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 15, 09:45 UTC
[BugFix][DP] Send DP wave completion only from `dp_rank==0` (#18502)
Signed-off-by: Nick Hill <nhill@redhat.com> Co-authored-by: kourosh hakhamaneshi <kourosh@anyscale.com>
Nick Hill committed
May 21, 2025 at 20:25 UTC
cf5984b2fe33e0ec56eca4c83c440d03943750e6
1 file changed
+2
-2
vllm/v1/engine/core.py
+2
-2
@@ -701,7 +701,7 @@ class DPEngineCoreProc(EngineCoreProc):
701
for i in range(local_dp_rank * world_size, (local_dp_rank + 1) *
702
world_size))
703
704
- self.local_dp_rank = local_dp_rank
704
+ self.dp_rank = dp_rank
705
self.dp_group = vllm_config.parallel_config.stateless_init_dp_group()
706
self.current_wave = 0
707
@@ -774,7 +774,7 @@ class DPEngineCoreProc(EngineCoreProc):
774
local_unfinished_reqs)
775
776
if not self.engines_running:
777
- if self.local_dp_rank == 0:
777
+ if self.dp_rank == 0:
778
# Notify client that we are pausing the loop.
779
logger.debug("Wave %d finished, pausing engine loop.",
780
self.current_wave)