A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix][V1][P/D]Fix the uneven polling issue in the toy proxy for P2pNcclConnector (#21819)
Signed-off-by: Abatom <abzhonghua@gmail.com>
Zhonghua Deng committed
Aug 5, 2025 at 04:17 UTC
c09efff9767ad26ecf99a6e6c13243612c278df3
1 file changed
+3
-2
examples/online_serving/disaggregated_serving_p2p_nccl_xpyd/disagg_proxy_p2p_nccl_xpyd.py
+3
-2
@@ -46,7 +46,7 @@ def _listen_for_register(poller, router_socket):
46
global prefill_instances
47
global prefill_cv
48
with prefill_cv:
49
- node = prefill_instances.pop(data["http_address"], None)
49
+ node = prefill_instances.get(data["http_address"], None)
50
prefill_instances[data["http_address"]] = (
51
data["zmq_address"],
52
time.time() + DEFAULT_PING_SECONDS,
@@ -57,7 +57,7 @@ def _listen_for_register(poller, router_socket):
57
global decode_instances
58
global decode_cv
59
with decode_cv:
60
- node = decode_instances.pop(data["http_address"], None)
60
+ node = decode_instances.get(data["http_address"], None)
61
decode_instances[data["http_address"]] = (
62
data["zmq_address"],
63
time.time() + DEFAULT_PING_SECONDS,
@@ -69,6 +69,7 @@ def _listen_for_register(poller, router_socket):
69
remote_address,
70
data,
71
)
72
+ return
73
74
if node is None:
75
print(f"🔵Add [HTTP:{data['http_address']}, ZMQ:{data['zmq_address']}]")