fix typo and delete a part of mc2_mask

Signed-off-by: linfeng-yuan <[email protected]>

Author: linfeng-yuan

vllm_ascend/ops/fused_moe/token_dispatcher.py

@@ -227,7 +227,6 @@ def token_dispatch(self,
         context_metadata = {
             "topk_ids": topk_ids,
             "topk_weights": topk_weights,
-            "mc2_mask": mc2_mask,
             "expert_map": expert_map,
             "ep_recv_counts": ep_recv_counts,
             "tp_recv_counts": tp_recv_counts,
@@ -256,7 +255,6 @@ def get_combine_mc_kwargs(self, hidden_states: torch.Tensor,
         ep_recv_counts = context_metadata["ep_recv_counts"]
         tp_recv_counts = context_metadata["tp_recv_counts"]
         assist_info_for_combine = context_metadata["assist_info_for_combine"]
-        mc2_mask = context_metadata["mc2_mask"]
         expand_scales = context_metadata["expand_scales"]
 
         assert expert_map is not None

vllm_ascend/worker/model_runner_v1.py

@@ -968,7 +968,7 @@ def _sync_metadata_across_dp(
         # NOTE: Here we can skip the all_reduce operation and avoid paading tokens
         # to max_tokens_acrodd_dp in D nodes. In MoE models, we must ensure that
         # num_tokens DOES NOT exceed mc2_tokens_capacity which means that moe_comm_method
-        # of each rank is MC2. It is recommended to enable recompute scheduler for D Noes.
+        # of each rank is MC2. It is recommended to enable recompute scheduler for D Nodes.
         if self.is_kv_consumer and not self.in_profile_run:
             num_tokens_after_padding = torch.tensor([num_tokens] *
                                                     self.dp_size,