xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-06 04:22:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

updated

Browse Source 
Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>
This commit is contained in:
rshaw@neuralmagic.com 2025-07-01 01:34:46 +00:00
parent 1172b70b79
commit 56939c835d
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
View File

@ -335,7 +335,7 @@ class NixlConnectorWorker:
# Agent. # Agent.
import os import os
num_workers = 16 num_workers = 32
# setting num_workers on the prefiller causes no notifs to be recved??? # setting num_workers on the prefiller causes no notifs to be recved???
# this is a hack to make sure we set num workers on the prefiller to 1. # this is a hack to make sure we set num workers on the prefiller to 1.
if os.getenv("VLLM_IS_PREFILL", "0") == "1": if os.getenv("VLLM_IS_PREFILL", "0") == "1":
@ -985,7 +985,7 @@ class NixlConnectorWorker:
assert len(local_block_descs_ids) == len(remote_block_descs_ids) assert len(local_block_descs_ids) == len(remote_block_descs_ids)
# Prepare transfer with Nixl. # Prepare transfer with Nixl.
CHUNK_SIZE = 1000 CHUNK_SIZE = 500
handles = [] handles = []
# NOTE: this is a hack to make make_prepped_xfer into threads so that # NOTE: this is a hack to make make_prepped_xfer into threads so that
# different workers are allocated for each chuck. Without this change, # different workers are allocated for each chuck. Without this change,