Merge branch 'v0.9.2-dev-disagg' into 'v0.9.2-dev'

Merge v0.9.2-dev-disagg into v0.9.2-dev

See merge request dcutoolkit/deeplearing/vllm!163

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py

@@ -143,7 +143,7 @@ class P2pNcclConnector(KVConnectorBase_V1):
                 request_id (str): request id for log
             """
             dst_kv_cache_layer_shape = dst_kv_cache_layer.shape
-            if isinstance(attn_metadata, MLACommonMetadata):
+            if isinstance(attn_metadata, MLACommonMetadata) or all(isinstance(value, MLACommonMetadata) for value in attn_metadata.values()):
                 num_pages = dst_kv_cache_layer_shape[0]
                 page_size = dst_kv_cache_layer_shape[1]
                 dst_kv_cache_layer = dst_kv_cache_layer.reshape(
@@ -193,8 +193,16 @@ class P2pNcclConnector(KVConnectorBase_V1):
         # Load the KV for each request each layer
         for request in metadata.requests:
             for layer_name in forward_context.no_compile_layers:
-                attn_layer = forward_context.no_compile_layers[layer_name]
-                kv_cache_layer = attn_layer.kv_cache[ \
+                layer = forward_context.no_compile_layers[layer_name]
+
+                # Only process layers that have kv_cache
+                # attribute (attention layers) Skip non-attention
+                # layers like FusedMoE
+                kv_cache = getattr(layer, 'kv_cache', None)
+                if kv_cache is None:
+                    continue
+
+                kv_cache_layer = kv_cache[ \
                     forward_context.virtual_engine]
 
                 kv_cache = self.p2p_nccl_engine.recv_tensor(