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
[Model][Gemma3] Simplify image input validation (#18710)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Lukas Geiger committed
May 27, 2025 at 04:13 UTC
0eebd74842d56a2c84374ee2559ef99f9c841774
1 file changed
+6
-12
vllm/model_executor/models/gemma3_mm.py
+6
-12
@@ -504,18 +504,12 @@ class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP,
504
return next(self.parameters()).dtype
505
506
def _validate_pixel_values(self, data: torch.Tensor) -> torch.Tensor:
507
- h = w = self.config.vision_config.image_size
508
- expected_dims = (3, h, w)
509
-
510
- def _validate_shape(d: torch.Tensor):
511
- if d.shape != expected_dims:
512
- raise ValueError(
513
- "The expected shape of pixel values per image per batch "
514
- f"is {expected_dims}. You supplied {tuple(d.shape)}.")
515
-
516
- for d in data:
517
- _validate_shape(d)
518
-
507
+ image_size = self.config.vision_config.image_size
508
+ expected_dims = (3, image_size, image_size)
509
+ if data.shape[1:] != expected_dims:
510
+ raise ValueError(
511
+ "The expected shape of pixel values per image per batch is "
512
+ f"{expected_dims}. You supplied {tuple(data.shape)}.")
513
return data
514
515
def _parse_and_validate_image_input(