It seems like ROCm 6.4-7.0 still support gfx906 but one needs to compile TensileLibrary for gfx906 manually (or download it from somewhere else). Instructions can be found in this issue: ROCm/ROCm#4625
-
rocblas still supports gfx906, you can build from source to get the TensileLibrary file for gfx906 for now. You can edit this line to only build for gfx906, this should speed up the build.
-
Just for anyone who won't be able to build from source. I think it is possible to take TensileLibrary_lazy_gfx906.dat from this package: https://archlinux.org/packages/extra/x86_64/rocblas/ Download it and extract.
-
Someone managed to run ComfyUI with ROCm 6.4.3. Just need to copy TensileLibrary to pytorch package folder inside local virtualenv dir https://www.reddit.com/r/linux4noobs/comments/1ly8rq6/comment/nd7f2iw/
- git clone https://github.com/ROCmSoftwarePlatform/Tensile.git
- cd Tensile
- git checkout (ROCm version)
- git clone https://github.com/ROCmSoftwarePlatform/rocBLAS.git
- cd rocBLAS
- git checkout (ROCm version)
- ./install.sh --no_hipblaslt --architecture "gfx906;gfx906:xnack-;gfx906:xnack+" -j14 -t (Tensile directory)
- compiled TensileLibrary files should be in ./build/release/rocblas-install/lib/rocblas/library
https://github.com/kprinssu/Kokoro-FastAPI/tree/rocm-release
This works with triton-gfx906 fork: https://github.com/AuleTechnologies/Aule-Attention