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 for PixtralHF models without spatial_merge_size (#16513)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Apr 11, 2025 at 17:32 UTC
87b836ba777c07acf35da802be0aacbfc51da3b5
1 file changed
+3
-2
vllm/model_executor/models/pixtral.py
+3
-2
@@ -926,8 +926,9 @@ class PixtralHFEncoderInfo(VisionEncoderInfo[PixtralVisionConfig]):
926
return self.vision_config.image_size
927
928
def get_patch_size(self) -> int:
929
- return (self.vision_config.patch_size *
930
- self.vision_config.spatial_merge_size)
929
+ spatial_merge_size = getattr(self.vision_config, "spatial_merge_size",
930
+ 1)
931
+ return (self.vision_config.patch_size * spatial_merge_size)
932
933
def get_patch_grid_length(self) -> int:
934
image_size, patch_size = self.get_image_size(), self.get_patch_size()