From 8283d7b85c7d8f485e6d02864fc24fc9334b2ad6 Mon Sep 17 00:00:00 2001 From: "rshaw@neuralmagic.com" Date: Mon, 30 Jun 2025 01:45:03 +0000 Subject: [PATCH] updated Signed-off-by: rshaw@neuralmagic.com --- .../kv_transfer/kv_connector/v1/nixl_connector.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) 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 c5ff321cd9f53..9adb561a1802e 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py @@ -333,8 +333,8 @@ class NixlConnectorWorker: self.nixl_wrapper = NixlWrapper(str(uuid.uuid4()), None, num_workers=None, - # num_shared_workers=16) # setting this > 0 causes the notifs to be recved - num_shared_workers=None) + num_shared_workers=16) # setting this > 0 causes the notifs to be recved + # num_shared_workers=None) # Map of engine_id -> {rank0: agent_name0, rank1: agent_name1..}. self._remote_agents: dict[str, dict[int, str]] = defaultdict(dict) @@ -864,7 +864,7 @@ class NixlConnectorWorker: end = time.perf_counter() print(f"========= SEND NOTIF TIME: {end - start} =========") else: - transfers[req_id] = (new_handles, notif_id, agent_name) + transfers[req_id] = (new_handles, agent_name, notif_id) return done_req_ids @@ -977,7 +977,7 @@ class NixlConnectorWorker: assert len(local_block_descs_ids) == len(remote_block_descs_ids) # Prepare transfer with Nixl. - CHUNK_SIZE = 1000 + CHUNK_SIZE = 100 handles = [] for i in range(0, len(local_block_descs_ids), CHUNK_SIZE): handles.append(