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 4, 01:15 UTC
[Bugfix] Fix test_whisper distributed test process handling (#42038)
Signed-off-by: dqzhengAP <dqzheng1996@gmail.com>
David Zheng committed
May 8, 2026 at 20:37 UTC
845ca327ce140d71f2829c95aee49c39f8723269
1 file changed
+3
tests/models/multimodal/generation/test_whisper.py
+3
@@ -41,6 +41,7 @@ def run_test(
41
tensor_parallel_size: int,
42
distributed_executor_backend: str | None = None,
43
enforce_eager: bool = True,
44
+ gpu_memory_utilization: float = 0.9,
45
) -> None:
46
"""Inference result should be the same between hf and vllm.
47
@@ -57,6 +58,7 @@ def run_test(
58
distributed_executor_backend=distributed_executor_backend,
59
limit_mm_per_prompt={"audio": 2},
60
enforce_eager=enforce_eager,
61
+ gpu_memory_utilization=gpu_memory_utilization,
62
disable_custom_all_reduce=True,
63
) as vllm_model:
64
vllm_outputs_per_case = [
@@ -319,6 +321,7 @@ def test_models_distributed(
321
tensor_parallel_size=2,
322
distributed_executor_backend=distributed_executor_backend,
323
enforce_eager=False,
324
+ gpu_memory_utilization=0.65,
325
)
326
327