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][DP] Fix creating too many DP Placement Groups (#26880)
Signed-off-by: Kebe <mail@kebe7jun.com> Signed-off-by: Rui Qiao <ruisearch42@gmail.com> Co-authored-by: Rui Qiao <ruisearch42@gmail.com>
Kebe committed
Oct 24, 2025 at 05:16 UTC
51dd14ac2b9a17cbc8eb91fe0f39311f1d594f63
1 file changed
+9
vllm/v1/engine/utils.py
+9
@@ -497,6 +497,8 @@ class CoreEngineActorManager:
497
)
498
placement_groups.append(pg)
499
local_dp_ranks.append(i)
500
+ if len(placement_groups) == dp_size:
501
+ break
502
503
if len(placement_groups) < dp_size:
504
raise ValueError(
@@ -506,6 +508,13 @@ class CoreEngineActorManager:
508
"Available resources: "
509
f"{available_resources}"
510
)
511
+ assert len(placement_groups) == dp_size, (
512
+ f"Created {len(placement_groups)} DP placement groups, expected {dp_size}"
513
+ )
514
+ assert len(local_dp_ranks) == dp_size, (
515
+ f"local_dp_ranks length {len(local_dp_ranks)} does not match "
516
+ f"expected {dp_size}"
517
+ )
518
return placement_groups, local_dp_ranks
519
520
@staticmethod