - - -class CreateChatCompletionRequest(BaseModel): - messages: List[llama_cpp.ChatCompletionRequestMessage] = Field( - default=[], description="A list of messages to generate completions for." - ) - functions: Optional[List[llama_cpp.ChatCompletionFunction]] = Field( - default=None, - description="A list of functions to apply to the generated completions.", - ) - function_call: Optional[llama_cpp.ChatCompletionRequestFunctionCall] = Field( - default=None, - description="A function to apply to the generated completions.", - ) - tools: Optional[List[llama_cpp.ChatCompletionTool]] = Field( - default=None, - description="A list of tools to apply to the generated completions.", - ) - tool_choice: Optional[llama_cpp.ChatCompletionToolChoiceOption] = Field( - default=None, - description="A tool to apply to the generated completions.", - ) # TODO: verify - max_tokens: Optional[int] = Field( - default=None, - description="The maximum number of tokens to generate. Defaults to inf", - ) - temperature: float = temperature_field - top_p: float = top_p_field - min_p: float = min_p_field - stop: Optional[Union[str, List[str]]] = stop_field - stream: bool = stream_field - presence_penalty: Optional[float] = presence_penalty_field - frequency_penalty: Optional[float] = frequency_penalty_field - logit_bias: Optional[Dict[str, float]] = Field(None) - seed: Optional[int] = Field(None) - response_format: Optional[llama_cpp.ChatCompletionRequestResponseFormat] = Field( - default=None, - ) - - # ignored or currently unsupported - model: Optional[str] = model_field - n: Optional[int] = 1 - user: Optional[str] = Field(None) - - # llama.cpp specific parameters - top_k: int = top_k_field - repeat_penalty: float = repeat_penalty_field - logit_bias_type: Optional[Literal["input_ids", "tokens"]] = Field(None) - mirostat_mode: int = mirostat_mode_field - mirostat_tau: float = mirostat_tau_field - mirostat_eta: float = mirostat_eta_field - grammar: Optional[str] = None - - model_config = { - "json_schema_extra": { - "examples": [ - { - "messages": [ - ChatCompletionRequestMessage( - role="system", content="You are a helpful assistant." - ).model_dump(), - ChatCompletionRequestMessage( - role="user", content="What is the capital of France?" - ).model_dump(), - ] - } - ] - } - } - - @router.post( "/v1/chat/completions", - summary="Chat" + summary="Chat", + dependencies=[Depends(authenticate)], + response_model=Union[llama_cpp.ChatCompletion, str], + responses={ + "200": { + "description": "Successful Response", + "content": { + "application/json": { + "schema": { + "anyOf": [ + { + "$ref": "#/components/schemas/CreateChatCompletionResponse" + } + ], + "title": "Completion response, when stream=False", + } + }, + "text/event-stream": { + "schema": { + "type": "string", + "title": "Server Side Streaming response, when stream=True" + + "See SSE format: https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#Event_stream_format", # noqa: E501 + "example": """data: {... see CreateChatCompletionResponse ...} \\n\\n data: ... \\n\\n ... data: [DONE]""", + } + }, + }, + } + }, + tags=[openai_v1_tag], ) async def create_chat_completion( request: Request, - body: CreateChatCompletionRequest, - llama: llama_cpp.Llama = Depends(get_llama), - settings: Settings = Depends(get_settings), + body: CreateChatCompletionRequest = Body( + openapi_examples={ + "normal": { + "summary": "Chat Completion", + "value": { + "model": "gpt-3.5-turbo", + "messages": [ + {"role": "system", "content": "You are a helpful assistant."}, + {"role": "user", "content": "What is the capital of France?"}, + ], + }, + }, + "json_mode": { + "summary": "JSON Mode", + "value": { + "model": "gpt-3.5-turbo", + "messages": [ + {"role": "system", "content": "You are a helpful assistant."}, + {"role": "user", "content": "Who won the world series in 2020"}, + ], + "response_format": {"type": "json_object"}, + }, + }, + "tool_calling": { + "summary": "Tool Calling", + "value": { + "model": "gpt-3.5-turbo", + "messages": [ + {"role": "system", "content": "You are a helpful assistant."}, + {"role": "user", "content": "Extract Jason is 30 years old."}, + ], + "tools": [ + { + "type": "function", + "function": { + "name": "User", + "description": "User record", + "parameters": { + "type": "object", + "properties": { + "name": {"type": "string"}, + "age": {"type": "number"}, + }, + "required": ["name", "age"], + }, + }, + } + ], + "tool_choice": { + "type": "function", + "function": { + "name": "User", + }, + }, + }, + }, + "logprobs": { + "summary": "Logprobs", + "value": { + "model": "gpt-3.5-turbo", + "messages": [ + {"role": "system", "content": "You are a helpful assistant."}, + {"role": "user", "content": "What is the capital of France?"}, + ], + "logprobs": True, + "top_logprobs": 10, + }, + }, + } + ), ) -> llama_cpp.ChatCompletion: + # This is a workaround for an issue in FastAPI dependencies + # where the dependency is cleaned up before a StreamingResponse + # is complete. + # https://github.com/tiangolo/fastapi/issues/11143 + + body_model = body.model exclude = { "n", "logit_bias_type", "user", + "min_tokens", } kwargs = body.model_dump(exclude=exclude) - if body.logit_bias is not None: - kwargs["logit_bias"] = ( - _logit_bias_tokens_to_input_ids(llama, body.logit_bias) - if body.logit_bias_type == "tokens" - else body.logit_bias - ) - - if body.grammar is not None: - kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar) - - iterator_or_completion: Union[ - llama_cpp.ChatCompletion, Iterator[llama_cpp.ChatCompletionChunk] - ] = await run_in_threadpool(llama.create_chat_completion, **kwargs) - - if isinstance(iterator_or_completion, Iterator): - # EAFP: It's easier to ask for forgiveness than permission - first_response = await run_in_threadpool(next, iterator_or_completion) - - # If no exception was raised from first_response, we can assume that - # the iterator is valid and we can use it to stream the response. - def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: - yield first_response - yield from iterator_or_completion - + # handle streaming request + if kwargs.get("stream", False): send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( recv_chan, @@ -873,40 +501,97 @@ def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: get_event_publisher, request=request, inner_send_chan=send_chan, - iterator=iterator(), + body=body, + body_model=body_model, + llama_call=llama_cpp.Llama.create_chat_completion, + kwargs=kwargs, ), + sep="\n", + ping_message_factory=_ping_message_factory, ) - else: - return iterator_or_completion + # handle regular request + async with contextlib.asynccontextmanager(get_llama_proxy)() as llama_proxy: + llama = prepare_request_resources(body, llama_proxy, body_model, kwargs) -class ModelData(TypedDict): - id: str - object: Literal["model"] - owned_by: str - permissions: List[str] - + if await request.is_disconnected(): + print( + f"Disconnected from client (via refresh/close) before llm invoked {request.client}" + ) + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail="Client closed request", + ) -class ModelList(TypedDict): - object: Literal["list"] - data: List[ModelData] + return await run_in_threadpool(llama.create_chat_completion, **kwargs) -@router.get("/v1/models", summary="Models") +@router.get( + "/v1/models", + summary="Models", + dependencies=[Depends(authenticate)], + tags=[openai_v1_tag], +) async def get_models( - settings: Settings = Depends(get_settings), + llama_proxy: LlamaProxy = Depends(get_llama_proxy), ) -> ModelList: - assert llama is not None return { "object": "list", "data": [ { - "id": settings.model_alias - if settings.model_alias is not None - else llama.model_path, + "id": model_alias, "object": "model", "owned_by": "me", "permissions": [], } + for model_alias in llama_proxy ], } + + +extras_tag = "Extras" + + +@router.post( + "/extras/tokenize", + summary="Tokenize", + dependencies=[Depends(authenticate)], + tags=[extras_tag], +) +async def tokenize( + body: TokenizeInputRequest, + llama_proxy: LlamaProxy = Depends(get_llama_proxy), +) -> TokenizeInputResponse: + tokens = llama_proxy(body.model).tokenize(body.input.encode("utf-8"), special=True) + + return TokenizeInputResponse(tokens=tokens) + + +@router.post( + "/extras/tokenize/count", + summary="Tokenize Count", + dependencies=[Depends(authenticate)], + tags=[extras_tag], +) +async def count_query_tokens( + body: TokenizeInputRequest, + llama_proxy: LlamaProxy = Depends(get_llama_proxy), +) -> TokenizeInputCountResponse: + tokens = llama_proxy(body.model).tokenize(body.input.encode("utf-8"), special=True) + + return TokenizeInputCountResponse(count=len(tokens)) + + +@router.post( + "/extras/detokenize", + summary="Detokenize", + dependencies=[Depends(authenticate)], + tags=[extras_tag], +) +async def detokenize( + body: DetokenizeInputRequest, + llama_proxy: LlamaProxy = Depends(get_llama_proxy), +) -> DetokenizeInputResponse: + text = llama_proxy(body.model).detokenize(body.tokens).decode("utf-8") + + return DetokenizeInputResponse(text=text) diff --git a/llama_cpp/server/cli.py b/llama_cpp/server/cli.py new file mode 100644 index 000000000..3dd007676 --- /dev/null +++ b/llama_cpp/server/cli.py @@ -0,0 +1,97 @@ +from __future__ import annotations + +import argparse + +from typing import List, Literal, Union, Any, Type, TypeVar + +from pydantic import BaseModel + + +def _get_base_type(annotation: Type[Any]) -> Type[Any]: + if getattr(annotation, "__origin__", None) is Literal: + assert hasattr(annotation, "__args__") and len(annotation.__args__) >= 1 # type: ignore + return type(annotation.__args__[0]) # type: ignore + elif getattr(annotation, "__origin__", None) is Union: + assert hasattr(annotation, "__args__") and len(annotation.__args__) >= 1 # type: ignore + non_optional_args: List[Type[Any]] = [ + arg for arg in annotation.__args__ if arg is not type(None) # type: ignore + ] + if non_optional_args: + return _get_base_type(non_optional_args[0]) + elif ( + getattr(annotation, "__origin__", None) is list + or getattr(annotation, "__origin__", None) is List + ): + assert hasattr(annotation, "__args__") and len(annotation.__args__) >= 1 # type: ignore + return _get_base_type(annotation.__args__[0]) # type: ignore + return annotation + + +def _contains_list_type(annotation: Type[Any] | None) -> bool: + origin = getattr(annotation, "__origin__", None) + + if origin is list or origin is List: + return True + elif origin in (Literal, Union): + return any(_contains_list_type(arg) for arg in annotation.__args__) # type: ignore + else: + return False + + +def _parse_bool_arg(arg: str | bytes | bool) -> bool: + if isinstance(arg, bytes): + arg = arg.decode("utf-8") + + true_values = {"1", "on", "t", "true", "y", "yes"} + false_values = {"0", "off", "f", "false", "n", "no"} + + arg_str = str(arg).lower().strip() + + if arg_str in true_values: + return True + elif arg_str in false_values: + return False + else: + raise ValueError(f"Invalid boolean argument: {arg}") + + +def add_args_from_model(parser: argparse.ArgumentParser, model: Type[BaseModel]): + """Add arguments from a pydantic model to an argparse parser.""" + + for name, field in model.model_fields.items(): + description = field.description + if field.default and description and not field.is_required(): + description += f" (default: {field.default})" + base_type = ( + _get_base_type(field.annotation) if field.annotation is not None else str + ) + list_type = _contains_list_type(field.annotation) + if base_type is not bool: + parser.add_argument( + f"--{name}", + dest=name, + nargs="*" if list_type else None, + type=base_type, + help=description, + ) + if base_type is bool: + parser.add_argument( + f"--{name}", + dest=name, + type=_parse_bool_arg, + help=f"{description}", + ) + + +T = TypeVar("T", bound=Type[BaseModel]) + + +def parse_model_from_args(model: T, args: argparse.Namespace) -> T: + """Parse a pydantic model from an argparse namespace.""" + return model( + **{ + k: v + for k, v in vars(args).items() + if v is not None and k in model.model_fields + } + ) diff --git a/llama_cpp/server/errors.py b/llama_cpp/server/errors.py new file mode 100644 index 000000000..d0eda5664 --- /dev/null +++ b/llama_cpp/server/errors.py @@ -0,0 +1,212 @@ +from __future__ import annotations + +import sys +import traceback +import time +from re import compile, Match, Pattern +from typing import Callable, Coroutine, Optional, Tuple, Union, Dict +from typing_extensions import TypedDict + + +from fastapi import ( + Request, + Response, + HTTPException, +) +from fastapi.responses import JSONResponse +from fastapi.routing import APIRoute + +from llama_cpp.server.types import ( + CreateCompletionRequest, + CreateEmbeddingRequest, + CreateChatCompletionRequest, +) + + +class ErrorResponse(TypedDict): + """OpenAI style error response""" + + message: str + type: str + param: Optional[str] + code: Optional[str] + + +class ErrorResponseFormatters: + """Collection of formatters for error responses. + + Args: + request (Union[CreateCompletionRequest, CreateChatCompletionRequest]): + Request body + match (Match[str]): Match object from regex pattern + + Returns: + Tuple[int, ErrorResponse]: Status code and error response + """ + + @staticmethod + def context_length_exceeded( + request: Union["CreateCompletionRequest", "CreateChatCompletionRequest"], + match, # type: Match[str] # type: ignore + ) -> Tuple[int, ErrorResponse]: + """Formatter for context length exceeded error""" + + context_window = int(match.group(2)) + prompt_tokens = int(match.group(1)) + completion_tokens = request.max_tokens + if hasattr(request, "messages"): + # Chat completion + message = ( + "This model's maximum context length is {} tokens. " + "However, you requested {} tokens " + "({} in the messages, {} in the completion). " + "Please reduce the length of the messages or completion." + ) + else: + # Text completion + message = ( + "This model's maximum context length is {} tokens, " + "however you requested {} tokens " + "({} in your prompt; {} for the completion). " + "Please reduce your prompt; or completion length." + ) + return 400, ErrorResponse( + message=message.format( + context_window, + (completion_tokens or 0) + prompt_tokens, + prompt_tokens, + completion_tokens, + ), # type: ignore + type="invalid_request_error", + param="messages", + code="context_length_exceeded", + ) + + @staticmethod + def model_not_found( + request: Union["CreateCompletionRequest", "CreateChatCompletionRequest"], + match, # type: Match[str] # type: ignore + ) -> Tuple[int, ErrorResponse]: + """Formatter for model_not_found error""" + + model_path = str(match.group(1)) + message = f"The model `{model_path}` does not exist" + return 400, ErrorResponse( + message=message, + type="invalid_request_error", + param=None, + code="model_not_found", + ) + + +class RouteErrorHandler(APIRoute): + """Custom APIRoute that handles application errors and exceptions""" + + # key: regex pattern for original error message from llama_cpp + # value: formatter function + pattern_and_formatters: Dict[ + "Pattern[str]", + Callable[ + [ + Union["CreateCompletionRequest", "CreateChatCompletionRequest"], + "Match[str]", + ], + Tuple[int, ErrorResponse], + ], + ] = { + compile( + r"Requested tokens \((\d+)\) exceed context window of (\d+)" + ): ErrorResponseFormatters.context_length_exceeded, + compile( + r"Model path does not exist: (.+)" + ): ErrorResponseFormatters.model_not_found, + } + + def error_message_wrapper( + self, + error: Exception, + body: Optional[ + Union[ + "CreateChatCompletionRequest", + "CreateCompletionRequest", + "CreateEmbeddingRequest", + ] + ] = None, + ) -> Tuple[int, ErrorResponse]: + """Wraps error message in OpenAI style error response""" + if body is not None and isinstance( + body, + ( + CreateCompletionRequest, + CreateChatCompletionRequest, + ), + ): + # When text completion or chat completion + for pattern, callback in self.pattern_and_formatters.items(): + match = pattern.search(str(error)) + if match is not None: + return callback(body, match) + + # Only print the trace on unexpected exceptions + print(f"Exception: {str(error)}", file=sys.stderr) + traceback.print_exc(file=sys.stderr) + + # Wrap other errors as internal server error + return 500, ErrorResponse( + message=str(error), + type="internal_server_error", + param=None, + code=None, + ) + + def get_route_handler( + self, + ) -> Callable[[Request], Coroutine[None, None, Response]]: + """Defines custom route handler that catches exceptions and formats + in OpenAI style error response""" + + original_route_handler = super().get_route_handler() + + async def custom_route_handler(request: Request) -> Response: + try: + start_sec = time.perf_counter() + response = await original_route_handler(request) + elapsed_time_ms = int((time.perf_counter() - start_sec) * 1000) + response.headers["openai-processing-ms"] = f"{elapsed_time_ms}" + return response + except HTTPException as unauthorized: + # api key check failed + raise unauthorized + except Exception as exc: + json_body = await request.json() + try: + if "messages" in json_body: + # Chat completion + body: Optional[ + Union[ + CreateChatCompletionRequest, + CreateCompletionRequest, + CreateEmbeddingRequest, + ] + ] = CreateChatCompletionRequest(**json_body) + elif "prompt" in json_body: + # Text completion + body = CreateCompletionRequest(**json_body) + else: + # Embedding + body = CreateEmbeddingRequest(**json_body) + except Exception: + # Invalid request body + body = None + + # Get proper error message from the exception + ( + status_code, + error_message, + ) = self.error_message_wrapper(error=exc, body=body) + return JSONResponse( + {"error": error_message}, + status_code=status_code, + ) + + return custom_route_handler diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py new file mode 100644 index 000000000..c6716f919 --- /dev/null +++ b/llama_cpp/server/model.py @@ -0,0 +1,298 @@ +from __future__ import annotations + +import json + +from typing import Dict, Optional, Union, List + +import llama_cpp +import llama_cpp.llama_speculative as llama_speculative +import llama_cpp.llama_tokenizer as llama_tokenizer + +from llama_cpp.server.settings import ModelSettings + + +class LlamaProxy: + def __init__(self, models: List[ModelSettings]) -> None: + assert len(models) > 0, "No models provided!" + + self._model_settings_dict: dict[str, ModelSettings] = {} + for model in models: + if not model.model_alias: + model.model_alias = model.model + self._model_settings_dict[model.model_alias] = model + + self._current_model: Optional[llama_cpp.Llama] = None + self._current_model_alias: Optional[str] = None + + self._default_model_settings: ModelSettings = models[0] + self._default_model_alias: str = self._default_model_settings.model_alias # type: ignore + + # Load default model + self._current_model = self.load_llama_from_model_settings( + self._default_model_settings + ) + self._current_model_alias = self._default_model_alias + + def __call__(self, model: Optional[str] = None) -> llama_cpp.Llama: + if model is None: + model = self._default_model_alias + + if model not in self._model_settings_dict: + model = self._default_model_alias + + if model == self._current_model_alias: + if self._current_model is not None: + return self._current_model + + if self._current_model: + self._current_model.close() + self._current_model = None + + settings = self._model_settings_dict[model] + self._current_model = self.load_llama_from_model_settings(settings) + self._current_model_alias = model + return self._current_model + + def __getitem__(self, model: str): + return self._model_settings_dict[model].model_dump() + + def __setitem__(self, model: str, settings: Union[ModelSettings, str, bytes]): + if isinstance(settings, (bytes, str)): + settings = ModelSettings.model_validate_json(settings) + self._model_settings_dict[model] = settings + + def __iter__(self): + for model in self._model_settings_dict: + yield model + + def free(self): + if self._current_model: + self._current_model.close() + del self._current_model + + @staticmethod + def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama: + chat_handler = None + if settings.chat_format == "llava-1-5": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.Llava15ChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.Llava15ChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "obsidian": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.ObsidianChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.ObsidianChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "llava-1-6": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.Llava16ChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.Llava16ChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "moondream": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.MoondreamChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.MoondreamChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "nanollava": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.NanoLlavaChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.NanoLlavaChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "llama-3-vision-alpha": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.Llama3VisionAlpha.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.Llama3VisionAlpha( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "minicpm-v-2.6": + assert settings.clip_model_path is not None, "clip model not found" + if settings.hf_model_repo_id is not None: + chat_handler = ( + llama_cpp.llama_chat_format.MiniCPMv26ChatHandler.from_pretrained( + repo_id=settings.hf_model_repo_id, + filename=settings.clip_model_path, + verbose=settings.verbose, + ) + ) + else: + chat_handler = llama_cpp.llama_chat_format.MiniCPMv26ChatHandler( + clip_model_path=settings.clip_model_path, verbose=settings.verbose + ) + elif settings.chat_format == "hf-autotokenizer": + assert ( + settings.hf_pretrained_model_name_or_path is not None + ), "hf_pretrained_model_name_or_path must be set for hf-autotokenizer" + chat_handler = ( + llama_cpp.llama_chat_format.hf_autotokenizer_to_chat_completion_handler( + settings.hf_pretrained_model_name_or_path + ) + ) + elif settings.chat_format == "hf-tokenizer-config": + assert ( + settings.hf_tokenizer_config_path is not None + ), "hf_tokenizer_config_path must be set for hf-tokenizer-config" + chat_handler = llama_cpp.llama_chat_format.hf_tokenizer_config_to_chat_completion_handler( + json.load(open(settings.hf_tokenizer_config_path)) + ) + + tokenizer: Optional[llama_cpp.BaseLlamaTokenizer] = None + if settings.hf_pretrained_model_name_or_path is not None: + tokenizer = llama_tokenizer.LlamaHFTokenizer.from_pretrained( + settings.hf_pretrained_model_name_or_path + ) + + draft_model = None + if settings.draft_model is not None: + draft_model = llama_speculative.LlamaPromptLookupDecoding( + num_pred_tokens=settings.draft_model_num_pred_tokens + ) + + kv_overrides: Optional[Dict[str, Union[bool, int, float, str]]] = None + if settings.kv_overrides is not None: + assert isinstance(settings.kv_overrides, list) + kv_overrides = {} + for kv in settings.kv_overrides: + key, value = kv.split("=") + if ":" in value: + value_type, value = value.split(":") + if value_type == "bool": + kv_overrides[key] = value.lower() in ["true", "1"] + elif value_type == "int": + kv_overrides[key] = int(value) + elif value_type == "float": + kv_overrides[key] = float(value) + elif value_type == "str": + kv_overrides[key] = value + else: + raise ValueError(f"Unknown value type {value_type}") + + import functools + + kwargs = {} + + if settings.hf_model_repo_id is not None: + create_fn = functools.partial( + llama_cpp.Llama.from_pretrained, + repo_id=settings.hf_model_repo_id, + filename=settings.model, + ) + else: + create_fn = llama_cpp.Llama + kwargs["model_path"] = settings.model + + _model = create_fn( + **kwargs, + # Model Params + n_gpu_layers=settings.n_gpu_layers, + split_mode=settings.split_mode, + main_gpu=settings.main_gpu, + tensor_split=settings.tensor_split, + vocab_only=settings.vocab_only, + use_mmap=settings.use_mmap, + use_mlock=settings.use_mlock, + kv_overrides=kv_overrides, + rpc_servers=settings.rpc_servers, + # Context Params + seed=settings.seed, + n_ctx=settings.n_ctx, + n_batch=settings.n_batch, + n_ubatch=settings.n_ubatch, + n_threads=settings.n_threads, + n_threads_batch=settings.n_threads_batch, + rope_scaling_type=settings.rope_scaling_type, + rope_freq_base=settings.rope_freq_base, + rope_freq_scale=settings.rope_freq_scale, + yarn_ext_factor=settings.yarn_ext_factor, + yarn_attn_factor=settings.yarn_attn_factor, + yarn_beta_fast=settings.yarn_beta_fast, + yarn_beta_slow=settings.yarn_beta_slow, + yarn_orig_ctx=settings.yarn_orig_ctx, + mul_mat_q=settings.mul_mat_q, + logits_all=settings.logits_all, + embedding=settings.embedding, + offload_kqv=settings.offload_kqv, + flash_attn=settings.flash_attn, + # Sampling Params + last_n_tokens_size=settings.last_n_tokens_size, + # LoRA Params + lora_base=settings.lora_base, + lora_path=settings.lora_path, + # Backend Params + numa=settings.numa, + # Chat Format Params + chat_format=settings.chat_format, + chat_handler=chat_handler, + # Speculative Decoding + draft_model=draft_model, + # KV Cache Quantization + type_k=settings.type_k, + type_v=settings.type_v, + # Tokenizer + tokenizer=tokenizer, + # Misc + verbose=settings.verbose, + ) + if settings.cache: + if settings.cache_type == "disk": + if settings.verbose: + print(f"Using disk cache with size {settings.cache_size}") + cache = llama_cpp.LlamaDiskCache(capacity_bytes=settings.cache_size) + else: + if settings.verbose: + print(f"Using ram cache with size {settings.cache_size}") + cache = llama_cpp.LlamaRAMCache(capacity_bytes=settings.cache_size) + _model.set_cache(cache) + return _model diff --git a/llama_cpp/server/settings.py b/llama_cpp/server/settings.py new file mode 100644 index 000000000..13c951241 --- /dev/null +++ b/llama_cpp/server/settings.py @@ -0,0 +1,240 @@ +from __future__ import annotations + +import multiprocessing + +from typing import Optional, List, Literal, Union, Dict, cast +from typing_extensions import Self + +from pydantic import Field, model_validator +from pydantic_settings import BaseSettings + +import llama_cpp + +# Disable warning for model and model_alias settings +BaseSettings.model_config["protected_namespaces"] = () + + +class ModelSettings(BaseSettings): + """Model settings used to load a Llama model.""" + + model: str = Field( + description="The path to the model to use for generating completions." + ) + model_alias: Optional[str] = Field( + default=None, + description="The alias of the model to use for generating completions.", + ) + # Model Params + n_gpu_layers: int = Field( + default=0, + ge=-1, + description="The number of layers to put on the GPU. The rest will be on the CPU. Set -1 to move all to GPU.", + ) + split_mode: int = Field( + default=llama_cpp.LLAMA_SPLIT_MODE_LAYER, + description="The split mode to use.", + ) + main_gpu: int = Field( + default=0, + ge=0, + description="Main GPU to use.", + ) + tensor_split: Optional[List[float]] = Field( + default=None, + description="Split layers across multiple GPUs in proportion.", + ) + vocab_only: bool = Field( + default=False, description="Whether to only return the vocabulary." + ) + use_mmap: bool = Field( + default=llama_cpp.llama_supports_mmap(), + description="Use mmap.", + ) + use_mlock: bool = Field( + default=llama_cpp.llama_supports_mlock(), + description="Use mlock.", + ) + kv_overrides: Optional[List[str]] = Field( + default=None, + description="List of model kv overrides in the format key=type:value where type is one of (bool, int, float). Valid true values are (true, TRUE, 1), otherwise false.", + ) + rpc_servers: Optional[str] = Field( + default=None, + description="comma seperated list of rpc servers for offloading", + ) + # Context Params + seed: int = Field( + default=llama_cpp.LLAMA_DEFAULT_SEED, description="Random seed. -1 for random." + ) + n_ctx: int = Field(default=2048, ge=0, description="The context size.") + n_batch: int = Field( + default=512, ge=1, description="The batch size to use per eval." + ) + n_ubatch: int = Field( + default=512, ge=1, description="The physical batch size used by llama.cpp" + ) + n_threads: int = Field( + default=max(multiprocessing.cpu_count() // 2, 1), + ge=1, + description="The number of threads to use. Use -1 for max cpu threads", + ) + n_threads_batch: int = Field( + default=max(multiprocessing.cpu_count(), 1), + ge=0, + description="The number of threads to use when batch processing. Use -1 for max cpu threads", + ) + rope_scaling_type: int = Field( + default=llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED + ) + rope_freq_base: float = Field(default=0.0, description="RoPE base frequency") + rope_freq_scale: float = Field( + default=0.0, description="RoPE frequency scaling factor" + ) + yarn_ext_factor: float = Field(default=-1.0) + yarn_attn_factor: float = Field(default=1.0) + yarn_beta_fast: float = Field(default=32.0) + yarn_beta_slow: float = Field(default=1.0) + yarn_orig_ctx: int = Field(default=0) + mul_mat_q: bool = Field( + default=True, description="if true, use experimental mul_mat_q kernels" + ) + logits_all: bool = Field(default=True, description="Whether to return logits.") + embedding: bool = Field(default=False, description="Whether to use embeddings.") + offload_kqv: bool = Field( + default=True, description="Whether to offload kqv to the GPU." + ) + flash_attn: bool = Field( + default=False, description="Whether to use flash attention." + ) + # Sampling Params + last_n_tokens_size: int = Field( + default=64, + ge=0, + description="Last n tokens to keep for repeat penalty calculation.", + ) + # LoRA Params + lora_base: Optional[str] = Field( + default=None, + description="Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model.", + ) + lora_path: Optional[str] = Field( + default=None, + description="Path to a LoRA file to apply to the model.", + ) + # Backend Params + numa: Union[bool, int] = Field( + default=False, + description="Enable NUMA support.", + ) + # Chat Format Params + chat_format: Optional[str] = Field( + default=None, + description="Chat format to use.", + ) + clip_model_path: Optional[str] = Field( + default=None, + description="Path to a CLIP model to use for multi-modal chat completion.", + ) + # Cache Params + cache: bool = Field( + default=False, + description="Use a cache to reduce processing times for evaluated prompts.", + ) + cache_type: Literal["ram", "disk"] = Field( + default="ram", + description="The type of cache to use. Only used if cache is True.", + ) + cache_size: int = Field( + default=2 << 30, + description="The size of the cache in bytes. Only used if cache is True.", + ) + # Tokenizer Options + hf_tokenizer_config_path: Optional[str] = Field( + default=None, + description="The path to a HuggingFace tokenizer_config.json file.", + ) + hf_pretrained_model_name_or_path: Optional[str] = Field( + default=None, + description="The model name or path to a pretrained HuggingFace tokenizer model. Same as you would pass to AutoTokenizer.from_pretrained().", + ) + # Loading from HuggingFace Model Hub + hf_model_repo_id: Optional[str] = Field( + default=None, + description="The model repo id to use for the HuggingFace tokenizer model.", + ) + # Speculative Decoding + draft_model: Optional[str] = Field( + default=None, + description="Method to use for speculative decoding. One of (prompt-lookup-decoding).", + ) + draft_model_num_pred_tokens: int = Field( + default=10, + description="Number of tokens to predict using the draft model.", + ) + # KV Cache Quantization + type_k: Optional[int] = Field( + default=None, + description="Type of the key cache quantization.", + ) + type_v: Optional[int] = Field( + default=None, + description="Type of the value cache quantization.", + ) + # Misc + verbose: bool = Field( + default=True, description="Whether to print debug information." + ) + + @model_validator( + mode="before" + ) # pre=True to ensure this runs before any other validation + def set_dynamic_defaults(self) -> Self: + # If n_threads or n_threads_batch is -1, set it to multiprocessing.cpu_count() + cpu_count = multiprocessing.cpu_count() + values = cast(Dict[str, int], self) + if values.get("n_threads", 0) == -1: + values["n_threads"] = cpu_count + if values.get("n_threads_batch", 0) == -1: + values["n_threads_batch"] = cpu_count + return self + + +class ServerSettings(BaseSettings): + """Server settings used to configure the FastAPI and Uvicorn server.""" + + # Uvicorn Settings + host: str = Field(default="localhost", description="Listen address") + port: int = Field(default=8000, description="Listen port") + ssl_keyfile: Optional[str] = Field( + default=None, description="SSL key file for HTTPS" + ) + ssl_certfile: Optional[str] = Field( + default=None, description="SSL certificate file for HTTPS" + ) + # FastAPI Settings + api_key: Optional[str] = Field( + default=None, + description="API key for authentication. If set all requests need to be authenticated.", + ) + interrupt_requests: bool = Field( + default=True, + description="Whether to interrupt requests when a new request is received.", + ) + disable_ping_events: bool = Field( + default=False, + description="Disable EventSource pings (may be needed for some clients).", + ) + root_path: str = Field( + default="", + description="The root path for the server. Useful when running behind a reverse proxy.", + ) + + +class Settings(ServerSettings, ModelSettings): + pass + + +class ConfigFileSettings(ServerSettings): + """Configuration file format settings.""" + + models: List[ModelSettings] = Field(default=[], description="Model configs") diff --git a/llama_cpp/server/types.py b/llama_cpp/server/types.py new file mode 100644 index 000000000..fdd164456 --- /dev/null +++ b/llama_cpp/server/types.py @@ -0,0 +1,316 @@ +from __future__ import annotations + +from typing import List, Optional, Union, Dict +from typing_extensions import TypedDict, Literal + +from pydantic import BaseModel, Field + +import llama_cpp + + +model_field = Field( + description="The model to use for generating completions.", default=None +) + +max_tokens_field = Field( + default=16, ge=1, description="The maximum number of tokens to generate." +) + +min_tokens_field = Field( + default=0, + ge=0, + description="The minimum number of tokens to generate. It may return fewer tokens if another condition is met (e.g. max_tokens, stop).", +) + +temperature_field = Field( + default=0.8, + description="Adjust the randomness of the generated text.\n\n" + + "Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.", +) + +top_p_field = Field( + default=0.95, + ge=0.0, + le=1.0, + description="Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P.\n\n" + + "Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top_p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text.", +) + +min_p_field = Field( + default=0.05, + ge=0.0, + le=1.0, + description="Sets a minimum base probability threshold for token selection.\n\n" + + "The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter min_p represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with min_p=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out.", +) + +stop_field = Field( + default=None, + description="A list of tokens at which to stop generation. If None, no stop tokens are used.", +) + +stream_field = Field( + default=False, + description="Whether to stream the results as they are generated. Useful for chatbots.", +) + +top_k_field = Field( + default=40, + ge=0, + description="Limit the next token selection to the K most probable tokens.\n\n" + + "Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top_k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text.", +) + +repeat_penalty_field = Field( + default=1.1, + ge=0.0, + description="A penalty applied to each token that is already generated. This helps prevent the model from repeating itself.\n\n" + + "Repeat penalty is a hyperparameter used to penalize the repetition of token sequences during text generation. It helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient.", +) + +presence_penalty_field = Field( + default=0.0, + ge=-2.0, + le=2.0, + description="Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics.", +) + +frequency_penalty_field = Field( + default=0.0, + ge=-2.0, + le=2.0, + description="Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim.", +) + +mirostat_mode_field = Field( + default=0, + ge=0, + le=2, + description="Enable Mirostat constant-perplexity algorithm of the specified version (1 or 2; 0 = disabled)", +) + +mirostat_tau_field = Field( + default=5.0, + ge=0.0, + le=10.0, + description="Mirostat target entropy, i.e. the target perplexity - lower values produce focused and coherent text, larger values produce more diverse and less coherent text", +) + +mirostat_eta_field = Field( + default=0.1, ge=0.001, le=1.0, description="Mirostat learning rate" +) + +grammar = Field( + default=None, + description="A CBNF grammar (as string) to be used for formatting the model's output.", +) + + +class CreateCompletionRequest(BaseModel): + prompt: Union[str, List[str]] = Field( + default="", description="The prompt to generate completions for." + ) + suffix: Optional[str] = Field( + default=None, + description="A suffix to append to the generated text. If None, no suffix is appended. Useful for chatbots.", + ) + max_tokens: Optional[int] = Field( + default=16, ge=0, description="The maximum number of tokens to generate." + ) + min_tokens: int = min_tokens_field + temperature: float = temperature_field + top_p: float = top_p_field + min_p: float = min_p_field + echo: bool = Field( + default=False, + description="Whether to echo the prompt in the generated text. Useful for chatbots.", + ) + stop: Optional[Union[str, List[str]]] = stop_field + stream: bool = stream_field + logprobs: Optional[int] = Field( + default=None, + ge=0, + description="The number of logprobs to generate. If None, no logprobs are generated.", + ) + presence_penalty: Optional[float] = presence_penalty_field + frequency_penalty: Optional[float] = frequency_penalty_field + logit_bias: Optional[Dict[str, float]] = Field(None) + seed: Optional[int] = Field(None) + + # ignored or currently unsupported + model: Optional[str] = model_field + n: Optional[int] = 1 + best_of: Optional[int] = 1 + user: Optional[str] = Field(default=None) + + # llama.cpp specific parameters + top_k: int = top_k_field + repeat_penalty: float = repeat_penalty_field + logit_bias_type: Optional[Literal["input_ids", "tokens"]] = Field(None) + mirostat_mode: int = mirostat_mode_field + mirostat_tau: float = mirostat_tau_field + mirostat_eta: float = mirostat_eta_field + grammar: Optional[str] = None + + model_config = { + "json_schema_extra": { + "examples": [ + { + "prompt": "\n\n### Instructions:\nWhat is the capital of France?\n\n### Response:\n", + "stop": ["\n", "###"], + } + ] + } + } + + +class CreateEmbeddingRequest(BaseModel): + model: Optional[str] = model_field + input: Union[str, List[str]] = Field(description="The input to embed.") + user: Optional[str] = Field(default=None) + + model_config = { + "json_schema_extra": { + "examples": [ + { + "input": "The food was delicious and the waiter...", + } + ] + } + } + + +class ChatCompletionRequestMessage(BaseModel): + role: Literal["system", "user", "assistant", "function"] = Field( + default="user", description="The role of the message." + ) + content: Optional[str] = Field( + default="", description="The content of the message." + ) + + +class CreateChatCompletionRequest(BaseModel): + messages: List[llama_cpp.ChatCompletionRequestMessage] = Field( + default=[], description="A list of messages to generate completions for." + ) + functions: Optional[List[llama_cpp.ChatCompletionFunction]] = Field( + default=None, + description="A list of functions to apply to the generated completions.", + ) + function_call: Optional[llama_cpp.ChatCompletionRequestFunctionCall] = Field( + default=None, + description="A function to apply to the generated completions.", + ) + tools: Optional[List[llama_cpp.ChatCompletionTool]] = Field( + default=None, + description="A list of tools to apply to the generated completions.", + ) + tool_choice: Optional[llama_cpp.ChatCompletionToolChoiceOption] = Field( + default=None, + description="A tool to apply to the generated completions.", + ) # TODO: verify + max_tokens: Optional[int] = Field( + default=None, + description="The maximum number of tokens to generate. Defaults to inf", + ) + min_tokens: int = min_tokens_field + logprobs: Optional[bool] = Field( + default=False, + description="Whether to output the logprobs or not. Default is True", + ) + top_logprobs: Optional[int] = Field( + default=None, + ge=0, + description="The number of logprobs to generate. If None, no logprobs are generated. logprobs need to set to True.", + ) + temperature: float = temperature_field + top_p: float = top_p_field + min_p: float = min_p_field + stop: Optional[Union[str, List[str]]] = stop_field + stream: bool = stream_field + presence_penalty: Optional[float] = presence_penalty_field + frequency_penalty: Optional[float] = frequency_penalty_field + logit_bias: Optional[Dict[str, float]] = Field(None) + seed: Optional[int] = Field(None) + response_format: Optional[llama_cpp.ChatCompletionRequestResponseFormat] = Field( + default=None, + ) + + # ignored or currently unsupported + model: Optional[str] = model_field + n: Optional[int] = 1 + user: Optional[str] = Field(None) + + # llama.cpp specific parameters + top_k: int = top_k_field + repeat_penalty: float = repeat_penalty_field + logit_bias_type: Optional[Literal["input_ids", "tokens"]] = Field(None) + mirostat_mode: int = mirostat_mode_field + mirostat_tau: float = mirostat_tau_field + mirostat_eta: float = mirostat_eta_field + grammar: Optional[str] = None + + model_config = { + "json_schema_extra": { + "examples": [ + { + "messages": [ + ChatCompletionRequestMessage( + role="system", content="You are a helpful assistant." + ).model_dump(), + ChatCompletionRequestMessage( + role="user", content="What is the capital of France?" + ).model_dump(), + ] + } + ] + } + } + + +class ModelData(TypedDict): + id: str + object: Literal["model"] + owned_by: str + permissions: List[str] + + +class ModelList(TypedDict): + object: Literal["list"] + data: List[ModelData] + + +class TokenizeInputRequest(BaseModel): + model: Optional[str] = model_field + input: str = Field(description="The input to tokenize.") + + model_config = { + "json_schema_extra": {"examples": [{"input": "How many tokens in this query?"}]} + } + + +class TokenizeInputResponse(BaseModel): + tokens: List[int] = Field(description="A list of tokens.") + + model_config = {"json_schema_extra": {"example": {"tokens": [123, 321, 222]}}} + + +class TokenizeInputCountResponse(BaseModel): + count: int = Field(description="The number of tokens in the input.") + + model_config = {"json_schema_extra": {"example": {"count": 5}}} + + +class DetokenizeInputRequest(BaseModel): + model: Optional[str] = model_field + tokens: List[int] = Field(description="A list of toekns to detokenize.") + + model_config = {"json_schema_extra": {"example": [{"tokens": [123, 321, 222]}]}} + + +class DetokenizeInputResponse(BaseModel): + text: str = Field(description="The detokenized text.") + + model_config = { + "json_schema_extra": {"example": {"text": "How many tokens in this query?"}} + } diff --git a/pyproject.toml b/pyproject.toml index b5affaa9d..9983ef777 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,5 +1,5 @@ [build-system] -requires = ["scikit-build-core[pyproject]>=0.5.1"] +requires = ["scikit-build-core[pyproject]>=0.9.2"] build-backend = "scikit_build_core.build" [project] @@ -15,6 +15,7 @@ dependencies = [ "typing-extensions>=4.5.0", "numpy>=1.20.0", "diskcache>=5.6.1", + "jinja2>=2.11.3", ] requires-python = ">=3.8" classifiers = [ @@ -34,11 +35,17 @@ server = [ "pydantic-settings>=2.0.1", "sse-starlette>=1.6.1", "starlette-context>=0.3.6,<0.4", + "PyYAML>=5.1", ] test = [ "pytest>=7.4.0", "httpx>=0.24.1", "scipy>=1.10", + "fastapi>=0.100.0", + "sse-starlette>=1.6.1", + "starlette-context>=0.3.6,<0.4", + "pydantic-settings>=2.0.1", + "huggingface-hub>=0.23.0" ] dev = [ "black>=23.3.0", @@ -58,7 +65,7 @@ wheel.packages = ["llama_cpp"] cmake.verbose = true cmake.minimum-version = "3.21" minimum-version = "0.5.1" -sdist.include = [".git", "vendor/llama.cpp/.git"] +sdist.include = [".git", "vendor/llama.cpp/*"] [tool.scikit-build.metadata.version] provider = "scikit_build_core.metadata.regex" @@ -71,5 +78,4 @@ Documentation = "https://llama-cpp-python.readthedocs.io/en/latest/" Changelog = "https://llama-cpp-python.readthedocs.io/en/latest/changelog/" [tool.pytest.ini_options] -addopts = "--ignore=vendor" - +testpaths = "tests" diff --git a/scripts/get-releases.sh b/scripts/get-releases.sh new file mode 100755 index 000000000..4c904da78 --- /dev/null +++ b/scripts/get-releases.sh @@ -0,0 +1,47 @@ +#!/bin/bash + +# Function to get all releases +get_all_releases() { + local page=1 + local per_page=100 + local releases="" + local new_releases + + # Prepare headers + local headers=(-H "Accept: application/vnd.github.v3+json") + if [ -n "$GITHUB_TOKEN" ]; then + headers+=(-H "Authorization: Bearer $GITHUB_TOKEN") + fi + + while true; do + response=$(curl -s "${headers[@]}" \ + "https://api.github.com/repos/abetlen/llama-cpp-python/releases?page=$page&per_page=$per_page") + + # Check if the response is valid JSON + if ! echo "$response" | jq empty > /dev/null 2>&1; then + echo "Error: Invalid response from GitHub API" >&2 + echo "Response: $response" >&2 + return 1 + fi + + new_releases=$(echo "$response" | jq -r '.[].tag_name') + if [ -z "$new_releases" ]; then + break + fi + releases="$releases $new_releases" + ((page++)) + done + + echo $releases +} + +# Get all releases and save to file +releases=$(get_all_releases) +if [ $? -ne 0 ]; then + echo "Failed to fetch releases. Please check your internet connection and try again later." >&2 + exit 1 +fi + +echo "$releases" | tr ' ' '\n' > all_releases.txt + +echo "All releases have been saved to all_releases.txt" diff --git a/scripts/releases-to-pep-503.sh b/scripts/releases-to-pep-503.sh new file mode 100755 index 000000000..71910efcb --- /dev/null +++ b/scripts/releases-to-pep-503.sh @@ -0,0 +1,104 @@ +#!/bin/bash + +# Enable exit on error +set -e + +# Function for logging +log_error() { + echo "ERROR: $1" >&2 +} + +log_info() { + echo "INFO: $1" +} + +# Get output directory or default to index/whl/cpu +output_dir=${1:-"index/whl/cpu"} + +# Get pattern from second arg or default to valid python package version pattern +pattern=${2:-"^[v]?[0-9]+\.[0-9]+\.[0-9]+$"} + +# Get the current directory (where the script is run from) +current_dir="$(pwd)" + +# Check if all_releases.txt exists +if [ ! -f "$current_dir/all_releases.txt" ]; then + log_error "all_releases.txt not found in the current directory." + exit 1 +fi + +# Create output directory +mkdir -p "$output_dir" + +# Create an index html file +cat << EOF > "$output_dir/index.html" + + + + + llama-cpp-python +
+ + + +EOF + +# Create llama-cpp-python directory +mkdir -p "$output_dir/llama-cpp-python" + +# Create an index html file in llama-cpp-python directory +cat << EOF > "$output_dir/llama-cpp-python/index.html" + + + +

Links for llama-cpp-python

+EOF + +# Filter releases by pattern +releases=$(grep -E "$pattern" "$current_dir/all_releases.txt") + +# Prepare curl headers +headers=('--header' 'Accept: application/vnd.github.v3+json') +if [ -n "$GITHUB_TOKEN" ]; then + headers+=('--header' "authorization: Bearer $GITHUB_TOKEN") +fi +headers+=('--header' 'content-type: application/json') + +# For each release, get all assets +for release in $releases; do + log_info "Processing release: $release" + response=$(curl -s "${headers[@]}" \ + "https://api.github.com/repos/abetlen/llama-cpp-python/releases/tags/$release") + + if [ -z "$response" ]; then + log_error "Empty response from GitHub API for release $release" + continue + fi + + if ! echo "$response" | jq -e '.assets' > /dev/null 2>&1; then + log_error "Invalid or unexpected response from GitHub API for release $release" + log_error "Response: $response" + continue + fi + + # Get release version from release ie v0.1.0-cu121 -> v0.1.0 + release_version=$(echo "$release" | grep -oE "^[v]?[0-9]+\.[0-9]+\.[0-9]+") + echo "

$release_version

" >> "$output_dir/llama-cpp-python/index.html" + + wheel_urls=$(echo "$response" | jq -r '.assets[] | select(.name | endswith(".whl")) | .browser_download_url') + if [ -z "$wheel_urls" ]; then + log_error "No wheel files found for release $release" + continue + fi + + echo "$wheel_urls" | while read -r asset; do + echo " $asset" >> "$output_dir/llama-cpp-python/index.html" + echo "
" >> "$output_dir/llama-cpp-python/index.html" + done +done + +echo " " >> "$output_dir/llama-cpp-python/index.html" +echo "" >> "$output_dir/llama-cpp-python/index.html" +echo "" >> "$output_dir/llama-cpp-python/index.html" + +log_info "Index generation complete. Output directory: $output_dir" diff --git a/tests/test_grammar.py b/tests/test_grammar.py deleted file mode 100644 index 2e24903c2..000000000 --- a/tests/test_grammar.py +++ /dev/null @@ -1,13 +0,0 @@ -import llama_cpp - -tree = """ -leaf ::= "." -node ::= leaf | "(" node node ")" -root ::= node -""" - -def test_grammar_from_string(): - grammar = llama_cpp.LlamaGrammar.from_string(tree) - assert grammar._n_rules == 3 - assert grammar._start_rule_index == 2 - assert grammar.grammar is not None diff --git a/tests/test_llama.py b/tests/test_llama.py index dac33b74d..fc182ae20 100644 --- a/tests/test_llama.py +++ b/tests/test_llama.py @@ -1,12 +1,22 @@ import ctypes +import multiprocessing import numpy as np -import pytest from scipy.special import log_softmax +from huggingface_hub import hf_hub_download + +import pytest + import llama_cpp +import llama_cpp._internals as internals -MODEL = "./vendor/llama.cpp/models/ggml-vocab-llama.gguf" + +MODEL = "./vendor/llama.cpp/models/ggml-vocab-llama-spm.gguf" + + +def test_llama_cpp_version(): + assert llama_cpp.__version__ def test_llama_cpp_tokenization(): @@ -47,247 +57,162 @@ def test_llama_cpp_tokenization(): @pytest.fixture -def mock_llama(monkeypatch): - def setup_mock(llama: llama_cpp.Llama, output_text: str): - n_ctx = llama.n_ctx() - n_vocab = llama.n_vocab() - output_tokens = llama.tokenize( - output_text.encode("utf-8"), add_bos=True, special=True - ) - logits = (llama_cpp.c_float * (n_vocab * n_ctx))(-100.0) - for i in range(n_ctx): - output_idx = i + 1 # logits for first tokens predict second token - if output_idx < len(output_tokens): - logits[i * n_vocab + output_tokens[output_idx]] = 100.0 - else: - logits[i * n_vocab + llama.token_eos()] = 100.0 - n = 0 - last_n_tokens = 0 - - def mock_decode(ctx: llama_cpp.llama_context_p, batch: llama_cpp.llama_batch): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - assert batch.n_tokens > 0, "no tokens in batch" - assert all( - batch.n_seq_id[i] == 1 for i in range(batch.n_tokens) - ), "n_seq >1 not supported by mock_llama" - assert all( - batch.seq_id[i][0] == 0 for i in range(batch.n_tokens) - ), "n_seq >1 not supported by mock_llama" - assert batch.logits[ - batch.n_tokens - 1 - ], "logits not allocated for last token" - # Update the mock context state - nonlocal n - nonlocal last_n_tokens - n = max(batch.pos[i] for i in range(batch.n_tokens)) + 1 - last_n_tokens = batch.n_tokens - return 0 - - def mock_get_logits(ctx: llama_cpp.llama_context_p): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - assert n > 0, "mock_llama_decode not called" - assert last_n_tokens > 0, "mock_llama_decode not called" - # Return view of logits for last_n_tokens - return (llama_cpp.c_float * (last_n_tokens * n_vocab)).from_address( - ctypes.addressof(logits) - + (n - last_n_tokens) * n_vocab * ctypes.sizeof(llama_cpp.c_float) - ) - - monkeypatch.setattr("llama_cpp.llama_cpp.llama_decode", mock_decode) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_get_logits", mock_get_logits) - - def mock_kv_cache_clear(ctx: llama_cpp.llama_context_p): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - def mock_kv_cache_seq_rm( - ctx: llama_cpp.llama_context_p, - seq_id: llama_cpp.llama_seq_id, - pos0: llama_cpp.llama_pos, - pos1: llama_cpp.llama_pos, - ): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - def mock_kv_cache_seq_cp( - ctx: llama_cpp.llama_context_p, - seq_id_src: llama_cpp.llama_seq_id, - seq_id_dst: llama_cpp.llama_seq_id, - pos0: llama_cpp.llama_pos, - pos1: llama_cpp.llama_pos, - ): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - def mock_kv_cache_seq_keep( - ctx: llama_cpp.llama_context_p, - seq_id: llama_cpp.llama_seq_id, - ): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - def mock_kv_cache_seq_shift( - ctx: llama_cpp.llama_context_p, - seq_id: llama_cpp.llama_seq_id, - pos0: llama_cpp.llama_pos, - pos1: llama_cpp.llama_pos, - ): - # Test some basic invariants of this mocking technique - assert ctx == llama._ctx.ctx, "context does not match mock_llama" - return - - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_clear", mock_kv_cache_clear) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_rm", mock_kv_cache_seq_rm) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_cp", mock_kv_cache_seq_cp) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_keep", mock_kv_cache_seq_keep) - monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_shift", mock_kv_cache_seq_shift) - - return setup_mock - - -def test_llama_patch(mock_llama): - n_ctx = 128 - llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True, n_ctx=n_ctx) - n_vocab = llama_cpp.llama_n_vocab(llama._model.model) - assert n_vocab == 32000 - - text = "The quick brown fox" - output_text = " jumps over the lazy dog." - all_text = text + output_text - - ## Test basic completion from bos until eos - mock_llama(llama, all_text) - completion = llama.create_completion("", max_tokens=36) - assert completion["choices"][0]["text"] == all_text - assert completion["choices"][0]["finish_reason"] == "stop" - - ## Test basic completion until eos - mock_llama(llama, all_text) - completion = llama.create_completion(text, max_tokens=20) - assert completion["choices"][0]["text"] == output_text - assert completion["choices"][0]["finish_reason"] == "stop" - - ## Test streaming completion until eos - mock_llama(llama, all_text) - chunks = list(llama.create_completion(text, max_tokens=20, stream=True)) - assert "".join(chunk["choices"][0]["text"] for chunk in chunks) == output_text - assert chunks[-1]["choices"][0]["finish_reason"] == "stop" - - ## Test basic completion until stop sequence - mock_llama(llama, all_text) - completion = llama.create_completion(text, max_tokens=20, stop=["lazy"]) - assert completion["choices"][0]["text"] == " jumps over the " - assert completion["choices"][0]["finish_reason"] == "stop" - - ## Test streaming completion until stop sequence - mock_llama(llama, all_text) - chunks = list( - llama.create_completion(text, max_tokens=20, stream=True, stop=["lazy"]) - ) - assert ( - "".join(chunk["choices"][0]["text"] for chunk in chunks) == " jumps over the " +def llama_cpp_model_path(): + repo_id = "Qwen/Qwen2-0.5B-Instruct-GGUF" + filename = "qwen2-0_5b-instruct-q8_0.gguf" + model_path = hf_hub_download(repo_id, filename) + return model_path + + +def test_real_model(llama_cpp_model_path): + import os + assert os.path.exists(llama_cpp_model_path) + + params = llama_cpp.llama_model_default_params() + params.use_mmap = llama_cpp.llama_supports_mmap() + params.use_mlock = llama_cpp.llama_supports_mlock() + params.check_tensors = False + + model = internals.LlamaModel(path_model=llama_cpp_model_path, params=params) + + cparams = llama_cpp.llama_context_default_params() + cparams.n_ctx = 16 + cparams.n_batch = 16 + cparams.n_ubatch = 16 + cparams.n_threads = multiprocessing.cpu_count() + cparams.n_threads_batch = multiprocessing.cpu_count() + cparams.logits_all = False + cparams.flash_attn = True + + context = internals.LlamaContext(model=model, params=cparams) + tokens = model.tokenize(b"Hello, world!", add_bos=True, special=True) + + assert tokens == [9707, 11, 1879, 0] + + tokens = model.tokenize(b"The quick brown fox jumps", add_bos=True, special=True) + + batch = internals.LlamaBatch(n_tokens=len(tokens), embd=0, n_seq_max=1) + + seed = 1337 + sampler = internals.LlamaSampler() + sampler.add_top_k(50) + sampler.add_top_p(0.9, 1) + sampler.add_temp(0.8) + sampler.add_dist(seed) + + result = tokens + n_eval = 0 + for _ in range(4): + batch.set_batch(tokens, n_past=n_eval, logits_all=False) + context.decode(batch) + n_eval += len(tokens) + token_id = sampler.sample(context, -1) + tokens = [token_id] + result += tokens + + output = result[5:] + output_text = model.detokenize(output, special=True) + assert output_text == b" over the lazy dog" + +def test_real_llama(llama_cpp_model_path): + model = llama_cpp.Llama( + llama_cpp_model_path, + n_ctx=32, + n_batch=32, + n_ubatch=32, + n_threads=multiprocessing.cpu_count(), + n_threads_batch=multiprocessing.cpu_count(), + logits_all=False, + flash_attn=True, ) - assert chunks[-1]["choices"][0]["finish_reason"] == "stop" - - ## Test basic completion until length - mock_llama(llama, all_text) - completion = llama.create_completion(text, max_tokens=2) - assert completion["choices"][0]["text"] == " jumps" - assert completion["choices"][0]["finish_reason"] == "length" - - ## Test streaming completion until length - mock_llama(llama, all_text) - chunks = list(llama.create_completion(text, max_tokens=2, stream=True)) - assert "".join(chunk["choices"][0]["text"] for chunk in chunks) == " jumps" - assert chunks[-1]["choices"][0]["finish_reason"] == "length" - - -def test_llama_pickle(): - import pickle - import tempfile - - fp = tempfile.TemporaryFile() - llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True) - pickle.dump(llama, fp) - fp.seek(0) - llama = pickle.load(fp) - assert llama - assert llama.ctx is not None - - text = b"Hello World" - - assert llama.detokenize(llama.tokenize(text)) == text - - -def test_utf8(mock_llama): - llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True, logits_all=True) + output = model.create_completion( + "The quick brown fox jumps", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + seed=1337 + ) + assert output["choices"][0]["text"] == " over the lazy dog" + + + output = model.create_completion( + "The capital of france is paris, 'true' or 'false'?:\n", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + seed=1337, + grammar=llama_cpp.LlamaGrammar.from_string(""" +root ::= "true" | "false" +""") + ) + assert output["choices"][0]["text"] == "true" - output_text = "😀" + suffix = b"rot" + tokens = model.tokenize(suffix, add_bos=True, special=True) + def logit_processor_func(input_ids, logits): + for token in tokens: + logits[token] *= 1000 + return logits - ## Test basic completion with utf8 multibyte - mock_llama(llama, output_text) - completion = llama.create_completion("", max_tokens=4) - assert completion["choices"][0]["text"] == output_text + logit_processors = llama_cpp.LogitsProcessorList( + [logit_processor_func] + ) - ## Test basic completion with incomplete utf8 multibyte - mock_llama(llama, output_text) - completion = llama.create_completion("", max_tokens=1) - assert completion["choices"][0]["text"] == "" + output = model.create_completion( + "The capital of france is par", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + seed=1337, + logits_processor=logit_processors + ) + assert output["choices"][0]["text"].lower().startswith("rot") + model.set_seed(1337) -def test_llama_server(): - from fastapi.testclient import TestClient - from llama_cpp.server.app import create_app, Settings + state = model.save_state() - settings = Settings( - model=MODEL, - vocab_only=True, + output = model.create_completion( + "Pick a number from 1 to 10?:\n", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + grammar=llama_cpp.LlamaGrammar.from_string(""" +root ::= "1" | "2" | "3" | "4" | "5" | "6" | "7" | "8" | "9" | "10" +""") ) - app = create_app(settings) - client = TestClient(app) - response = client.get("/v1/models") - assert response.json() == { - "object": "list", - "data": [ - { - "id": MODEL, - "object": "model", - "owned_by": "me", - "permissions": [], - } - ], - } - - -@pytest.mark.parametrize( - "size_and_axis", - [ - ((32_000,), -1), # last token's next-token logits - ((10, 32_000), -1), # many tokens' next-token logits, or batch of last tokens - ((4, 10, 32_000), -1), # batch of texts - ], -) -@pytest.mark.parametrize("convert_to_list", [True, False]) -def test_logits_to_logprobs(size_and_axis, convert_to_list: bool, atol: float = 1e-7): - size, axis = size_and_axis - logits: np.ndarray = -np.random.uniform(low=0, high=60, size=size) - logits = logits.astype(np.single) - if convert_to_list: - # Currently, logits are converted from arrays to lists. This may change soon - logits = logits.tolist() - log_probs = llama_cpp.Llama.logits_to_logprobs(logits, axis=axis) - log_probs_correct = log_softmax(logits, axis=axis) - assert log_probs.dtype == np.single - assert log_probs.shape == size - assert np.allclose(log_probs, log_probs_correct, atol=atol) - + number_1 = output["choices"][0]["text"] + + output = model.create_completion( + "Pick a number from 1 to 10?:\n", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + grammar=llama_cpp.LlamaGrammar.from_string(""" +root ::= "1" | "2" | "3" | "4" | "5" | "6" | "7" | "8" | "9" | "10" +""") + ) + number_2 = output["choices"][0]["text"] + + model.load_state(state) + + output = model.create_completion( + "Pick a number from 1 to 10?:\n", + max_tokens=4, + top_k=50, + top_p=0.9, + temperature=0.8, + grammar=llama_cpp.LlamaGrammar.from_string(""" +root ::= "1" | "2" | "3" | "4" | "5" | "6" | "7" | "8" | "9" | "10" +""") + ) + number_3 = output["choices"][0]["text"] -def test_llama_cpp_version(): - assert llama_cpp.__version__ + assert number_1 != number_2 + assert number_1 == number_3 diff --git a/tests/test_llama_chat_format.py b/tests/test_llama_chat_format.py new file mode 100644 index 000000000..f031bf72b --- /dev/null +++ b/tests/test_llama_chat_format.py @@ -0,0 +1,89 @@ +import json + +import jinja2 + +from llama_cpp import ( + ChatCompletionRequestUserMessage, +) +import llama_cpp.llama_types as llama_types +import llama_cpp.llama_chat_format as llama_chat_format + +from llama_cpp.llama_chat_format import hf_tokenizer_config_to_chat_formatter + +def test_mistral_instruct(): + chat_template = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" + chat_formatter = jinja2.Template(chat_template) + messages = [ + llama_types.ChatCompletionRequestUserMessage(role="user", content="Instruction"), + llama_types.ChatCompletionRequestAssistantMessage(role="assistant", content="Model answer"), + llama_types.ChatCompletionRequestUserMessage(role="user", content="Follow-up instruction"), + ] + response = llama_chat_format.format_mistral_instruct( + messages=messages, + ) + prompt = ("" if response.added_special else "") + response.prompt + reference = chat_formatter.render( + messages=messages, + bos_token="", + eos_token="", + ) + assert prompt == reference + + +mistral_7b_tokenizer_config = """{ + "add_bos_token": true, + "add_eos_token": false, + "added_tokens_decoder": { + "0": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "1": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "2": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + } + }, + "additional_special_tokens": [], + "bos_token": "", + "clean_up_tokenization_spaces": false, + "eos_token": "", + "legacy": true, + "model_max_length": 1000000000000000019884624838656, + "pad_token": null, + "sp_model_kwargs": {}, + "spaces_between_special_tokens": false, + "tokenizer_class": "LlamaTokenizer", + "unk_token": "", + "use_default_system_prompt": false, + "chat_template": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" +}""" + + +def test_hf_tokenizer_config_str_to_chat_formatter(): + tokenizer_config = json.loads(mistral_7b_tokenizer_config) + chat_formatter = hf_tokenizer_config_to_chat_formatter( + tokenizer_config + ) + chat_formatter_respoonse = chat_formatter( + messages=[ + ChatCompletionRequestUserMessage(role="user", content="Hello, world!"), + ] + ) + + assert chat_formatter_respoonse.prompt == ("[INST] Hello, world! [/INST]" "") diff --git a/tests/test_llama_grammar.py b/tests/test_llama_grammar.py new file mode 100644 index 000000000..34ef2874d --- /dev/null +++ b/tests/test_llama_grammar.py @@ -0,0 +1,78 @@ +import llama_cpp +import json + +tree = """ +leaf ::= "." +node ::= leaf | "(" node node ")" +root ::= node +""" + + +def test_grammar_from_string(): + grammar = llama_cpp.LlamaGrammar.from_string(tree) + # assert grammar._n_rules == 3 + # assert grammar._start_rule_index == 2 + # assert grammar.grammar is not None + + +def test_composed_pydantic_grammar(): + """ + from pydantic import BaseModel + + class A(BaseModel): + a: int + + class B(BaseModel): + a: A + b: int + """ + + # This schema corresponds to the grammar in the comment above. + # We don't use the pydantic models directly to avoid the dependency. + schema = { + "$defs": { + "A": { + "properties": {"a": {"title": "A", "type": "integer"}}, + "required": ["a"], + "title": "A", + "type": "object", + } + }, + "properties": { + "a": {"$ref": "#/$defs/A"}, + "b": {"title": "B", "type": "integer"}, + }, + "required": ["a", "b"], + "title": "B", + "type": "object", + } + + grammar = llama_cpp.LlamaGrammar.from_json_schema(json.dumps(schema)) + + # assert grammar.grammar is not None + + +def test_grammar_anyof(): + sch = { + "properties": { + "temperature": { + "description": "The temperature mentioned", + "type": "number", + }, + "unit": { + "anyOf": [ + { + "description": "Unit for temperature", + "enum": ["celsius", "fahrenheit"], + "type": "string", + }, + {"type": "null"}, + ], + }, + }, + "type": "object", + } + + grammar = llama_cpp.LlamaGrammar.from_json_schema(json.dumps(sch)) + + # assert grammar.grammar is not None diff --git a/tests/test_llama_speculative.py b/tests/test_llama_speculative.py new file mode 100644 index 000000000..b5d450567 --- /dev/null +++ b/tests/test_llama_speculative.py @@ -0,0 +1,16 @@ +import numpy as np + +from llama_cpp.llama_speculative import LlamaPromptLookupDecoding + +def test_find_candidate_pred_tokens(): + find_candidate_pred_tokens = LlamaPromptLookupDecoding.find_candidate_pred_tokens + + # Test Case 1: Matching ngram is found + input_ids1 = np.array([1, 2, 3, 1, 2, 3, 1, 2, 3]) + result1 = find_candidate_pred_tokens(input_ids1, max_ngram_size=3, num_pred_tokens=2) + assert np.array_equal(result1, np.array([1, 2])) + + # Test Case 2: Matching ngram is not found + input_ids2 = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9]) + result2 = find_candidate_pred_tokens(input_ids2, max_ngram_size=3, num_pred_tokens=2) + assert np.array_equal(result2, np.array([])) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 0e18b2e7d..8733e0cf6 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 0e18b2e7d0b5c0a509ea40098def234b8d4a938a +Subproject commit 8733e0cf6eefc7c7752297cc22d0836706f4222c