Skip to content

[core] MLA performance boost for AMD GPUs and tuned MoE config for MI…#13439

Closed
qli88 wants to merge 3 commits intovllm-project:mainfrom
ROCm:mla_perf_boost_for_amd
Closed

[core] MLA performance boost for AMD GPUs and tuned MoE config for MI…#13439
qli88 wants to merge 3 commits intovllm-project:mainfrom
ROCm:mla_perf_boost_for_amd

Conversation

@qli88
Copy link
Copy Markdown
Contributor

@qli88 qli88 commented Feb 18, 2025 •
edited by github-actions bot
Loading

  1. Params tweak in MLA kernel for AMD GPUs to improve performance;
  2. Tuned MoE config for DS v3/r1 on MI300X

@github-actions
Copy link
Copy Markdown

github-actions bot commented Feb 18, 2025

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

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 either: Add ready label to the PR or enable auto-merge.

🚀

@qli88 qli88 force-pushed the mla_perf_boost_for_amd branch from 7e481ee to 7a76f70 Compare February 18, 2025 00:22
@hongxiayang hongxiayang added the rocm Related to AMD ROCm label Feb 18, 2025
mgoin
mgoin reviewed Feb 19, 2025
View reviewed changes
Copy link
Copy Markdown
Member

@mgoin mgoin left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM with a small fix needed, cc @LucasWilkinson

@qli88 qli88 force-pushed the mla_perf_boost_for_amd branch 2 times, most recently from 70a9795 to 4f2422a Compare February 19, 2025 20:00
hongxiayang
hongxiayang reviewed Feb 20, 2025
View reviewed changes
qli88 and others added 3 commits February 20, 2025 16:31
@qli88
[core] MLA performance boost for AMD GPUs and tuned MoE config for MI…
c3c4997 
…300X

Signed-off-by: qli88 <qiang.li2@amd.com>
@qli88
remove the redundant line
a77294e 
Signed-off-by: qli88 <qiang.li2@amd.com>
@qli88 @hongxiayang
Update vllm/attention/ops/triton_decode_attention.py
148e877 
Co-authored-by: Hongxia Yang <62075498+hongxiayang@users.noreply.github.com>
Signed-off-by: qli88 <qiang.li2@amd.com>
@qli88 qli88 force-pushed the mla_perf_boost_for_amd branch from c131305 to 148e877 Compare February 20, 2025 16:32
@hongxiayang
Copy link
Copy Markdown
Collaborator

hongxiayang commented Feb 21, 2025

cc @houseroad

hongxiayang
hongxiayang reviewed Feb 21, 2025
View reviewed changes
vllm/attention/ops/triton_decode_attention.py
Comment on lines 181 to +184
BLOCK = 64
if is_hip_:
BLOCK = 8

Copy link
Copy Markdown
Collaborator

@hongxiayang hongxiayang Feb 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
BLOCK = 64
if is_hip_:
BLOCK = 8
BLOCK = 64 if not is_hip_ else 8

@qli88 qli88 closed this Feb 23, 2025
@houseroad
Copy link
Copy Markdown
Collaborator

houseroad commented Feb 23, 2025

why do we close the PR?

@qli88
Copy link
Copy Markdown
Contributor Author

qli88 commented Feb 23, 2025 •
edited
Loading

why do we close the PR?

@houseroad I created a new PR to adapt the commit landed yesterday (#12639). Please take a look at that one (#13718)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@mgoin mgoin mgoin left review comments

+1 more reviewer

@hongxiayang hongxiayang hongxiayang approved these changes

Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

rocm Related to AMD ROCm

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@qli88 @hongxiayang @houseroad @mgoin