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, 08:30 UTC
[Core][Multimodal] Convert PIL Image to array without data copy when hashing (#18682)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Lukas Geiger committed
May 25, 2025 at 18:33 UTC
6071e989df1531b59ef35568f83f7351afb0b51e
2 files changed
+3
-3
vllm/multimodal/hasher.py
+2
-2
@@ -36,8 +36,8 @@ class MultiModalHasher:
36
return np.array(obj).tobytes()
37
38
if isinstance(obj, Image.Image):
39
- return cls.item_to_bytes("image",
40
- np.array(convert_image_mode(obj, "RGBA")))
39
+ return cls.item_to_bytes(
40
+ "image", np.asarray(convert_image_mode(obj, "RGBA")))
41
if isinstance(obj, torch.Tensor):
42
return cls.item_to_bytes("tensor", obj.numpy())
43
if isinstance(obj, np.ndarray):
vllm/multimodal/video.py
+1
-1
@@ -164,7 +164,7 @@ class VideoMediaIO(MediaIO[npt.NDArray]):
164
)
165
166
return np.stack([
167
- np.array(load_frame(frame_data))
167
+ np.asarray(load_frame(frame_data))
168
for frame_data in data.split(",")
169
])
170