[BugFix] Fix PP performance and PP kv connector output regression (#28768)

Signed-off-by: Nick Hill <nhill@redhat.com>

vllm/v1/worker/gpu_worker.py

@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """A GPU worker class."""
 
-import copy
 import gc
 import os
 from contextlib import AbstractContextManager, nullcontext
@@ -45,7 +44,6 @@ from vllm.v1.core.sched.output import GrammarOutput
 from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
 from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
 from vllm.v1.outputs import (
-    EMPTY_MODEL_RUNNER_OUTPUT,
     AsyncModelRunnerOutput,
     DraftTokenIds,
     ModelRunnerOutput,
@@ -581,18 +579,7 @@ class Worker(WorkerBase):
             all_gather_tensors=all_gather_tensors,
         )
 
-        kv_connector_output = output.kv_connector_output
-        if not kv_connector_output:
-            return None
-
-        # In case of PP with kv transfer, we need to pass through the
-        # kv_connector_output
-        if kv_connector_output.is_empty():
-            return EMPTY_MODEL_RUNNER_OUTPUT
-
-        output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
-        output.kv_connector_output = kv_connector_output
-        return output
+        return None
 
     def take_draft_token_ids(self) -> DraftTokenIds | None:
         return self.model_runner.take_draft_token_ids()