1 Pre-release version can run various models supported by the main branch of Ollama. only tested for windows .
test env:
Intel(R) Core(TM) Ultra 9 285H
Windows11 24H2
OLLAMA_INTEL_GPU : true = use intel gpu; false normal ollama
OLLAMA_INTEL_IF_TYPE : SYCL = use SYCL of ggml lib to find gpu device and auto parse memory size; ONEAPI = use level zero lib to find gpu device (can't get memory size)
OLLAMA_NUM_GPU : when use OLLAMA_INTEL_IF_TYPE=ONEAPI, this param is max number of layers offload to gpu; default = 64. if use OLLAMA_INTEL_IF_TYPE=SYCL, layers offload to gpu is auto calculate.
.\ollama-intel-gpu.bat.\ollama.exe run gemma3:12b --verboseThis document recode process of merge ggml-sycl from llama.cpp. to support Intel-Gpu.
Only tested in windows and intel integrated Graphics Card.
A portable package in https://github.com/chnxq/ollama/releases
Install Intel-OneApi: from OneApiBaseToolkit
Other ref: SYCL document for detail.
(default install in C:\Program Files (x86)\Intel\oneAPI) for next example.
cmd.exe "/K" '"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" && powershell'build libraries: on ollama root directory:
cmake -B build -G "Ninja" -DGGML_SYCL=ON -DGGML_SYCL_TARGET=INTEL -DGGML_CPU_ALL_VARIANTS=ON -DGGML_BACKEND_DL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Releasecmake --build build --config Release -j source /opt/intel/oneapi/setvars.shbuild libraries: on ollama root directory:
cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=INTEL -DGGML_CPU_ALL_VARIANTS=ON -DGGML_BACKEND_DL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DCMAKE_BUILD_TYPE=Releasecmake --build build --config Release -j -vbuild go src:
go build -o ollama.exeon ollama root directory
cmd.exe "/K" '"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" && powershell'
set OLLAMA_INTEL_GPU=true
set OLLAMA_INTEL_IF_TYPE=SYCL
set OLLAMA_NUM_GPU=999
set SYCL_CACHE_PERSISTENT=1
set OLLAMA_LIBRARY_PATH=./build/lib/ollama
# run ollama server
.\ollama.exe serveor use shell ollama-intel-gpu.bat:
set OLLAMA_INTEL_GPU=true
set OLLAMA_INTEL_IF_TYPE=SYCL
set OLLAMA_NUM_GPU=64
set SYCL_CACHE_PERSISTENT=1
set OLLAMA_LIBRARY_PATH=./build/lib/ollama
set ONEAPI_ROOT=C:\Program Files (x86)\Intel\oneAPI
set PATH=%PATH%;%ONEAPI_ROOT%\2025.1\bin;./build/lib/ollama;
.\ollama.exe servenote:
- set OLLAMA_NUM_GPU=xxx xxx: It needs to be manually set. According to the number of model layers that the video memory can load.for example my T140 has 16G shared video memory,i set it to 64.
- Next 2 env is necessary if use sycl for discover intel gpu: set OLLAMA_INTEL_GPU=true set OLLAMA_INTEL_IF_TYPE=SYCL (env OLLAMA_INTEL_IF_TYPE is used in go and c code of ollama and llama.cpp,same name as build param)
- When use pure CPU inference,a known bug need to delete ggml_sycl library temporay.
# run ollama test client 1
.\ollama.exe run deepseek-r1:1.5b --verbose# run ollama test client 2
.\ollama.exe run qwen3:4b-fp16 --verbose# run ollama test client 3
.\ollama.exe run gemma3:12b --verbose