qwen3_moe/qwen25 support torchair graph

[NicholasTao]

What this PR does / why we need it?

Added support for the TorchAir graph mode in qwen3_moe and qwen2.5

Does this PR introduce any user-facing change?

No

How was this patch tested?

llm = LLM(
    model=model,
    tensor_parallel_size=GPUs_per_dp_rank,
    enforce_eager=False,
    enable_expert_parallel=True,
    max_model_len=4096,
    max_num_seqs=16,
    trust_remote_code=trust_remote_code,
    gpu_memory_utilization=0.4,
    additional_config={
             "torchair_graph_config": {
                 "enabled": True,
                 "use_cached_graph": False,
                 "graph_batch_sizes_init": False,
                 "graph_batch_sizes": [16]
             },
             "ascend_scheduler_config": {
                 "enabled": True,
                 "chunked_prefill_enabled":True,
             },
             "refresh": True,
    },
)

• vLLM version: v0.10.0
• vLLM main: vllm-project/vllm@f729023

[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 adds support for DeepSeek models on the torchair backend, involving significant refactoring and the addition of new model implementation files. My review has identified two critical bugs in tensor padding logic which could lead to runtime errors in distributed environments. Additionally, there is a high-severity issue with an incorrect type hint in a method signature that could cause confusion and runtime errors. The remaining refactoring of attention backends and model registration appears to be in order.

[gemini-code-assist]

The padding logic here is incorrect. nn.functional.pad does not support negative values for adding padding; negative values are for cropping. The expression -num_tokens % self.tp_size will be negative if num_tokens is not divisible by self.tp_size, which will likely cause a runtime error. The intention is to pad the tensor to have a number of tokens divisible by self.tp_size for the reduce_scatter operation. The correct padding amount should be (self.tp_size - num_tokens % self.tp_size) % self.tp_size.

            if is_force_scatter and num_tokens % self.tp_size:
                pad_size = (self.tp_size - num_tokens % self.tp_size) % self.tp_size
                output_parallel = nn.functional.pad(
                    output_parallel, (0, 0, 0, pad_size))

[gemini-code-assist]

The padding logic is incorrect. nn.functional.pad with a negative padding value will either raise an error or crop the tensor, which is not the intended behavior. The goal is to pad residual so its number of tokens is divisible by tp_size. The correct padding amount should be calculated as (tp_size - num_tokens % tp_size) % tp_size.

            if num_tokens % tp_size:
                pad_size = (tp_size - num_tokens % tp_size) % tp_size
                residual = nn.functional.pad(residual,
                                             (0, 0, 0, pad_size))

[gemini-code-assist]

The type hint for kv_caches is torch.Tensor, but it is used as a list (kv_caches[current_step_idx]). This contradicts the usage, the base class DeepSeekMultiTokenPredictor which defines it as List[torch.Tensor], and the caller TorchairDeepSeekMTP which passes an Optional[List[torch.Tensor]]. To prevent potential runtime errors and improve code clarity, the type hint should be corrected to Optional[List[torch.Tensor]].

Suggested change

	kv_caches: torch.Tensor,
	kv_caches: Optional[List[torch.Tensor]],

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[wangxiyuan]

I'll merge this in quick once lint CI passed