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
[Log] Add Warning for `LLM(data_parallel_size=k)` single-process DP Usage (#27282)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Oct 22, 2025 at 08:12 UTC
1a0f4defb76c5c6d3b6ba49ca2f560aad5afdab4
1 file changed
+10
vllm/entrypoints/llm.py
+10
@@ -285,6 +285,16 @@ class LLM:
285
else:
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:
291
+ raise ValueError(
292
+ f"LLM(data_parallel_size={_dps}) is not supported for single-"
293
+ "process usage and may hang. Please use "
294
+ "the explicit multi-process data-parallel example at "
295
+ "'examples/offline_inference/data_parallel.py'."
296
+ )
297
+
298
engine_args = EngineArgs(
299
model=model,
300
runner=runner,