[Bugfix] Prevent crash from attempting to use a CUDA graph for a batch size that wasn't captured

[tjohnson31415]

The interaction of batch expansion in speculative decoding, with CUDA graph capture, and setting --max-num-seqs less than 256 can result in a crash due to a KeyError when the server looks for a CUDA graph batch size that is not captured.

The occurence of the crash comes from a few factors:

• CUDA graphs are created in capture_model supporting batch sizes up to max_num_seqs or 256 (_BATCH_SIZES_TO_CAPTURE[-1]) (whichever is smaller) REF
• max_num_seqs is user configurable and is intended to limit the maximum batch size
• however, with speculative decoding, batch expansion can make the internal batch size for the scorer model larger than max_num_seqs

The fix proposed in this PR is to add an additional check in _use_captured_graph() to not use a graph if the internal batch size is greater than get_graph_batch_size(max_num_seqs) so that we don't try to use a graph for a batch size that wasn't captured.

Example stacktrace:

...
  File "/workspace/my-vllm/lib64/python3.11/site-packages/vllm/spec_decode/spec_decode_worker.py", line 609, in _run_speculative_decoding_step
    proposal_scores = self.scorer.score_proposals(
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib64/python3.11/contextlib.py", line 81, in inner
    return func(*args, **kwds)
           ^^^^^^^^^^^^^^^^^^^
  File "/workspace/my-vllm/lib64/python3.11/site-packages/vllm/spec_decode/batch_expansion.py", line 85, in score_proposals
    target_sampler_output = self._scorer_worker.execute_model(
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/workspace/my-vllm/lib64/python3.11/site-packages/vllm/worker/worker_base.py", line 328, in execute_model
    output = self.model_runner.execute_model(
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/workspace/my-vllm/lib64/python3.11/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/workspace/my-vllm/lib64/python3.11/site-packages/vllm/worker/model_runner.py", line 1409, in execute_model
    model_executable = self.graph_runners[virtual_engine][
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
KeyError: 40

FIX #6306
FIX #7907

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.

🚀

[njhill]

Thanks @tjohnson31415 for this important fix! Might it be better to instead adjust the effective match batch size used in determining what graph sizes to capture upfront? I.e. multiply it by the max number of speculative tokens?

[tjohnson31415]

@njhill My understanding is that limiting CUDA graph compilation to support max_num_seqs is intended to limit the memory consumption of the graphs. I'm not sure if making that memory consumption dependent on the speculator type/configuration would be confusing or not. Of course we could add to the warning log messages to explain what is happening.

That still might be a good change, but I think it can be a separate change from this bugfix PR. Even if other parts of the code better handle the batch expansion, _use_captured_graph() should verify that a graph actually exists for the batch size requested (ideally it could check that more directly instead of replicating logic from capture_model).

[njhill]

Thanks @tjohnson31415, I agree it's good to fix this here regardless

[njhill]

Could we store the max as an instance field when capturing the graph in capture_model() (i.e. batch_size_capture_list[-1]) .. so that the check here is simpler/cheaper (it will replace both of these batch_size <= checks).

[njhill]

@njhill My understanding is that limiting CUDA graph compilation to support max_num_seqs is intended to limit the memory consumption of the graphs. I'm not sure if making that memory consumption dependent on the speculator type/configuration would be confusing or not. Of course we could add to the warning log messages to explain what is happening.

Limiting based on max_num_seqs is just because any graphs bigger than that would never be used so no point in spending the additional capture time + memory.

I don't think it would be confusing that the spec decoding config factors into this effective max batch calculation. There's anyhow a hard cap of 256, I'm not suggesting that we necessarily increase this .. i.e. that limit would still be hit if max_num_seqs was 64 and max spec tokens was 4...

[njhill]

@tjohnson31415 ah looks like this may also be covered by #7894

[youkaichao]

didn't expect #7894 conflicts with this pr 🤣

[tjohnson31415]

Ah, yep, and #7894 already includes the refactor to store the max batch size.

🚀

[njhill]

@tjohnson31415 if batch expansion means that the effective batch size can actually be max_spec_tokens * self.scheduler_config.max_num_seqs, I still think it would be worthwhile to take this into account when calculating max_batchsize_to_capture or else we may be leaving some performance on the table. Maybe we can create a separate PR for that like you said.