ai-agents-trainer-compatibilit-training-fails-with-typeerror-customtrainer-comput-64248860

THEN Update your custom compute_loss method signature to accept the new required argument num_items_in_batch with a default value of None. This is necessary because the Trainer in transformers >=4.49.0 now passes num_items_in_batch to compute_loss.

def compute_loss(self, model, inputs, num_items_in_batch=None, return_outputs=False):
    # your custom loss logic here
    ...