[Kernel] Integrate DeepGEMM dense block fp8

[mgoin]

WIP since the performance is not better on my systems (8xH100 CUDA 12.5 and 8xH200 CUDA 12.4) compared to our CUTLASS kernel, and I'm unsure how we can distribute DeepGEMM. See #13917 for microbenchmarks

Setup

I installed DeepGEMM in a parallel directory to vLLM, like so

git clone --recursive https://github.com/deepseek-ai/DeepGEMM
cd DeepGEMM
python setup.py install
uv pip install -e .

Usage

E2E evaluations/benchmarks on GSM8k with 8xH200 and CUDA 12.4:

Default (using our existing CUTLASS Block FP8):

lm_eval --model vllm --model_args pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto
...
Processed prompts: 100%|████████████████████████████████████| 1319/1319 [02:53<00:00,  7.59it/s, est. speed input: 6621.27 toks/s, output: 775.88 toks/s]
Running generate_until requests: 100%|██████████████████████| 1319/1319 [02:54<00:00,  7.57it/s]
vllm (pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9492|±  | 0.006|
|     |       |strict-match    |     5|exact_match|↑  |0.9492|±  | 0.006|

DeepGEMM:

VLLM_USE_DEEPGEMM=1 lm_eval --model vllm --model_args pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto
...
Processed prompts: 100%|████████████████████████████████████| 1319/1319 [02:52<00:00,  7.63it/s, est. speed input: 6654.56 toks/s, output: 778.43 toks/s]
Running generate_until requests: 100%|██████████████████████| 1319/1319 [02:53<00:00,  7.60it/s]
vllm (pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  | 0.95|±  | 0.006|
|     |       |strict-match    |     5|exact_match|↑  | 0.95|±  | 0.006|

[github-actions]

👋 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.

🚀

[houseroad]

Curious, do we see perf wins here?

[mgoin]

@houseroad it seems worse at small M but better at large M compared to our CUTLASS kernels, however this is only true for specific shapes. I need to do more careful benchmarking but in the above lm-eval gsm8k throughput test it seems to be ~same

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @mgoin.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[li2haipeng]

Hi, do we plan to merge this PR? We did a benchmark on 8xH200 and we saw ~5% otps gain on Deepseek-r1. Input len=1600, outputlen=600 and bs=1/4/8, tp=8.

[houseroad]

fi we do see perf wins, we can pick this PR up. cc: @chenyang78

[chenyang78]

fi we do see perf wins, we can pick this PR up. cc: @chenyang78

Thanks for the heads-up. I will look into this.