[Bugfix] fix qwen2.5-vl-72b shape ERROR during the _prepare_inputs phase under high concurrency.

[Levi-JQ]

What this PR does / why we need it?

qwen2.5-vl-72b reports a shape ERROR during the _prepare_inputs phase under high concurrency【 issue #4430 】

This PR fix it.

The related PR in main branch :#3612

The related commit in vllm : https://github.com/vllm-project/vllm/blob/17c540a993af88204ad1b78345c8a865cf58ce44/vllm/model_executor/models/interfaces.py

【The _get_text_embeddings function has been refactored to interfaces.pyin vllm.】

[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 aims to fix a shape error for qwen2.5-vl-72b under high concurrency by updating how multimodal embeddings are prepared. The changes involve modifying get_input_embeddings to accept an is_multimodal mask and updating the NPUModelRunner to construct and pass this mask. This aligns the codebase with newer vLLM practices for handling multimodal inputs.

My review identifies a critical issue where a necessary argument is missing in the call to get_input_embeddings, which likely leaves the original bug unfixed. The proposed change completes the fix by correctly enabling the handling of out-of-vocabulary multimodal tokens.

[gemini-code-assist]

The call to get_input_embeddings is missing the handle_oov_mm_token argument. Without it, the default value False will be used, and the logic to handle out-of-vocabulary (OOV) multimodal tokens will not be triggered. This can lead to errors when the input contains multimodal tokens with IDs outside the language model's vocabulary range, which is likely the root cause of the issue this PR aims to fix. You should pass handle_oov_mm_token=self.supports_mm_inputs to correctly handle these cases.

Suggested change

	inputs_embeds = self.model.get_input_embeddings(
	input_ids,
	multimodal_embeddings=mm_embeds,
	is_multimodal=is_mm_embed,
	)
	inputs_embeds = self.model.get_input_embeddings(
	input_ids,
	multimodal_embeddings=mm_embeds,
	is_multimodal=is_mm_embed,
	handle_oov_mm_token=self.supports_mm_inputs,
	)