[Bugfix] Resolve MTP > 1 issue when lm head tp > 1

[jianzs]

What this PR does / why we need it?

Previously, the dummy run executed compute_logits only once, regardless of num_speculative_tokens. This caused execute_model to hang on compute_logits when lm head tensor parallelism exceeded 1. The fix ensures compute_logits executes correctly during dummy run, matching num_speculative_tokens.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

• vLLM version: v0.11.0
• vLLM main: vllm-project/vllm@2918c1b

[gemini-code-assist]

Code Review

This pull request aims to fix an issue with speculative decoding (MTP) when tensor parallelism is used on the language model head. The core of the fix is to ensure the dummy run correctly simulates the multiple compute_logits calls that occur in a real run. While the fix is correctly applied for MtpProposer, it seems to be incomplete for EagleProposer, which could lead to the same issue in that scenario. Additionally, a refactoring in model_runner_v1.py appears to have introduced an AttributeError by calling a non-existent method on the drafter object. I've provided critical comments and suggestions for both issues.

[gemini-code-assist]

The dummy_drafter_compute_logits function calls self.drafter.compute_logits, but the compute_logits method is on the model attribute of the drafter object, not on the drafter itself. This will result in an AttributeError. The call should be self.drafter.model.compute_logits.

Suggested change

	return self.drafter.compute_logits(
	return self.drafter.model.compute_logits(

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[whx-sjtu]

We firstly fixed hanging issue of running MTP=1 with llm head tp in PR #3915. This PR refactors it to run dummy_compute_logits in drafter's dummy_run and further fixes MTP > 1 scenario. LGTM.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[zouyida2052]

I've tested it on deepseek and it proves to be useful, please make ci happy