cbipok
/

VideoLLaMA3-2B-fork

Video-Text-to-Text

videollama3_qwen2

text-generation

large-language-model

video-language-model

Model card Files Files and versions

lkhl commited on Mar 12, 2025

Commit

e129b11

·

verified ·

1 Parent(s): 16a7483

Update modeling_videollama3.py

Files changed (1) hide show

modeling_videollama3.py +1 -1

modeling_videollama3.py CHANGED Viewed

@@ -286,7 +286,7 @@ class Videollama3MetaForCausalLM(ABC):
         # 2. embed visual tokens
         batched_num_patches = grid_sizes.prod(dim=1).div(merge_sizes ** 2).long()
-        mm_features = self.encode_images(pixel_values, grid_sizes, merge_sizes)
         mm_features = self._get_valid_visual_tokens(mm_features, batched_num_patches, modals)
         compression_mask = self._get_compression_mask(

         # 2. embed visual tokens
         batched_num_patches = grid_sizes.prod(dim=1).div(merge_sizes ** 2).long()
+        mm_features = self.encode_images(pixel_values, grid_sizes, merge_sizes).to(input_ids.device)
         mm_features = self._get_valid_visual_tokens(mm_features, batched_num_patches, modals)
         compression_mask = self._get_compression_mask(