Option to build GFX9 ROCm support?

[nikAizuddin]

Can we have a build option to support older GPUs such as GFX9? So that we don't have to patch scripts/build_llama_and_whisper.sh file in order to build for GFX9:

--- a/container-images/scripts/build_llama_and_whisper.sh
+++ b/container-images/scripts/build_llama_and_whisper.sh
@@ -62,7 +62,7 @@ configure_common_flags() {
   common_flags=("-DGGML_NATIVE=OFF")
   case "$containerfile" in
     rocm)
-      common_flags+=("-DGGML_HIP=ON" "-DAMDGPU_TARGETS=${AMDGPU_TARGETS:-gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102}")
+      common_flags+=("-DGGML_HIP=ON" "-DAMDGPU_TARGETS=${AMDGPU_TARGETS:-gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102,gfx900}")
       ;;
     cuda)
       common_flags+=("-DGGML_CUDA=ON" "-DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined")
@@ -120,7 +120,7 @@ main() {
 
   clone_and_build_llama_cpp
   dnf -y clean all
-  rm -rf /var/cache/*dnf* /opt/rocm-*/lib/*/library/*gfx9*
+  rm -rf /var/cache/*dnf*
   ldconfig # needed for libraries
 }

[rhatdan]

The issue is size I believe. You should be able to specify AMD_GPU_TARGETS=gfx900 to get it to build,
but we would need a way to not remove the gfx9 files.

[nikAizuddin]

You should be able to specify AMD_GPU_TARGETS=gfx900

I just tried with make build IMAGE=rocm AMD_GPU_TARGETS=gfx900, it doesn't work. It seems AMD_GPU_TARGETS args from make is ignored because I see podman build --no-cache --platform linux/amd64 -t quay.io/ramalama/rocm -f rocm/Containerfile . command is used.

we would need a way to not remove the gfx9 files.

I think adding a simple regex check for gfx9** should be sufficient:

  if [[ ! "$AMDGPU_TARGETS" =~ gfx9[0-9a-z][0-9a-z] ]]; then
    rm -rf /opt/rocm-*/lib/*/library/*gfx9*
  fi

I can create a pull request for this

[nikAizuddin]

I think I'm no longer interested with ROCm build because I've found that Vulkan is faster than ROCm #705 (comment) at least on my Vega 56 GPU. So, I guess I'll close this issue.