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] Fix torchrun DP with LLM class (#27395)
Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
22quinn committed
Oct 24, 2025 at 01:11 UTC
e0ef8a29206d48e0af37fdaed6635b97db4c0c9f
1 file changed
+4
-3
vllm/entrypoints/llm.py
+4
-3
@@ -286,10 +286,11 @@ class LLM:
286
structured_outputs_instance = StructuredOutputsConfig()
287
288
# warn about single-process data parallel usage.
289
- _dps = int(kwargs.get("data_parallel_size", 1))
290
- if _dps > 1:
289
+ _dp_size = int(kwargs.get("data_parallel_size", 1))
290
+ _distributed_executor_backend = kwargs.get("distributed_executor_backend")
291
+ if _dp_size > 1 and not _distributed_executor_backend == "external_launcher":
292
raise ValueError(
292
- f"LLM(data_parallel_size={_dps}) is not supported for single-"
293
+ f"LLM(data_parallel_size={_dp_size}) is not supported for single-"
294
"process usage and may hang. Please use "
295
"the explicit multi-process data-parallel example at "
296
"'examples/offline_inference/data_parallel.py'."