biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[1/N] Elastic EP Milestone 2 (#34861)

Browse Source 
Signed-off-by: Yongji Wu <wuyongji317@gmail.com>
Signed-off-by: Itay Alroy <ialroy@nvidia.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: Ron Tourgeman <rtourgeman@nvidia.com>
Co-authored-by: Yongji Wu <wuyongji317@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Ron Tourgeman <rtourgeman@nvidia.com>
This commit is contained in:
Itay Alroy
2026-02-28 06:46:42 +02:00
committed by GitHub
parent 90805ff464
commit dea268336f
53 changed files with 3613 additions and 1016 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
vllm/v1/engine/coordinator.py
View File

@@ -71,6 +71,9 @@ class DPCoordinator:
)
local_only_eng = dp_size == parallel_config.data_parallel_size_local
# NOTE(yongji): handling scaling from intra-node to inter-node
if parallel_config.enable_elastic_ep:
local_only_eng = False
back_publish_address = get_engine_client_zmq_addr(local_only_eng, host)
back_output_address = get_engine_client_zmq_addr(local_only_eng, host)
@@ -201,6 +204,7 @@ class DPCoordinatorProc:
poller = zmq.Poller()
poller.register(publish_front, zmq.POLLIN)
poller.register(publish_back, zmq.POLLIN)
poller.register(output_back, zmq.POLLIN)
last_publish_time = 0
while True:
@@ -231,6 +235,22 @@ class DPCoordinatorProc:
events = dict(events)
wave_state_changed = False
if publish_back in events:
buffer = publish_back.recv()
if buffer == b"\x01":
# NOTE(yongji): newly started engine subscribed
# We need to send READY message here instead of receiving
# SCALE_ELASTIC_EP notification from engine core client
# as SCALE_ELASTIC_EP is only sent when
# new engines finished initialization.
# Subscription message, on the other hand, is sent
# by each engine during initialization
publish_back.send(b"READY")
else:
logger.error(
"DP Coordinator receives unexpected message from engines"
)
if publish_front in events:
buffer = publish_front.recv()
if buffer in (b"\x01", b"\x00"):
@@ -259,7 +279,6 @@ class DPCoordinatorProc:
# current_wave
# we note that 0 is the wave number for the new
# engine
engines_running = False
logger.info(
"DPCoordinator scaled up from %s to %s engines",
current_count,