[llama4] Pop aspect_ratios from image processor output in Llama4Processor

jmswen · jmswen · commit ca64ae5095c9 · 2025-03-31T19:27:35.000-07:00
Signed-off-by: Jon Swenson &lt;jmswen@gmail.com&gt;
diff --git a/src/transformers/models/llama4/processing_llama4.py b/src/transformers/models/llama4/processing_llama4.py
@@ -344,7 +344,7 @@ def __call__(
             num_patches_per_chunk = int(
                 (image_height // self.patch_size) * (image_width // self.patch_size) // self.downsample_ratio
             )
-            aspect_ratios = image_inputs["aspect_ratios"]
+            aspect_ratios = image_inputs.pop("aspect_ratios")
 
             total_placeholders = sum(prompt.count(self.fake_image_token) for prompt in text)
             if total_placeholders != len(images):

Original file line number	Diff line number	Diff line change
`@@ -344,7 +344,7 @@ def __call__(`
`344`	`344`	`num_patches_per_chunk = int(`
`345`	`345`	`(image_height // self.patch_size) * (image_width // self.patch_size) // self.downsample_ratio`
`346`	`346`	`)`
`347`		`- aspect_ratios = image_inputs["aspect_ratios"]`
	`347`	`+ aspect_ratios = image_inputs.pop("aspect_ratios")`
`348`	`348`
`349`	`349`	`total_placeholders = sum(prompt.count(self.fake_image_token) for prompt in text)`
`350`	`350`	`if total_placeholders != len(images):`