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 15, 14:45 UTC
[bugfix] Avoid the time consumption caused by creating dummy videos. (#16371)
Jintao committed
Apr 10, 2025 at 12:30 UTC
4aed0ca6a2de8c56838fe850f8d8f8559074744a
1 file changed
+6
vllm/multimodal/profiling.py
+6
@@ -78,6 +78,8 @@ class BaseDummyInputsBuilder(ABC, Generic[_I]):
78
length: int,
79
num_audios: int,
80
) -> list[npt.NDArray]:
81
+ if num_audios == 0:
82
+ return []
83
audio = np.zeros((length, ))
84
return [audio] * num_audios
85
@@ -88,6 +90,8 @@ class BaseDummyInputsBuilder(ABC, Generic[_I]):
90
height: int,
91
num_images: int,
92
) -> list[Image.Image]:
93
+ if num_images == 0:
94
+ return []
95
image = Image.new("RGB", (width, height), color=255)
96
return [image] * num_images
97
@@ -99,6 +103,8 @@ class BaseDummyInputsBuilder(ABC, Generic[_I]):
103
num_frames: int,
104
num_videos: int,
105
) -> list[npt.NDArray]:
106
+ if num_videos == 0:
107
+ return []
108
video = np.full((num_frames, width, height, 3), 255)
109
return [video] * num_videos
110