[Feature] Enhance all-reduce skipping logic for MoE models in NPUModelRunner (#5329)

Besides enabling `recompute_scheduler_enable`, we can skip all_reduce
when max_num_batched_tokens is below mc2's requirement.

- vLLM version: release/v0.13.0
- vLLM main:
bc0a5a0c08

---------

Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>

vllm_ascend/worker/model_runner_v1.py

@@ -398,24 +398,41 @@ class NPUModelRunner(GPUModelRunner):
         return self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE and self.compilation_config.mode == CompilationMode.VLLM_COMPILE and not self.model_config.enforce_eager
 
     def _skip_all_reduce_acorss_dp_group(self) -> bool:
-        # NOTE: 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
+        Decide whether to skip the all-reduce across the data-parallel (DP) group.
-        # num_tokens DOES NOT exceed mc2_tokens_capacity which means that moe_comm_method
+
-        # of each rank is MC2. For dense models, skipping all_reduce is not necessary
+        Skipping is only applicable for MoE models and only on ranks that act as
-        # since collective-communication is not time-consuming since dp_size in dense
+        KV consumers. We skip the DP all-reduce when either:
-        # model deployments is always small and can be overlapped by async scheduling.
+        - Both the prefill and decode communication methods are MC2 (or FUSED_MC2), or
-        if not is_moe_model(self.vllm_config):
+        - Decode requires MC2 and ascend_config.recompute_scheduler_enable is True.
+        """
+        # Only applicable to MoE models and KV consumer ranks.
+        if not is_moe_model(self.vllm_config) or not self.is_kv_consumer:
             return False
+
+        def needs_mc2(num_tokens: int) -> bool:
+            return select_moe_comm_method(num_tokens, self.vllm_config) in {
+                MoECommType.MC2, MoECommType.FUSED_MC2
+            }
+
+        # Determine whether decode must use MC2. Use max cudagraph capture size
+        # if available, otherwise use the maximal uniform decode token count.
         if self.compilation_config.cudagraph_capture_sizes:
-            potential_max_num_tokens = self.compilation_config.max_cudagraph_capture_size
+            potential_max_tokens = self.compilation_config.max_cudagraph_capture_size
         else:
-            potential_max_num_tokens = self.max_num_reqs * self.uniform_decode_query_len
+            potential_max_tokens = self.max_num_reqs * self.uniform_decode_query_len
-        # To ensure skipping all_reduce across dp group is valid, we need to ensure that
+        decode_must_use_mc2 = needs_mc2(potential_max_tokens)
-        # moe_comm_method of each rank is MC2 and recomputation would never happen in D
+
-        # nodes. So here we check whether recompute_scheduler_enable is True.
+        # For prefill, use the scheduler's max_num_batched_tokens for a single
-        return self.is_kv_consumer and self.ascend_config.recompute_scheduler_enable and select_moe_comm_method(
+        # batch.
-            potential_max_num_tokens,
+        prefill_must_use_mc2 = needs_mc2(
-            self.vllm_config) in {MoECommType.MC2, MoECommType.FUSED_MC2}
+            self.vllm_config.scheduler_config.max_num_batched_tokens)
+
+        # Skip all-reduce if decode requires MC2 and either prefill also
+        # requires MC2 or recompute-based scheduler is enabled.
+        return decode_must_use_mc2 and (
+            prefill_must_use_mc2
+            or self.ascend_config.recompute_scheduler_enable)
 
     def _sync_metadata_across_dp(
             self, num_tokens: int,