Skip to content

[Feature] Full Cuda Graph Support for Cutlass MLA and 6% E2E Throughput Improvement #22763

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.

Sign up for GitHub

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

Jump to bottom
Merged
merged 5 commits into from
Aug 15, 2025
Merged

[Feature] Full Cuda Graph Support for Cutlass MLA and 6% E2E Throughput Improvement #22763

merged 5 commits into from
Aug 15, 2025
+88 −2

Conversation

yewentao256
Copy link
Member

@yewentao256 yewentao256 commented Aug 12, 2025 •
edited by github-actions bot
Loading

Purpose

Support full cuda graph for cutlass MLA (SM100) and 6% E2E Throughput Improvement

Thanks for previous work of enabling cutlass MLA on SM100!

Test

vllm serve deepseek-ai/DeepSeek-V2-Lite --port 10256 --enable-expert-parallel --data-parallel-size 2 --trust_remote_code -O '{"full_cuda_graph": true}' --cuda-graph-sizes 16 32 64 128 256 512

Acc

lm_eval --model local-completions --model_args "base_url=http://127.0.0.1:10256/v1/completions,model=deepseek-ai/DeepSeek-V2-Lite,num_concurrent=256" --tasks gsm8k
# now
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.3753|±  |0.0133|
|     |       |strict-match    |     5|exact_match|↑  |0.3707|±  |0.0133|
# without cuda graph
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.3776|±  |0.0134|
|     |       |strict-match    |     5|exact_match|↑  |0.3745|±  |0.0133|

Perf

vllm bench serve --model deepseek-ai/DeepSeek-V2-Lite --dataset-name random --num-prompts 1000 --host 127.0.0.1 --port 10256 --endpoint-type openai --endpoint /v1/completions --max-concurrency 256 --random-input-len 32 --random-output-len 512
# full cuda graph
============ Serving Benchmark Result ============
Successful requests:                     1000      
Maximum request concurrency:             256       
Benchmark duration (s):                  85.27     
Total input tokens:                      30947     
Total generated tokens:                  506054    
Request throughput (req/s):              11.73     
Output token throughput (tok/s):         5934.46   
Total Token throughput (tok/s):          6297.38   
---------------Time to First Token----------------
Mean TTFT (ms):                          655.21    
Median TTFT (ms):                        734.63    
P99 TTFT (ms):                           1002.44   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          40.85     
Median TPOT (ms):                        40.07     
P99 TPOT (ms):                           46.15     
---------------Inter-token Latency----------------
Mean ITL (ms):                           40.58     
Median ITL (ms):                         35.89     
P99 ITL (ms):                            53.80     
==================================================
# without
============ Serving Benchmark Result ============
Successful requests:                     1000      
Maximum request concurrency:             256       
Benchmark duration (s):                  90.59     
Total input tokens:                      30947     
Total generated tokens:                  505342    
Request throughput (req/s):              11.04     
Output token throughput (tok/s):         5578.42   
Total Token throughput (tok/s):          5920.04   
---------------Time to First Token----------------
Mean TTFT (ms):                          846.60    
Median TTFT (ms):                        790.83    
P99 TTFT (ms):                           1818.41   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          43.14     
Median TPOT (ms):                        45.87     
P99 TPOT (ms):                           50.04     
---------------Inter-token Latency----------------
Mean ITL (ms):                           42.90     
Median ITL (ms):                         37.26     
P99 ITL (ms):                            55.26     
==================================================

@yewentao256
full-cuda-graph-support-for-cutlass-mla
f3cd295 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
@yewentao256 yewentao256 changed the title [Feature] Full Cuda Graph Support for Cutlass MLA and [Feature] Full Cuda Graph Support for Cutlass MLA and 6% E2E Throughput Improvement Aug 12, 2025
@mergify mergify bot added the v1 label Aug 12, 2025
gemini-code-assist[bot]
gemini-code-assist bot reviewed Aug 12, 2025
View reviewed changes
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot left a comment

Choose a reason for hiding this comment

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

Code Review

This pull request enables full CUDA graph support for Cutlass MLA in decode-only scenarios. The changes are minimal and correctly implemented by introducing a CutlassMLAMetadataBuilder that signals this capability. My review includes a suggestion to improve code style for better adherence to PEP 8.

@github-actions GitHub Actions
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.

🚀

@yewentao256
update pep8
0d0a52f 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
@yewentao256 yewentao256 self-assigned this Aug 12, 2025
@yewentao256 yewentao256 added the ready ONLY add when PR is ready to merge/full CI is needed label Aug 12, 2025
LucasWilkinson
LucasWilkinson approved these changes Aug 12, 2025
View reviewed changes
Copy link
Collaborator

@LucasWilkinson LucasWilkinson left a comment

Choose a reason for hiding this comment

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

LGTM! thanks for doing this!

@LucasWilkinson LucasWilkinson enabled auto-merge (squash) August 12, 2025 20:31
@mgoin mgoin added the performance Performance-related issues label Aug 12, 2025
mgoin
mgoin approved these changes Aug 12, 2025
View reviewed changes
Copy link
Member

@mgoin mgoin left a comment •
edited
Loading

Choose a reason for hiding this comment

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

😮 that is just about as clean as you can do it

Are there any unit tests we have for full cudagraph attention backends? Just thinking how we can test this over time

@robertgshaw2-redhat
Copy link
Collaborator

when it works out of the box >>

@yewentao256
add unit test
39ed2fa 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
@yewentao256
Copy link
Member Author

yewentao256 commented Aug 13, 2025 •
edited
Loading

😮 that is just about as clean as you can do it

Are there any unit tests we have for full cudagraph attention backends? Just thinking how we can test this over time

Sounds good, just adding a new unit test for it

pytest vllm/tests/compile/piecewise/test_full_cudagraph.py
-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
==================================== 2 passed, 22 skipped, 3 warnings in 119.77s (0:01:59) ====================================

@LucasWilkinson LucasWilkinson merged commit 5c3fbfe into vllm-project:main Aug 15, 2025
41 checks passed
yiliu30 pushed a commit to yiliu30/vllm-fork that referenced this pull request Aug 19, 2025
@yewentao256 @yiliu30
[Feature] Full Cuda Graph Support for Cutlass MLA and 6% E2E Throughp…
74e8f1c 
…ut Improvement (vllm-project#22763)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
divakar-amd pushed a commit to divakar-amd/vllm_upstream that referenced this pull request Aug 20, 2025
@yewentao256 @divakar-amd
[Feature] Full Cuda Graph Support for Cutlass MLA and 6% E2E Throughp…
91a8efc 
…ut Improvement (vllm-project#22763)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
djmmoss pushed a commit to djmmoss/vllm that referenced this pull request Aug 21, 2025
@yewentao256 @djmmoss
[Feature] Full Cuda Graph Support for Cutlass MLA and 6% E2E Throughp…
2e72687 
…ut Improvement (vllm-project#22763)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Duncan Moss <djm.moss@gmail.com>
epwalsh pushed a commit to epwalsh/vllm that referenced this pull request Aug 28, 2025
@yewentao256 @epwalsh
[Feature] Full Cuda Graph Support for Cutlass MLA and 6% E2E Throughp…
7c31c58 
…ut Improvement (vllm-project#22763)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
xiao-llm pushed a commit to xiao-llm/vllm that referenced this pull request Aug 28, 2025
@yewentao256 @amd-xiaoyu12
[Feature] Full Cuda Graph Support for Cutlass MLA and 6% E2E Throughp…
8a6c974 
…ut Improvement (vllm-project#22763)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Xiao Yu <xiao.yu@amd.com>
zhewenl pushed a commit to zhewenl/vllm that referenced this pull request Aug 28, 2025
@yewentao256 @zhewenl
[Feature] Full Cuda Graph Support for Cutlass MLA and 6% E2E Throughp…
b273d0a 
…ut Improvement (vllm-project#22763)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@mgoin mgoin mgoin approved these changes

@LucasWilkinson LucasWilkinson LucasWilkinson approved these changes

@WoosukKwon WoosukKwon Awaiting requested review from WoosukKwon

@robertgshaw2-redhat robertgshaw2-redhat Awaiting requested review from robertgshaw2-redhat

@njhill njhill Awaiting requested review from njhill

@ywang96 ywang96 Awaiting requested review from ywang96

@comaniac comaniac Awaiting requested review from comaniac

@alexm-redhat alexm-redhat Awaiting requested review from alexm-redhat

+2 more reviewers

@gemini-code-assist gemini-code-assist[bot] gemini-code-assist[bot] left review comments

@SageMoore SageMoore SageMoore approved these changes

Reviewers whose approvals may not affect merge requirements
Assignees

@yewentao256 yewentao256

Labels
performance Performance-related issues ready ONLY add when PR is ready to merge/full CI is needed v1
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
5 participants
@yewentao256 @robertgshaw2-redhat @mgoin @SageMoore @LucasWilkinson