[cherry-pick pr-4254] bugfix for mtp>1 when lm_head_tp>1

[zouyida2052]

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?

[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.

[gemini-code-assist]

Code Review

This pull request addresses a bug in speculative decoding with multi-token prediction (MTP) and language model head tensor parallelism (lm_head_tp > 1), where an incorrect number of compute_logits calls in the dummy run caused the model execution to hang. The fix introduces a dummy_compute_logits function that is passed down to the proposers' dummy_run methods and called within the speculative loop, ensuring the dummy run accurately reflects the actual execution flow. The changes also include related correctness fixes for tensor slicing and device synchronization. My review suggests a performance optimization to avoid an unnecessary host-to-device memory copy during the dummy run by creating a tensor directly on the target device.

[gemini-code-assist]

The dummy_indices tensor is created on the CPU by default and is then implicitly transferred to the NPU when used for indexing the hidden_states tensor. This host-to-device copy occurs in each dummy run, which can be inefficient. To improve performance, dummy_indices should be created directly on the NPU.

Suggested change

	dummy_indices = torch.zeros(max_num_reqs_across_dp,
	dtype=torch.int32)
	dummy_indices = torch.zeros(max_num_reqs_across_dp,
	dtype=torch.int32, device=self.device)

[github-actions]

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