[Performance] Introducing Prefix-Cached Chunked Prefill with flash-attn backend and 10% throughput gained under prompt <1K

[Juelianqvq]

Fix illegal memory access Error & Misaligned result under high concurrency in #6144
The root cause of this problem maybe the overlook of the token_chunk_size boundary check, leading to misaligned indices calculation.
The reviewers have made a mistake in that PR since offline consistency or single request 1-by-1 can be passed relatively easily but it can be wrong in high concurrency situations (batched inference + scheduler behaviour), even the prefix-caching length order matters. I use Locust to reproduce the problem.

[Performance Related]
having tested with following two extreme scenarios(vs Prefix Caching ONLY):

Llama-2-13B:
500-600 input tokens(300 cached) + 50 output tokens
reaching +15% throughput gained under 10-15 concurrency

Qwen2-72B-AWQ:
2k-6k input tokens(1k cached) + 512 output tokens
reaching +3.5% throughput gained under 10 concurrency
(Maybe Disaggregated is the better choice)

Since my GPU is not good enough, welcome to use this PR and share your results.

[Deficiency and Help Wanted]
1.Don't know why cached length (means value of len(computed blocks)) can be modified when a sequence is being split into different rounds of chunks. So a little dirty code is attached as hack here (see the comments).
2.Compatibility has not been verified in a hurry, have a try and any feedback is welcome

Hope to find out the answer together @sighingnow

cc @rkooo567 @comaniac @Yard1 @zhuohan123 @WoosukKwon

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 run fastcheck CI which consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your fast-check build on Buildkite UI.

Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).

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

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

🚀

[comaniac]

(Please ignore my previous comment as I was referring to the wrong PR)

Thanks for the investigation. I'll review the PR tomorrow.

[robertgshaw2-neuralmagic]

Awesome work to run this down!

[rkooo567]

I will take a more deeper look, but does it also work with block manager v2? we will enable that by default soon I think

[rkooo567]

remove it?

[Juelianqvq]

I will take a more deeper look, but does it also work with block manager v2? we will enable that by default soon I think

Does v2's performance exceed v1? If so I'll dig into it in my spare time.

[comaniac]

I will take a more deeper look, but does it also work with block manager v2? we will enable that by default soon I think

Does v2's performance exceed v1? If so I'll dig into it in my spare time.

v2 has the similar performance as v1, but v2 supports more features such as lookahead slot allocation which is required for speculative decoding (and other ongoing features). Since our ultimate goal is deprecating v1, supporting v2 is necessary.

[Juelianqvq]

v2 has the similar performance as v1, but v2 supports more features such as lookahead slot allocation which is required for speculative decoding (and other ongoing features). Since our ultimate goal is deprecating v1, supporting v2 is necessary.

FYI, I tried out the v2-block-manager and here is the conclusion:
① --use-v2-block-manager only
works well and achieve the similar performance as v1
② --use-v2-block-manager --enable-prefix-caching
much worse performace than v1 when outputs correctly, sometimes trigger CUDA errors, always see mismatched results halfway during pressure test
under low concurrency (i.e. 10) , like the option takes no effect since there is only very small acceleration. Moreover, it can get stuck, sometimes being preempted, sometimes triggers illegal memory access w.r.t varlen_forward kernel, some of the outputs become nonsense halfway and never return to normal any longer. Under high concurrency (i.e. 50), the same.
③ my PR which sets --use-v2-block-manager --enable-prefix-caching --enable-chunked-prefill
needs to adjust four lines at least
comment these https://github.com/vllm-project/vllm/blob/main/vllm/core/block/prefix_caching_block.py#L729-730
and append workaround code (don't know why, not familiar with v2 version)
if len(self._blocks) == first_block_idx: first_block_idx -= 1
here: https://github.com/vllm-project/vllm/blob/main/vllm/core/block/block_table.py#L152
never see messed outputs and never stuck yet.
④ I cannot guarantee that my PR and code change in ③ fix the problem in ②, it maybe just the coincidence which lowers the probability of various accidents. My suggestion is can we merge this PR into v1 first and take a deeper look on fixing v2's strange phenomena.
⑤ It seems that we seldom use pressure test with different kinds of datasets.All my dataset is 500-600 tokens with 300 of them can be cached, 30-50 output tokens on average. Current PR lacks the compatibility support of protecting prefix cache only from the combination of chunked prefill, which can also produces CUDA error. Besides that, everything looks good.

[sighingnow]

Hi @Juelianqvq, I have read the diff and if I understand it correctly, the key different between this PR with #6144 is: leaving at least 1 token for prefill for each sequence. I could add such logic into #6144.

[Juelianqvq]

@sighingnow Actually not only that, you have mentioned the correctness of keeping at least 1 tokens which you have missed before. Moreover, the modification in block_manager.py matters too. Just have a try and see whether you can have a inference speed up with only modifying keeping token logic. I've got the answer because I've developed in so many cases and pointed out the existing problem in my PR which certainly behaves faster using a work-around way.