Build issue (HIPBLAS for gfx600)

[danilabagroff]

Hi folks,

I am trying to build llama.cpp for my AMD FirePro D500 with -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx600 but stuck at the very beginning:
/opt/rocm-6.2.0/lib/llvm/bin/../../../include/hip/hip_runtime.h:41:2: error: HIP is not supported on the specified GPU ARCH with wavefront size 64

I have find out(https://clang.llvm.org/docs/AMDGPUSupport.html) that we have a macro:
__AMDGCN_WAVEFRONT_SIZE__ | Defines the wavefront size. Allowed values are 32 and 64. __AMDGCN_WAVEFRONT_SIZE | Alias to __AMDGCN_WAVEFRONT_SIZE__. To be deprecated.

I have tried to define __AMDGCN_WAVEFRONT_SIZE=32 within DCMAKE_CXX_FLAGS without any luck :(
Do I have any chance to build llama.cpp for my target GPU?

Thank you.

[8XXD8]

gfx600 is unsupported, and has been for years. Rocm 6.2 officially only supports gfx906 (Mi50 and Radeon VII) an up, you can make gfx900 (Mi25 and Vega) work but that is all, anything older is a no go sadly.
Maybe the Vulkan backend can work.

[Allan-Luu]

maybe an older version of ROCm would work? try using a docker container that contains maybe ROCm 5.7 or ROCm 6.0 if ROCm 6.2 doesn't support gfx600

[FlorianHeigl]

I heard that debian is running their own builders enabling missing AMD platforms. If you got the opportunity, you could try that. so debian os and amdgpu, rocm etc from the DEBIAN repos. _NOT from AMD.