[Bug]: Qwen/Qwen1.5-MoE-A2.7B-Chat模型启动报错

[Luo-Jinyan]

Your current environment

npu910B && vllm-ascend0.7.1rc1

Your output of above commands here

🐛 Describe the bug

报错信息

 File "vllm_ascend/ops/fused_moe.py", line 152, in forward_oot
    topk_weights, topk_ids = group_topk(
  File "vllm_ascend/ops/fused_moe.py", line 49, in group_topk
    torch_npu.npu_group_topk(input=scores, out=scores, group_num=num_expert_group, k=topk_group)
  File "venv/lib/python3.10/site-packages/torch/_ops.py", line 1116, in __call__
    return self._op(*args, **(kwargs or {}))
RuntimeError: npu::npu_group_topk() Expected a value of type 'int' for argument 'k' but instead found type 'NoneType'.

这个报错应该是因为在调用group_topk送了topk_group=None导致的，不确定这里是对支持的MoE模型有要求，还是其他原因.

[MengqingCao]

Thanks for this bug report, we indeed need a arg transfer if it is None.

However we still need to wait for the next CANN version to support MOE, including qwen moe and deepseek v3, etc.

FYI, #72 (comment)