convert : yield Gemma3N custom_map tensors directly (#19091)
This commit is contained in:
@@ -6145,7 +6145,8 @@ class Gemma3nVisionAudioModel(ConformerAudioModel):
|
|||||||
|
|
||||||
if name.startswith("model.vision_tower.timm_model.blocks."):
|
if name.startswith("model.vision_tower.timm_model.blocks."):
|
||||||
# Double-indexed block tensors through custom logic
|
# Double-indexed block tensors through custom logic
|
||||||
new_name = self.custom_map(name)
|
yield (self.custom_map(name), data_torch)
|
||||||
|
return
|
||||||
else:
|
else:
|
||||||
# Route non-repeating (conv_stem, msfa, embedding, etc.) and un-catched through tensor_mapping.py
|
# Route non-repeating (conv_stem, msfa, embedding, etc.) and un-catched through tensor_mapping.py
|
||||||
new_name = self.map_tensor_name(name)
|
new_name = self.map_tensor_name(name)
|
||||||
|
|||||||
Reference in New Issue
Block a user