Repository navigation
Releases
v3.9.6
Change logs
- Add environment variable
ZLUDA_CACHE_DIR, thanks @sfinktah - Fix build with the latest Rust compilers.
- Add
zluda_get_version()- returns Git commit hash. - Introduce HIP SDK 6.4 and 7.1 supports.
- Suppress ignorable compiler warnings and fix bugs.
v3.9.5
Change logs
- Add AMD Comgr 2.9.0 support.
Enable AMD Adrenaline driver 25.5.1 support.
Nightly Features
In order to enable cuBLASLt/cuDNN acceleration on supported device, you should download and unpack HIP SDK extension upon your existing HIP SDK 6.2 installation.
HIP SDK extension: Download
(unzip and paste folders upon path/to/AMD/ROCm/6.2)
v3.9.4
Change logs
- Rewrite
cuTexObjectDestroy()to mimic CUDA driver api more precisely. - Add basic ONNX Runtime support. (nightly)
- Add workaround for HIP SDK 6.2 bug.
- Enable COLMAP.
- Fix
torch.topk().
Nightly Features
In order to enable cuBLASLt/cuDNN acceleration on supported device, you should download and unpack HIP SDK extension upon your existing HIP SDK 6.2 installation.
HIP SDK extension: RUNTIME FULL
(unzip and paste folders upon path/to/AMD/ROCm/6.2)
HIP SDK extension runtime includes the minimal requirements that are needed to run applications.
HIP SDK extension full includes runtime, plus C/C++ headers, statically linked libraries, and more.
v3.9.3
Change logs
- [DNN] Add batch normalization forward/backward extended api.
- [DNN] Add basic paddleocr support.
Nightly Features
From ZLUDA v3.9.3, HIP SDK extension includes hipBLASLt (runtime/full) and composable kernel (full).
Also, MIOpen is rebuilt with composable kernel.
In order to enable cuBLASLt/cuDNN acceleration on supported device, you should download and unpack HIP SDK extension upon your existing HIP SDK 6.2 installation.
HIP SDK extension: RUNTIME FULL
(unzip and paste folders upon path/to/AMD/ROCm/6.2)
HIP SDK extension runtime includes the minimal requirements that are needed to run applications.
HIP SDK extension full includes runtime, plus C/C++ headers, statically linked libraries, and more.
v3.9.2
Change logs
- [Driver/Stream] Add cuStreamBeginCapture_v2.
- [Driver/Stream] Add cuStreamEndCapture.
- [Driver/Stream] Add cuThreadExchangeStreamCaptureMode.
- [Driver/Graph] Add cuGraphGetNodes.
- [Driver/Graph] Add cuGraphInstantiateWithFlags.
From ZLUDA v3.9.2, torch.compile() inductor and cuda_graph is now available.
(using custom triton)
cuDNN on Windows
There is no update for cuDNN. However, HIP SDK extension is re-built.
Supported Architectures
- gfx908, gfx90a, gfx942
- gfx1010
- gfx1030, gfx1031, gfx1032
- gfx1100, gfx1101, gfx1102, gfx1103
- gfx1150
- gfx1201
In order to enable cuDNN acceleration on supported device, you should download and unpack HIP SDK extension upon your existing HIP SDK 6.2 installation.
HIP SDK extension: DOWNLOAD
(unzip and paste folders upon path/to/AMD/ROCm/6.2)
โป HIP SDK extension does not include hipBLASLt.
v3.9.1
Change logs
- [PTX] Add lexeme kernel parameter attribute
.ptrfor ptx parser. - [Driver] Add cuStreamGetPriority.
- [cuBLAS] Add cublasStrmm_v2, cublasSgeam.
- [cuDNN] Add cudnnBatchNormalizationForwardInference.
- [nvRTC] Fix compilation and invalid program error.
cuDNN on Windows
From ZLUDA v3.9.1, cuDNN is now available on A1111 AMDGPU fork.
Supported Architectures
- gfx908, gfx90a, gfx942
- gfx1010
- gfx1030, gfx1031, gfx1032
- gfx1100, gfx1101, gfx1102, gfx1103
- gfx1150
- gfx1201
In order to enable cuDNN acceleration on supported device, you should download and unpack HIP SDK extension upon your existing HIP SDK 6.2 installation.
HIP SDK extension: DOWNLOAD
(unzip and paste folders upon path/to/AMD/ROCm/6.2)
โป HIP SDK extension does not include hipBLASLt.
v3.9.0
Change logs
- Add cublasStrsm_v2, cublasSgelsBatched.
- [Runtime] Do nothing instead of panicking. (release-mode only)
- Minor fixes.
- [Nightly] Enable FP16 Conv2d.
- [Nightly] Enable PyTorch Flash Attention 2.
cuDNN on Windows
From ZLUDA v3.9.0, the nightly build includes FP16 Conv2d and Flash Attention 2 support.
Supported Architectures
- gfx908, gfx90a
- gfx940, gfx941, gfx942
- gfx1030
- gfx1100, gfx1101, gfx1102
- gfx1150
In order to enable cuDNN acceleration on supported device, you should download and unpack HIP SDK extension upon your existing HIP SDK 6.2 installation.
HIP SDK extension: DOWNLOAD
(unzip and paste folders upon path/to/AMD/ROCm/6.2)
โป HIP SDK extension does not include hipBLASLt.
v3.8.8
Change logs
- Update README.
- Add missing
cudaDeviceSynchronizeandcudaGetErrorStringfor cuda runtime api. - Fix bug of cuda runtime api.
- [Nightly] Add
nvcuda.zluda_get_nightly_flag.
It returns1if ZLUDA was built with--nightlyflag. Otherwise, it returns0. - [Nightly] Unlock cuBLASLt matmul compute type.
- [Nightly] Experimental cuDNN support on Windows. (disabled by default)
cuDNN on Windows
Starting from v3.8.8, the nightly build will include cudnn.dll.
Supported Architectures
- gfx908, gfx90a
- gfx940, gfx941, gfx942
- gfx1030
- gfx1100, gfx1101, gfx1102
In order to enable cuDNN acceleration on supported device, you should download and unpack HIP SDK extension upon your existing HIP SDK 6.2 installation.
HIP SDK extension: DOWNLOAD
(unzip and paste folders upon path/to/AMD/ROCm/6.2)
โป HIP SDK extension does not include hipBLASLt.
v3.8.7
Change logs
- Add dummy cuFFTW dll file.
- Enable PyTorch fftn, ifftn, rfftn, and irfftn.
Known issues
v3.8.6
Change logs
- Update ZLUDA BLASLt including bindings for ROCm 6.x api changes.
- Enable hipBLASLt on Windows in nightly build.
You'll need to compile hipBLASLt yourself or get unofficial library for now. - Add cublasXt apis.
- Regenerate cublas bindings.
cublasHgemm*apis are now exported. - Introduce nightly build.
Build with--nightlyfor unstable features. Currently, only one unstable feature, hipBLASLt on Windows, is available. - Allow build for HIP SDK 5.5 and 6.1. (untested)
Any bug/issue report is welcome.