Skip to content

Sm100 blockwise fp8 swap ab #18564

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from

Conversation

IwakuraRein
Copy link
Contributor

@IwakuraRein IwakuraRein commented May 22, 2025

Add the swap and transpose tensor A/B path to sm100 block-wise fp8 GEMM. This reduces the overhead of extra paddings.

Also add a basic heuristic.

Note: looks like 18778 is duplicate with part of this PR.

Copy link

👋 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.

🚀

@IwakuraRein IwakuraRein force-pushed the sm100-blockwise-fp8-swap-ab branch from 559e27d to 83cea35 Compare May 24, 2025 21:27
@mgoin
Copy link
Member

mgoin commented May 28, 2025

Thanks @IwakuraRein, #18778 should be separate as it only touches scaled_mm_sm100_fp8_dispatch.cuh, not the blockwise fp8 path. Would appreciate help tuning for that file though, the cutlass profiler is giving me problems going to larger values.

Copy link
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.

The heuristics are fairly difficult to read now due to separating the tile scalars from the MmaTileShape, but overall these changes seem reasonable. Could you share an accuracy eval as a smoke test since we don't have sm100 in CI?

@mgoin mgoin added performance Performance-related issues ready ONLY add when PR is ready to merge/full CI is needed labels May 28, 2025
@IwakuraRein
Copy link
Contributor Author

@mgoin Hi. I have run the tests on following problem shapes and compared to the triton kernel triton_fp8_fp8_fp16_scaled_mm_blockwise. The results are all correct:

M N K
1-15, 64, 128, 172, 256, 384, 448 2112 7168
1-15, 64, 128, 172, 256, 384, 448 24576 1536
1-15, 64, 128, 172, 256, 384, 448 32768 512
1-15, 64, 128, 172, 256, 384, 448 7168 16384
1-15, 64, 128, 172, 256, 384, 448 4096 7168
1-15, 64, 128, 172, 256, 384, 448 7168 2048

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
performance Performance-related issues ready ONLY add when PR is ready to merge/full CI is needed
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants