[Misc] Remove pad_for_cudagraphs from config

[LucasWilkinson]

With #28579 we can strip pad_for_cudagraph from the config forcing cudagraph padding through the dispatch consistently making it easier to debug and enable support for separate sizes for FULL and PIECEWISE graphs

[gemini-code-assist]

Code Review

This pull request refactors the CUDA graph padding logic by removing pad_for_cudagraphs from the configuration and centralizing it within the CudagraphDispatcher. The changes are well-contained and improve the code structure by making the dispatcher solely responsible for padding, which simplifies debugging and future extensions.

My review found one potential issue in vllm/v1/spec_decode/eagle.py where the refactoring might unintentionally change the behavior for EagleProposer when cudagraph_mode is set to FULL. I've left a comment with a suggested fix to ensure the original behavior is preserved. Other than that, the changes look solid and achieve the stated goal.

[gemini-code-assist]

This change might introduce a regression. The previous implementation for EagleProposer only enabled CUDA graphs if PIECEWISE mode was supported. With this change, if the main model's cudagraph_mode is FULL, cudagraph_mode.mixed_mode() will return FULL, and EagleProposer will attempt to use FULL CUDA graphs, which it may not support (as indicated by the old logic and removed warning). This could lead to runtime errors.

To preserve the original behavior and ensure only PIECEWISE graphs are used for Eagle, I suggest restoring the check for PIECEWISE support.

Suggested change

	eagle_cudagraph_mode = (
	cudagraph_mode.mixed_mode()
	if not self.speculative_config.enforce_eager
	else CUDAGraphMode.NONE
	)
	if (not self.speculative_config.enforce_eager
	and cudagraph_mode.has_mode(CUDAGraphMode.PIECEWISE)):
	eagle_cudagraph_mode = CUDAGraphMode.PIECEWISE
	else:
	eagle_cudagraph_mode = CUDAGraphMode.NONE

[mergify]

Hi @LucasWilkinson, the pre-commit checks have failed. Please run:

uv pip install pre-commit
pre-commit install
pre-commit run --all-files

Then, commit the changes and push to your branch.

For future commits, pre-commit will run automatically on changed files before each commit.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Dispatch loop never terminates in hybrid padding test

In test_mamba_cache_cg_padding a CudagraphDispatcher is created and immediately used in the while condition without ever calling initialize_cudagraph_keys, so dispatch returns a BatchDescriptor with num_tokens unchanged when keys_initialized is False. The condition len(example_prompts) == ...num_tokens therefore stays true on every iteration and the loop appends forever, hanging the test before it exercises any logic. This test now never completes under any configuration.

Useful? React with 👍 / 👎.

[mergify]

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

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