[Doc] Add documentations for nightly benchmarks

[KuntaiDu]

FILL IN THE PR DESCRIPTION HERE

Add documentations for nightly benchmarks.
The latest results of nightly benchmark:
https://buildkite.com/vllm/performance-benchmark/builds/4068

BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE

---

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

• [Bugfix] for bug fixes.
• [CI/Build] for build or continuous integration improvements.
• [Doc] for documentation fixes and improvements.
• [Model] for adding a new model or improving an existing model. Model name should appear in the title.
• [Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
• [Kernel] for changes affecting CUDA kernels or other compute kernels.
• [Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
• [Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
• [Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

• We adhere to Google Python style guide and Google C++ style guide.
• Pass all linter checks. Please use format.sh to format your code.
• The code need to be well-documented to ensure future contributors can easily understand the code.
• Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
• Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

• After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
• After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
• After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
• Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only trigger fastcheck CI to run, which consists only a small and essential subset of tests to quickly catch errors with the flexibility to run extra individual tests on top (you can do this by unblocking test steps in the Buildkite run).

Full CI run is still required to merge this PR so once the PR is ready to go, please make sure to run it. If you need all test signals in between PR commits, you can trigger full CI as well.

To run full CI, you can do one of these:

• Comment /ready on the PR
• Add ready label to the PR
• Enable auto-merge.

🚀

[zhyncs]

@KuntaiDu By the way, why only test the cases of qps 2 and 4, without continuing to test qps 8, 16, etc.? I have tested vLLM before. My impression is that the performance deteriorates significantly during the increase in qps. I don't know if this meets expectations.

[KuntaiDu]

@KuntaiDu By the way, why only test the cases of qps 2 and 4, without continuing to test qps 8, 16, etc.? I have tested vLLM before. My impression is that the performance deteriorates significantly during the increase in qps. I don't know if this meets expectations.

I am not testing high QPS for now, to control the benchmark duration. Currently we are only testing one QPS per model. And the benchmark is already a bit slow (it takes 3.5 hours to run).

But yeah, it is definitely worthwhile to cover more QPS if we have more resources for CI in the future.

[zhyncs]

@KuntaiDu By the way, why only test the cases of qps 2 and 4, without continuing to test qps 8, 16, etc.? I have tested vLLM before. My impression is that the performance deteriorates significantly during the increase in qps. I don't know if this meets expectations.

I am not testing high QPS for now, to control the benchmark duration. Currently we are only testing one QPS per model. And the benchmark is already a bit slow (it takes 3.5 hours to run).

But yeah, it is definitely worthwhile to cover more QPS if we have more resources for CI in the future.

make sense

[zhyncs]

the support for FP8 benchmark on H100 is coming!

For frameworks that currently do not support FP8, how will they be handled? Will Int8 be used as a substitute or left empty?

[KuntaiDu]

The serving engines that I am benchmarking (vllm, lmdeploy, tgi and trt) all support fp8 now

[zhyncs]

The serving engines that I am benchmarking (vllm, lmdeploy, tgi and trt) all support fp8 now

LMDeploy TurboMind currently supports quantization for AWQ(W4A16), KV Cache Int4, and KV Cache Int8. FP8 is not supported yet but the feature is under way. We expect to first complete the optimization of GEMM(faster than cuBLAS at small batch sizes) and support for MOE. Targeted optimization for the Hopper architecture is also being planned, and it is expected to be faster than Flash Attention 3. The support for the new features of FP8 and Hopper has not been done before because we do not have an H100 development environment. It is highly likely that cloud servers will be used to support it in the future. I highly appreciate the work of W8A8(FP8) in vLLM and the optimization work of MOE, which is very worthy for us to learn. Cheers.

[KuntaiDu]

I will start developing fp8 benchmarks after LMDeploy supports it then.

[zhyncs]

But yeah, it is definitely worthwhile to cover more QPS if we have more resources for CI in the future.

Using the SOTA within 10B parameters, namely GLM 4 9B Chat, tested on a single A100 80G. The benchmark script used is from vLLM, with 1000 prompts. LMDeploy and vLLM both use the latest version. Hi @KuntaiDu The 3.5 hours you mentioned is quite long, I understand. Perhaps we can conduct more QPS benchmarks on smaller models based on the above. As far as I know, the models deployed by large-scale domestic companies are mainly below 10B. If you have any questions or suggestions, please feel free to contact me. Thanks.

[zhyncs]

@KuntaiDu I've also done the benchmark with Qwen 2 72B Instruct with TP 4.

From the data in the graph, vLLM's prefill has a slight advantage when the request rate is low. I am quite curious because I understand that vLLM uses custom all reduce. Why is the decoding ITL so much worse? Is this expected? If necessary, I can raise an issue separately to discuss this. @youkaichao

Don't get me wrong, I'm not here to criticize vLLM completely. In fact, the current strong support of vLLM on MOE and the work of W8A8(FP8) are impressive. If we have business to use in the short term, I would also recommend them to use vLLM. Cheers.

[zhyncs]

In my impression, when the batch size was not large on FasterTransformer before, custom all reduce was much better than NCCL.

[youkaichao]

I'm not surprised. the decoding phase has lots of overhead, the full scheduling and batching, sampling logic is written in python, which is slow. that's one of our recent goal to improve. welcome to contribute :)

[KuntaiDu]

@KuntaiDu I've also done the benchmark with Qwen 2 72B Instruct with TP 4.

From the data in the graph, vLLM's prefill has a slight advantage when the request rate is low. I am quite curious because I understand that vLLM uses custom all reduce. Why is the decoding ITL so much worse? Is this expected? If necessary, I can raise an issue separately to discuss this. @youkaichao

Don't get me wrong, I'm not here to criticize vLLM completely. In fact, the current strong support of vLLM on MOE and the work of W8A8(FP8) are impressive. If we have business to use in the short term, I would also recommend them to use vLLM. Cheers.

Thanks a lot for the measurement, and I am not expecting such a large performance gap... You are right, we need to benchmark vllm at high QPS then. I will raise a new PR for that.