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, 16:00 UTC
[Bugfix] Fix default behavior/fallback for pp in v1 (#16057)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Apr 4, 2025 at 11:58 UTC
4708f13a9c8758d08cd45df200d63aa97762e873
1 file changed
+3
-2
vllm/engine/arg_utils.py
+3
-2
@@ -1521,8 +1521,9 @@ class EngineArgs:
1521
# PP is supported on V1 with Ray distributed executor,
1522
# but off for MP distributed executor for now.
1523
if (self.pipeline_parallel_size > 1
1524
- and self.distributed_executor_backend == "mp"
1525
- and _warn_or_fallback("PP (MP distributed executor)")):
1524
+ and self.distributed_executor_backend != "ray"):
1525
+ name = "Pipeline Parallelism without Ray distributed executor"
1526
+ _raise_or_fallback(feature_name=name, recommend_to_remove=False)
1527
return False
1528
1529
# ngram is supported on V1, but off by default for now.