[Bugfix] Fix aclgraph sizes capture error in Qwen3 Moe model case

[lilinsiman]

What this PR does / why we need it?

1. Solved the problem that in the Qwen3 Moe model case, opening DP would use an extra stream, causing ACLgraph sizes capture error
2. After experimentation, it was found that in many cases, some operators would occupy more streams than expected. Therefore, the buffer area for streams in ACLgraph was not large enough. After discussion, extra 120 streams were added as buffer.

Does this PR introduce any user-facing change?

no

How was this patch tested?

ut

• vLLM version: main
• vLLM main: vllm-project/vllm@0ae43db

[gemini-code-assist]

Code Review

This pull request addresses an ACLgraph size capture error for MoE models with data parallelism by adjusting the stream buffer size and accounting for extra stream usage in MoE models. The changes look reasonable, but I've identified a potential issue in the newly introduced is_moe_model function. The current implementation for detecting MoE models is too broad and could lead to false positives. I've suggested a more robust implementation to improve accuracy.

[gemini-code-assist]

The current implementation to detect a Mixture-of-Experts (MoE) model is based on checking if 'experts' is a substring in any of the configuration keys. This approach is fragile and might lead to false positives if a non-MoE model has a configuration key containing this substring (e.g., use_shared_experts). This could cause incorrect logic to be applied, potentially leading to suboptimal performance or runtime errors.

A more robust approach would be to check for specific attributes that are characteristic of a MoE architecture, such as num_experts and num_experts_per_tok, and ensure they have meaningful values. This will make the detection more accurate and prevent incorrect behavior.

Suggested change

	def is_moe_model(vllm_config: VllmConfig):
	config = vllm_config.model_config.hf_config
	return any('experts' in key.lower() for key in config.to_dict())
	def is_moe_model(vllm_config: VllmConfig):
	config = vllm_config.model_config.hf_config
	# A more robust check for MoE models by verifying specific attributes
	# that are characteristic of such architectures.
	return (hasattr(config, "num_experts") and
	isinstance(config.num_experts, int) and config.num_experts > 0 and
	hasattr(config, "num_experts_per_tok"))

[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.

[MengqingCao]

we still need to check if is moe model right?

[yiz-liu]

Yeah, for MoE, we'll have to do prepare and finalize around quant_method.apply, which will have extra communication.

[MengqingCao]

ut failed due to some known issues, which are unrelated to this pr. Let's merge this after the full e2e test passed

[Yikun]

Please rewrite the PR title as a meaningful one. I didn't get any human readable info from original title.

[lilinsiman]

Please rewrite the PR title as a meaningful one. I didn't get any human readable info from original title.

Already done

[momo609]

Also update the note.