diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py index 709cd1cb5c..2dc4b88dfa 100644 --- a/unsloth/models/vision.py +++ b/unsloth/models/vision.py @@ -186,6 +186,10 @@ class FastBaseVisionModel: patch_saving_functions(model, vision = True) patch_saving_functions(tokenizer, vision = True) + # Fix gradient accumulation + from transformers.trainer import Trainer + patch_gradient_accumulation_fix(Trainer) + # Save tokenizer for inference purposes tokenizer.padding_side = "left" # Force inference tokenizer.tokenizer.padding_side = "left" # Force inference