convert : yield Gemma3N custom_map tensors directly (#19091)
This commit is contained in:
parent
bcb43163ae
commit
0bf5636938
|
|
@ -6145,7 +6145,8 @@ class Gemma3nVisionAudioModel(ConformerAudioModel):
|
|||
|
||||
if name.startswith("model.vision_tower.timm_model.blocks."):
|
||||
# Double-indexed block tensors through custom logic
|
||||
new_name = self.custom_map(name)
|
||||
yield (self.custom_map(name), data_torch)
|
||||
return
|
||||
else:
|
||||
# Route non-repeating (conv_stem, msfa, embedding, etc.) and un-catched through tensor_mapping.py
|
||||
new_name = self.map_tensor_name(name)
|
||||
|
|
|
|||
Loading…
Reference in New Issue