[Bugfix]: Pass scaling args to mc2

[jianzs]

Pass expert_scale and expand_scale args to the dispatch and combine functions.

btw, I think the current code for mc2 only works when graph mode is enabled.

[Copilot]

Pull Request Overview

This PR ensures that the newly computed scaling factors are correctly passed into the MC2 dispatch and combine routines for dynamic expert routing.

• Pass expert_scales into the initial npu_moe_distribute_dispatch call
• Unpack and capture expand_scales from the dispatch output
• Provide expand_scales to the subsequent npu_moe_combine invocation

Comments suppressed due to low confidence (4)

vllm_ascend/quantization/w8a8_dynamic.py:132

• The function’s docstring should be updated to document the new expert_scales and expand_scales parameters so their roles are clear to future maintainers.

def fused_experts_with_mc2(

vllm_ascend/quantization/w8a8_dynamic.py:164

• [nitpick] Unpacking with a placeholder _ and a magic slice range can be hard to follow—consider naming the sixth element or adding a comment to clarify what it represents.

expand_x, dynamic_scale, expand_idx, expert_token_nums, ep_recv_counts, _, expand_scales = output[0:7]

vllm_ascend/quantization/w8a8_dynamic.py:135

• There should be unit or integration tests verifying that expert_scales and expand_scales are correctly applied in the MC2 routines to prevent regressions.

"expert_scales": topk_weights.to(torch.float32),

vllm_ascend/quantization/w8a8_dynamic.py:135

• [nitpick] Calling .to(torch.float32) inside the hot path may incur extra device transfers; consider precomputing or caching the float32 conversion if this function is called repeatedly.

"expert_scales": topk_weights.to(torch.float32),