[P/D]kv_output_aggregator support P TP > D TP (#23917)

Signed-off-by: LCAIZJ <leichao139636@163.com>
Co-authored-by: leichao.lc <leichao.lc@antgroup.com>

vllm/v1/engine/core.py

@@ -128,6 +128,9 @@ class EngineCore:
             log_stats=self.log_stats,
         )
         self.use_spec_decode = vllm_config.speculative_config is not None
+        if self.scheduler.connector is not None:  # type: ignore
+            self.model_executor.init_kv_output_aggregator(
+                self.scheduler.connector.get_finished_count())  # type: ignore
 
         self.mm_registry = mm_registry = MULTIMODAL_REGISTRY
         self.mm_receiver_cache = engine_receiver_cache_from_config(