[0.13.0][Bugfix] fix pcp aclgraph qwen FIA bug

[weiguihua2]

What this PR does / why we need it?

[releases/v0.13.0] In the pcp full graph Qwen model scenario, the inconsistency between the Q shape and actual q len of the FIA operator is fixed.
PR for main branch: #6037

Does this PR introduce any user-facing change?

No

How was this patch tested?

vLLM version: v0.13.0

[gemini-code-assist]

Code Review

This pull request aims to fix a bug related to actual_seq_lengths_q in the update_attn_dcp_pcp_params function. While removing the incorrect slicing logic is correct, removing the padding logic as well might introduce a new issue. The query tensor's shape is tied to runtime_shape, and actual_seq_lengths_q needs to have a matching length. I've provided a suggestion to re-introduce the padding logic without the incorrect slicing to prevent potential runtime errors due to shape mismatch. Please review the suggestion.

[gemini-code-assist]

This change removes the padding logic for actual_seq_lengths_q, which could lead to a shape mismatch at runtime. It's implied that attn_metadata.actual_seq_lengths_q should now have a length equal to runtime_shape.

However, the existing test test_update_attn_dcp_pcp_params suggests this might not always be the case (runtime_shape=4 vs. len(actual_seq_lengths_q)=2). This would cause a mismatch with the query tensor's shape inside the npu_fused_infer_attention_score operator, leading to a runtime error. The test only passes because the operator is mocked.

While the original slicing was likely buggy, the padding seems necessary. The suggestion below reintroduces the padding without the incorrect slicing.

Suggested change

	actual_seq_lengths_q = attn_metadata.actual_seq_lengths_q
	actual_seq_lengths_q = list(attn_metadata.actual_seq_lengths_q)
	pad_len = runtime_shape - len(actual_seq_lengths_q)
	if pad_len > 0:
	actual_seq_lengths_q.extend([actual_seq_lengths_q[-1]] * pad_len)