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
[Bugfix] Allow dynamic number of patches for llava_onevision (#23525)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Aug 25, 2025 at 16:34 UTC
ebafb0936d95c3926979f925ac65525a7c8778f7
1 file changed
+1
-1
vllm/model_executor/models/llava_onevision.py
+1
-1
@@ -75,7 +75,7 @@ class LlavaOnevisionImagePixelInputs(TensorSchema):
75
76
pixel_values: Annotated[
77
Union[torch.Tensor, list[torch.Tensor]],
78
- TensorShape("bn", "np", 3, "h", "w"),
78
+ TensorShape("bn", "np", 3, "h", "w", dynamic_dims={"np"}),
79
]
80
81
image_sizes: Annotated[Optional[torch.Tensor], TensorShape("bn", 2)]