[P/D] force with_prefill true after allreduce in kv producer (vllm-project#3768)

### What this PR does / why we need it?
force with_prefill true after allreduce in kv producer

- vLLM version: v0.11.0rc3
- vLLM main:
vllm-project/vllm@c9461e0

---------

Signed-off-by: liziyu <[email protected]>
Signed-off-by: nsdie <[email protected]>

Author: liziyu179

vllm_ascend/distributed/mooncake_layerwise_connector.py

@@ -1151,7 +1151,7 @@ def save_kv_layer(self, layer_name: str, kv_layer: Tuple[torch.Tensor,
                       connector_metadata: MooncakeLayerwiseConnectorMetadata,
                       **kwargs) -> None:
         """MooncakeLayerwiseConnector does not save explicitly."""
-        if self.kv_role == 'kv_producer':
+        if self.kv_role == 'kv_producer' and connector_metadata.request.keys():
             if self.pd_head_ratio != 1:
                 if self.current_layer != 0:
                     self.completion_event.wait()

vllm_ascend/worker/model_runner_v1.py

@@ -2645,6 +2645,10 @@ def _dummy_run(
             tp_size = self.vllm_config.parallel_config.tensor_parallel_size
             num_tokens = math.ceil(num_tokens / tp_size) * tp_size
 
+        # Force dummy run on prefill stage when this node is deemed as kv producer.
+        if self.is_kv_producer and not self.is_kv_consumer:
+            with_prefill = True
+
         # Padding for DP
         (num_tokens, num_tokens_across_dp,
          with_prefill) = self._sync_metadata_across_dp(num_tokens,
@@ -2693,10 +2697,6 @@ def _dummy_run(
         num_scheduled_tokens = np.array(num_scheduled_tokens_list,
                                         dtype=np.int32)
 
-        # Force dummy run on prefill stage when this node is deemed as kv producer.
-        if self.is_kv_producer and not self.is_kv_consumer:
-            with_prefill = True
-
         if not self.in_profile_run and self.dynamic_eplb:
             self.eplb_updator.forward_before()