[bugfix][npugraph_ex]add the extra check for allreduce rmsnorm fusion pass

[ChenCangtao]

What this PR does / why we need it?

Allreduce rmsnorm fusion pass has an additional check condition, which requires fusion of the Fx graph only when the start of compile_range is greater than 512. We previously overlooked this check.

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.14.1
• vLLM main: vllm-project/vllm@dc917cc

[gemini-code-assist]

Code Review

This pull request adds a missing check to the allreduce rmsnorm fusion pass to ensure it only runs when compile_range.start is above a certain threshold. The implementation correctly adds this check to the extra_check function for the torchair pattern. However, I've identified a critical issue where an existing stream scope check is replaced instead of being augmented, which could lead to incorrect fusions. I've provided a suggestion to combine both checks. Additionally, I've noted a code duplication issue that should be addressed to improve maintainability.

[gemini-code-assist]

This new function introduces the intended check, but there are a couple of concerns:

1. (Critical) The PR description mentions adding an additional check, but the current implementation replaces the existing extra_stream_scope_check. This check prevents fusion across different streams and is likely a necessary safety guardrail that should be preserved. It seems both checks are required. This function should probably also call extra_stream_scope_check, which would require re-adding its import.

2. (High) The condition compile_range.start > ALLREDUCE_NORM_FUSE_THREHOLD is also present in GraphEXMatmulAllReduceAddRMSNormPass.is_applicable_for_range at the end of this file. This code duplication can lead to maintenance issues. It would be best to extract this logic into a shared helper function to be used in both places.

My suggestion below addresses the critical issue. Please also consider refactoring to address the code duplication.

Suggested change

	def extra_check_for_allreduce_rmsnorm_fusion_pass(match: Match) -> bool:
	compile_range = get_pass_context().compile_range
	return compile_range.start > ALLREDUCE_NORM_FUSE_THREHOLD
	def extra_check_for_allreduce_rmsnorm_fusion_pass(match: Match) -> bool:
	compile_range = get_pass_context().compile_range
	return extra_stream_scope_check(match) and compile_range.start > ALLREDUCE_NORM_FUSE_THREHOLD

[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.