Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

AMD ROCm support #5

Merged
merged 9 commits into from
Jan 26, 2024
Merged

Conversation

IlyasMoutawwakil
Copy link
Contributor

@IlyasMoutawwakil IlyasMoutawwakil commented Jan 22, 2024

  • Adapt setup.py
  • Adapt build.yaml workflow

@IlyasMoutawwakil IlyasMoutawwakil changed the title ROCm support AMD ROCm support Jan 22, 2024
@casper-hansen
Copy link
Owner

I have no AMD ROCm GPUs and no way of testing this. Is it correctly understood that it's exclusively ExLlama kernels that will work with AMD GPUs on ROCm 5 (but not 6)?

@IlyasMoutawwakil
Copy link
Contributor Author

IlyasMoutawwakil commented Jan 22, 2024

I tested on ROCm 5.6.1
@seungrokj do these changes build on both 5 and 6 without impacting performance ?
If so I can add them in this PR.

@seungrokj
Copy link

I tested on ROCm 5.6.1 @seungrokj do these changes build on both 5 and 6 without impacting performance ? If so I can add them in this PR.

@IlyasMoutawwakil I just checked AutoGPTQ/AutoGPTQ#515 in rocm5.7 env. It is compiled and there should be no perf issue because they are basically same syntaxes. (Somehow the old syntax is not compatible with rocm6.0 though)

@IlyasMoutawwakil
Copy link
Contributor Author

@casper-hansen for the other part of your question, yes for now Exllama is going to be the only way to run AWQ model on AMD GPUs. Untill we get a hipifiable, ROCm native or performant triton GEMM.

@casper-hansen casper-hansen merged commit 3aed4bf into casper-hansen:main Jan 26, 2024
# for free to join this conversation on GitHub. Already have an account? # to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants