[BugFix] Fix ascend scheduler assert error

[fan2956]

What this PR does / why we need it?

Running multimodal model with ascend scheduler may cause assert error 【assert (request.num_tokens - request.num_computed_tokens) == 1】

Does this PR introduce any user-facing change?

No

How was this patch tested?

• vLLM version: v0.10.2
• vLLM main: vllm-project/vllm@17b4c66

[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 provides a critical fix for an assertion error that occurs when running multimodal models with the Ascend scheduler. The change correctly prevents partial prefilling of multimodal requests when their encoder inputs cannot be scheduled, which was the root cause of the issue. The fix is well-targeted and effectively ensures the stability of the scheduler for multimodal workloads. The change is correct and I approve of it.

[gemini-code-assist]

This change is critical to prevent a crash when scheduling multimodal requests. Previously, if encoder inputs couldn't be scheduled due to budget constraints, the scheduler would still attempt to prefill the text tokens. This partial prefilling would lead to an inconsistent request state, causing an assertion failure later in the decoding phase. By adding the len(encoder_inputs_to_schedule) == 0 check, you ensure that the request is deferred until all its necessary components can be scheduled, which correctly resolves the issue.