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 13, 22:00 UTC
[Core][TPU] Support TPU Data Parallalism (#27365)
Signed-off-by: wenxindongwork <wenxindong@google.com>
wenxindongwork committed
Nov 1, 2025 at 11:14 UTC
af6e19f50f1d5d0c3801948c3ab17b2af231c259
1 file changed
+6
-1
vllm/entrypoints/llm.py
+6
-1
@@ -67,6 +67,7 @@ from vllm.outputs import (
67
RequestOutput,
68
ScoringRequestOutput,
69
)
70
+from vllm.platforms import current_platform
71
from vllm.pooling_params import PoolingParams
72
from vllm.sampling_params import BeamSearchParams, RequestOutputKind, SamplingParams
73
from vllm.tasks import PoolingTask
@@ -289,7 +290,11 @@ class LLM:
290
# warn about single-process data parallel usage.
291
_dp_size = int(kwargs.get("data_parallel_size", 1))
292
_distributed_executor_backend = kwargs.get("distributed_executor_backend")
292
- if _dp_size > 1 and not _distributed_executor_backend == "external_launcher":
293
+ if (
294
+ _dp_size > 1
295
+ and not _distributed_executor_backend == "external_launcher"
296
+ and not current_platform.is_tpu()
297
+ ):
298
raise ValueError(
299
f"LLM(data_parallel_size={_dp_size}) is not supported for single-"
300
"process usage and may hang. Please use "