diff --git a/tests/v1/kv_connector/unit/test_nixl_connector.py b/tests/v1/kv_connector/unit/test_nixl_connector.py index 8eaa352544..760a21c4c8 100644 --- a/tests/v1/kv_connector/unit/test_nixl_connector.py +++ b/tests/v1/kv_connector/unit/test_nixl_connector.py @@ -584,7 +584,7 @@ class TestNixlHandshake: def test_pcp_producer_uses_canonical_replica( self, default_vllm_config, dist_init, pcp_rank ): - """Only PCP rank zero publishes and reports sending completion.""" + """Only PCP rank zero publishes, but every rank reports completion.""" from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend vllm_config = create_vllm_config(kv_role="kv_producer") @@ -622,13 +622,18 @@ class TestNixlHandshake: payload = MagicMock(spec=NixlHandshakePayload) worker.xfer_handshake_metadata = payload - worker.get_finished = MagicMock(return_value=({"sent"}, set())) + worker.transfer_topo = MagicMock() + worker._get_new_notifs = MagicMock( + side_effect=lambda: {"sent"} if pcp_rank == 0 else set() + ) expected_payload = payload if pcp_rank == 0 else None assert connector.get_handshake_metadata() is expected_payload done_sending, done_recving = connector.get_finished(set()) - assert done_sending == ({"sent"} if pcp_rank == 0 else set()) + assert done_sending == ({"sent"} if pcp_rank == 0 else {req_id}) assert done_recving == set() + if pcp_rank > 0: + assert connector.get_finished(set()) == (set(), set()) @patch( "vllm.distributed.kv_transfer.kv_connector.v1.nixl.base_worker.NixlWrapper", diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py index 08a63592d3..5cb76932d4 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py @@ -656,6 +656,9 @@ class NixlBaseConnectorWorker: self._recving_transfers = defaultdict[ReqId, list[TransferHandle]](list) # Track the expiration time of requests that are waiting to be sent. self._reqs_to_send: dict[ReqId, float] = {} + # Replicated-KV PCP ranks > 0 never transfer; requests they are asked + # to send are reported finished immediately (see get_finished). + self._replicated_pcp_done_sending: set[ReqId] = set() # Set of requests that have been part of a batch, regardless of status. self._reqs_to_process: set[ReqId] = set() @@ -2543,6 +2546,9 @@ class NixlBaseConnectorWorker: done_sending = self._get_new_notifs() done_recving = self._pop_done_transfers(self._recving_transfers) + done_sending.update(self._replicated_pcp_done_sending) + self._replicated_pcp_done_sending.clear() + # Drain queue of requests where handshake or transfer setup failed. failed_recv_reqs = set[ReqId]() while not self._failed_recv_reqs.empty(): diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/connector.py b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/connector.py index ee7b991799..7914b7ab3a 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/connector.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/connector.py @@ -156,15 +156,6 @@ class NixlBaseConnector(KVConnectorBase_V1, SupportsHMA): ############################################################ def get_finished_count(self) -> int | None: - parallel_config = self._vllm_config.parallel_config - if ( - self.kv_transfer_config.kv_role == "kv_producer" - and parallel_config.prefill_context_parallel_size > 1 - ): - return ( - parallel_config.tensor_parallel_size - * parallel_config.pipeline_parallel_size - ) return None def get_num_new_matched_tokens( @@ -241,13 +232,7 @@ class NixlBaseConnector(KVConnectorBase_V1, SupportsHMA): def get_finished(self, finished_req_ids: set[str]) -> tuple[set[str], set[str]]: """Get the finished recving and sending requests.""" assert self.connector_worker is not None - done_sending, done_recving = self.connector_worker.get_finished() - if ( - self.kv_transfer_config.kv_role == "kv_producer" - and self.connector_worker.pcp_rank > 0 - ): - done_sending.clear() - return done_sending, done_recving + return self.connector_worker.get_finished() def get_block_ids_with_load_errors(self) -> set[int]: """Get block IDs that failed to load via NIXL.""" diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/pull_worker.py b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/pull_worker.py index a3bbfc636f..758f5c9fab 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/pull_worker.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/pull_worker.py @@ -80,6 +80,11 @@ class NixlPullConnectorWorker(NixlBaseConnectorWorker): self._read_blocks_for_req(*self._ready_requests.get_nowait()) if self.pcp_rank > 0: + # Replicated-KV PCP: only PCP rank 0 serves the KV, so this rank + # has nothing to send. Report the requests as sent right away so + # the scheduler-side aggregation (world_size workers, and any + # sibling connector inside a MultiConnector) still completes. + self._replicated_pcp_done_sending.update(metadata.reqs_to_send) return # Keep around the requests that have been part of a batch. This is