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 12, 16:00 UTC
[NIXL] Use config to enable telemetry + NIXL version bump (#29305)
Signed-off-by: NickLucche <nlucches@redhat.com>
Nicolò Lucchesi committed
Nov 24, 2025 at 18:18 UTC
26a465584ae14a04a1f6e9b36621d70e9d907c10
2 files changed
+4
-7
requirements/kv_connectors.txt
+1
-1
@@ -1,2 +1,2 @@
1
lmcache
2
-nixl >= 0.6.0 # Required for disaggregated prefill
2
+nixl >= 0.7.1 # Required for disaggregated prefill
vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+3
-6
@@ -4,7 +4,6 @@ import contextlib
4
import copy
5
import logging
6
import math
7
-import os
7
import queue
8
import threading
9
import time
@@ -810,9 +809,6 @@ class NixlConnectorWorker:
809
self.nixl_backends = vllm_config.kv_transfer_config.get_from_extra_config(
810
"backends", ["UCX"]
811
)
813
- # TODO temporary, once nixl allows for telemetry flag in config
814
- # (next release), we can remove this env var.
815
- os.environ["NIXL_TELEMETRY_ENABLE"] = "1"
812
813
# Agent.
814
non_ucx_backends = [b for b in self.nixl_backends if b != "UCX"]
@@ -828,10 +824,11 @@ class NixlConnectorWorker:
824
if nixl_agent_config is None:
825
config = None
826
else:
827
+ # Enable telemetry by default for NIXL 0.7.1 and above.
828
config = (
832
- nixl_agent_config(backends=self.nixl_backends)
829
+ nixl_agent_config(backends=self.nixl_backends, capture_telemetry=True)
830
if len(non_ucx_backends) > 0
834
- else nixl_agent_config(num_threads=num_threads)
831
+ else nixl_agent_config(num_threads=num_threads, capture_telemetry=True)
832
)
833
834
self.nixl_wrapper = NixlWrapper(str(uuid.uuid4()), config)