423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-10-14 15:52:49 +08:00
Code Issues Packages Projects Releases Wiki Activity

add some

Browse Source 
Former-commit-id: 6233764d18f31365e9ba450408306fad55567ffc
This commit is contained in:
fzc8578 2025-01-11 01:10:24 +08:00
parent 9dc7b6c7ac
commit 12c83e00fc
1 changed files with 2 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
src/llamafactory/data/collator.py
View File

@ -152,14 +152,11 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
features.update(mm_inputs)
if isinstance(features.get("pixel_values"), list): # for pixtral inputs
features = features.data # use default_collate() instead of BatchEncoding.to()
if "image_bound" in features: # for minicpmv inputs
features["position_ids"] = [torch.arange(input_ids.size(0)).long() for input_ids in features["input_ids"]]
features["position_ids"] = pad_sequence(features["position_ids"], batch_first=True, padding_value=0)
features["labels"] = pad_sequence(features["labels"], batch_first=True, padding_value=-100)
features["attention_mask"] = pad_sequence(features["attention_mask"], batch_first=True, padding_value=0)
new_features = {}
new_features.update({"data": features})
new_features = {"data": features}
new_features.update(features)
features = new_features