Fix Docker ROCM builds, use AMDGPU_TARGETS instead of GPU_TARGETS #9641
Conversation
github-actions
bot
added
the
devops
|
Seems AMDGPU_TARGETS will be deprecated but not for now. |
|
Is this change required? Any comment from people experienced with ROCm would be appreciated. I don't think there are any maintainers working with ROCm at the moment, so we depend on feedback from the community to review these changes. |
|
Also I just noticed that similar issue happened when I try build with multiple gfx IDs in |
Hello I started to use it, and it required due it does not work properly (even prebuilt containers does not work properly from here - https://github.com/ggerganov/llama.cpp/blob/master/docs/docker.md) also documentation describe AMDGPU_TARGETS not GPU_TARGETS - see https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md#hipblas |
|
When I build this project with my 7900XTX, I use: Not sure, if all the parameters are still needed, but I also use AMDGPU_TARGETS, and it works |
|
Thank you all! |
…erganov#9641) * Fix Docker ROCM builds, use AMDGPU_TARGETS instead of GPU_TARGETS * Set ROCM_DOCKER_ARCH as string due it incorrectly build and cause OOM exit code
…erganov#9641) * Fix Docker ROCM builds, use AMDGPU_TARGETS instead of GPU_TARGETS * Set ROCM_DOCKER_ARCH as string due it incorrectly build and cause OOM exit code
…erganov#9641) * Fix Docker ROCM builds, use AMDGPU_TARGETS instead of GPU_TARGETS * Set ROCM_DOCKER_ARCH as string due it incorrectly build and cause OOM exit code
Small description: When use GPU_TARGETS it crash like out of memory, but with AMDGPU_TARGETS it properly works. Also in documentations described AMDGPU_TARGETS not GPU_TARGETS