-
-
Notifications
You must be signed in to change notification settings - Fork 5.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[ROCm][AMD][Model] llama 3.2 support upstreaming #12421
[ROCm][AMD][Model] llama 3.2 support upstreaming #12421
Conversation
Signed-off-by: Aleksandr Malyshev <[email protected]>
Signed-off-by: Aleksandr Malyshev <[email protected]>
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can do one of these:
🚀 |
Signed-off-by: Aleksandr Malyshev <[email protected]>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for contributing. Added some comments for this PR.
vllm/model_executor/models/mllama.py
Outdated
i, | ||
) | ||
elif self.attn.backend in (_Backend.XFORMERS, _Backend.TORCH_SDPA): | ||
if current_platform.is_rocm(): |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can you merge this code path with if self.attn.backend in (_Backend.FLASH_ATTN, _Backend.FLASH_ATTN_VLLM_V1):
?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
normally I would not change code as purpose of this PR it is to match rocm repo and upstream in order not to bring more problems. But code seems identical in two branches and no reason to keep it both. Thank you for it!
Signed-off-by: Aleksandr Malyshev <[email protected]>
LGTM, can you merge from main to fix the CI failures? |
Signed-off-by: Aleksandr Malyshev <[email protected]>
@DarkLight1337 do you have any other concerns about this change? would be nice to land it asap to prevent merge conflicts. Thanks |
I have started the merge process. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM, thanks for your patience!
PR to propagate multimodal llama3.2 support into upstream for rocm arch