SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[modular] Follow global indexing and attribute setting, and their dependencies (#39180)

Browse Source 
* export global indexing statements

* add example

* style

* examples
This commit is contained in:
Cyril Vallez
2025-07-07 14:36:43 +02:00
committed by GitHub
parent 8570bc29f3
commit 5348fbc005
8 changed files with 254 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
examples/modular-transformers/modeling_new_task_model.py
View File

@@ -118,6 +118,8 @@ class NewTaskModelPreTrainedModel(PreTrainedModel):
)
class NewTaskModelModel(NewTaskModelPreTrainedModel):
_checkpoint_conversion_mapping = {"language_model.model": "language_model"}
# we are filtering the logits/labels so we shouldn't divide the loss based on num_items_in_batch
accepts_loss_kwargs = False
def __init__(self, config: NewTaskModelConfig):
super().__init__(config)
@@ -313,9 +315,11 @@ class NewTaskModelModel(NewTaskModelPreTrainedModel):
special_image_mask = inputs_embeds == self.get_input_embeddings()(
torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
)
special_image_mask = special_image_mask.all(-1)
else:
special_image_mask = (input_ids == self.config.image_token_id).unsqueeze(-1)
special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
special_image_mask = input_ids == self.config.image_token_id
special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
image_tokens_in_text = (special_image_mask).sum(dim=1).sum(dim=0)[0]