[Bugfix][KV Connector][NIXL] Support PCP producers (#52779)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Lucas Wilkinson
2026-08-21 16:22:01 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 3e47a9a824
commit 0a3a4f7f35
9 changed files with 173 additions and 11 deletions
@@ -283,12 +283,11 @@ def test_multi_example_connector_consistency():
events = get_connector_events()
storage1_scheduler_events = _ignore_event_collection(events["storage1-SCHEDULER"])
storage2_scheduler_events = _ignore_event_collection(events["storage2-SCHEDULER"])
# First event is bind_gpu_block_pool from initialization, then
# set_xfer_handshake_metadata_pp_aware, then on_new_request when the request is
# enqueued, then get_num_new_matched_tokens and update_state_after_alloc from
# generate().
assert storage1_scheduler_events[:6] == [
# Initial events bind the block pool, query completion counts, and exchange
# handshake metadata before the request is enqueued.
assert storage1_scheduler_events[:7] == [
"bind_gpu_block_pool",
"get_finished_count",
"set_xfer_handshake_metadata_pp_aware",
"on_new_request",
"get_num_new_matched_tokens 0",
@@ -307,8 +306,9 @@ def test_multi_example_connector_consistency():
"wait_for_layer_load",
"save_kv_layer",
]
assert storage2_scheduler_events[:6] == [
assert storage2_scheduler_events[:7] == [
"bind_gpu_block_pool",
"get_finished_count",
"set_xfer_handshake_metadata_pp_aware",
"on_new_request",
"get_num_new_matched_tokens 0",
@@ -570,6 +570,58 @@ class FakeNixlConnectorWorker(NixlConnectorWorker):
class TestNixlHandshake:
@pytest.mark.parametrize("pcp_rank", [0, 1])
@patch(
"vllm.distributed.kv_transfer.kv_connector.v1.nixl.base_worker.NixlWrapper",
FakeNixlWrapper,
)
def test_pcp_producer_uses_canonical_replica(
self, default_vllm_config, dist_init, pcp_rank
):
"""Only PCP rank zero publishes and reports sending completion."""
vllm_config = create_vllm_config(kv_role="kv_producer")
vllm_config.parallel_config.prefill_context_parallel_size = 2
with (
patch(
"vllm.distributed.kv_transfer.kv_connector.v1.nixl."
"base_worker.get_current_attn_backends",
return_value=[FlashAttentionBackend],
),
patch(
"vllm.distributed.kv_transfer.kv_connector.v1.nixl."
"base_worker.get_pcp_group"
) as mock_get_pcp_group,
):
mock_get_pcp_group.return_value.rank_in_group = pcp_rank
connector = NixlConnector(
vllm_config,
KVConnectorRole.WORKER,
make_kv_cache_config(block_size=16),
)
worker = connector.connector_worker
assert worker is not None
assert worker.pcp_rank == pcp_rank
req_id = "req"
metadata = NixlConnectorMetadata()
metadata.reqs_in_batch.add(req_id)
metadata.reqs_to_send[req_id] = time.perf_counter() + 10
worker.start_load_kv(metadata)
expected_tracked = pcp_rank == 0
assert (req_id in worker._reqs_to_process) == expected_tracked
assert (req_id in worker._reqs_to_send) == expected_tracked
payload = MagicMock(spec=NixlHandshakePayload)
worker.xfer_handshake_metadata = payload
worker.get_finished = MagicMock(return_value=({"sent"}, 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_recving == set()
@patch(
"vllm.distributed.kv_transfer.kv_connector.v1.nixl.base_worker.NixlWrapper",
FakeNixlWrapper,
@@ -334,6 +334,7 @@ class _StubWriterWorker(NixlPushConnectorWorker):
w._reqs_to_send = {}
w.consumer_notification_counts_by_req = defaultdict(int)
w.tp_rank = 0
w.pcp_rank = 0
w.world_size = 1
w.engine_id = "test-decode-engine"
w._remote_agents = {}
@@ -514,6 +515,24 @@ class TestPushWriterStartLoadKv:
assert w._push_writer_wake.is_set()
assert w.start_push_calls == []
def test_noncanonical_pcp_rank_skips_producer_work(self):
w = _StubWriterWorker.fresh()
w.pcp_rank = 1
w._send_heartbeats = MagicMock()
meta = NixlConnectorMetadata()
meta.push_finished_blocks["req"] = ([1, 2],)
meta.reqs_in_batch.add("req")
meta.reqs_to_send["req"] = time.perf_counter() + 10
w.start_load_kv(meta)
assert w._finished_blocks_inbox.empty()
assert "req" not in w._reqs_to_process
assert "req" not in w._reqs_to_send
assert not w._push_writer_wake.is_set()
w._send_heartbeats.assert_not_called()
# The P→D handshake must run on the base worker's background executor, never
# blocking the writer thread: ``_do_start_push_kv`` defers the WRITE until the