Fix of DeepSeek Error in KV Pool Mixed Deployment Scenario (#3087)

### What this PR does / why we need it?
A new kv_role "kv_both" is added to run mixed deployment scenarios. The
mixed deployment will involve a decode phase, where with_prefill should
be false.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.10.2
- vLLM main:
c60e6137f0

Signed-off-by: fems14 <1804143737@qq.com>

vllm_ascend/distributed/mooncake/mooncake_engine.py

@@ -119,7 +119,7 @@ class MooncakeEngine:
 
         if self.use_layerwise:
             self.get_event = threading.Event()
-            if self.kv_role == 'kv_producer':
+            if self.kv_role in ['kv_producer', 'kv_both']:
                 ready_event_sending = threading.Event()
                 self.kv_send_thread = KVCacheStoreLayerSendingThread(
                     self.tp_rank, self.tp_size, self.m_store,
@@ -135,7 +135,7 @@ class MooncakeEngine:
             self.kv_recv_thread.start()
             ready_event.wait()
         else:
-            if self.kv_role == 'kv_producer':
+            if self.kv_role in ['kv_producer', 'kv_both']:
                 ready_event_sending = threading.Event()
                 self.kv_send_thread = KVCacheStoreSendingThread(
                     self.tp_rank, self.tp_size, self.m_store,
@@ -429,7 +429,7 @@ class MooncakeEngine:
         done_sending = (
             self.kv_send_thread.
             get_and_clear_finished_requests(  # type: ignore[union-attr]
-            ) if self.kv_role == 'kv_producer' else set())
+            ) if self.kv_role in ['kv_producer', 'kv_both'] else set())
         done_recving = self.kv_recv_thread.get_and_clear_finished_requests(  # type: ignore[union-attr]
         )
 

vllm_ascend/worker/model_runner_v1.py

@@ -2406,7 +2406,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
                                         dtype=np.int32)
 
         # Force dummy run on prefill stage when this node is deemed as kv producer.
-        if self.is_kv_producer:
+        if self.is_kv_producer and not self.is_kv_consumer:
             with_prefill = True
 
         attn_metadata = self._build_attention_metadata(