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
[CI/Build] Also check DP in benchmarks throughput script (#23038)
Co-authored-by: Simon Mo <simon.mo@hey.com>
Zhewen Li committed
Aug 19, 2025 at 21:09 UTC
f72902327246bc68ff0d196a89cc81262f46de1b
2 files changed
+10
-2
benchmarks/benchmark_throughput.py
+2
-2
@@ -597,8 +597,8 @@ def validate_args(args):
597
# https://github.com/vllm-project/vllm/issues/16222
598
if args.data_parallel_size > 1:
599
raise ValueError(
600
- "Data parallel is not supported in offline benchmark, \
601
- please use benchmark serving instead"
600
+ "Data parallel is not supported in offline benchmark, "
601
+ "please use benchmark serving instead"
602
)
603
604
vllm/benchmarks/throughput.py
+8
@@ -434,6 +434,14 @@ def validate_args(args):
434
if args.backend == "mii" and args.tokenizer != args.model:
435
raise ValueError(
436
"Tokenizer must be the same as the model for MII backend.")
437
+
438
+ # --data-parallel is not supported currently.
439
+ # https://github.com/vllm-project/vllm/issues/16222
440
+ if args.data_parallel_size > 1:
441
+ raise ValueError(
442
+ "Data parallel is not supported in offline benchmark, "
443
+ "please use benchmark serving instead"
444
+ )
445
446
447
def add_cli_args(parser: argparse.ArgumentParser):