Releases: blace-ai/blace-ai
v0.9.96: Torchvision fix
Changes
- fix some models not being able to be run on CUDA backend
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA
v0.9.95: Add more operators
Changes
- add SumOp and AddOp
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA
v0.9.94: Neat Management of Worker Side Errors
Changes
- IPCEvaluator returns error code if computation is not finished after 2m (timeout) or if worker crashes.
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA
v0.9.92: Fix for debug builds on Windows
Changes
- fix slim library linking to wrong logger in debug builds leading to crashes
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA
v0.9.91: Debug builds on Windows & IPC worker logging
Changes
- debug-build targets can now link against the IPC version of blace.ai
- IPC worker now logs errors to system
std::filesystem::temp_directory_path() / "blace-ai-logs"
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA
v0.9.90: CUDA IPC Memory sharing & various IPC improvements
Changes
- allow IPC process to share CUDA memory on Windows and Linux, bringing performance on par with non-IPC version. A blace::RawMemoryObject holding native CUDA memory will natively share memory via the GPU. Use the ToDeviceOp with target CPU to force cpu memory sharing. Please check the docs regarding driver requirements
- evaluators now return a tuple <blace::ml_core::ReturnCode, std::shared_ptrblace::RawMemoryObject> with the first entry giving feedback about the execution result
- terminating the host process automatically terminates the IPC worker on all OS systems now
- ml_core::ProgressCallback was simplified, progress can be fetched on both IPC and non-IPC evaluators via get_nodes_processed()
- smaller bugfixes
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA
v0.9.88: (non-functional changes) Optional faster Model Loading
Changes
- make faster model loading optional to ensure backwards compatibility
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA
v0.9.87: Model Loading
Changes
- more efficient model loading if target device is CUDA or MPS
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA
v0.9.86: Model Wizard licensing
Changes
- the model wizard now can consume licenses to create unrestricted model artifacts
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA
v0.9.85: Library fix on Linux
Changes
- Ubuntu: fix a pytorch library that could not be loaded by the blace ai runtime in some cases
By downloading below artifacts you accept the license agreement of this repository.
Downloads
Windows CPU+DirectML
Windows CUDA+DirectML
MacOS Universal
Ubuntu CPU
Ubuntu CUDA