Loading vagen/trainer/ppo/ray_trainer.py +2 −2 Original line number Diff line number Diff line Loading @@ -293,8 +293,8 @@ def reduce_metrics(metrics: dict): def _compute_response_info(batch): if "loss_mask" in batch.batch.keys(): # end_of_response_position_mask=batch.batch["end_of_response_position_mask"] response_length = batch.batch['loss_mask'].sum(-1) prompt_length = batch.batch['attention_mask'].sum(-1)-batch.batch['loss_mask'].sum(-1) response_length = batch.batch['loss_mask'].sum(-1).float() prompt_length = (batch.batch['attention_mask'].sum(-1)-batch.batch['loss_mask'].sum(-1)).float() response_part_length = batch.batch['responses'].shape[-1] response_mask = batch.batch['loss_mask'][:, -response_part_length:] else: Loading Loading
vagen/trainer/ppo/ray_trainer.py +2 −2 Original line number Diff line number Diff line Loading @@ -293,8 +293,8 @@ def reduce_metrics(metrics: dict): def _compute_response_info(batch): if "loss_mask" in batch.batch.keys(): # end_of_response_position_mask=batch.batch["end_of_response_position_mask"] response_length = batch.batch['loss_mask'].sum(-1) prompt_length = batch.batch['attention_mask'].sum(-1)-batch.batch['loss_mask'].sum(-1) response_length = batch.batch['loss_mask'].sum(-1).float() prompt_length = (batch.batch['attention_mask'].sum(-1)-batch.batch['loss_mask'].sum(-1)).float() response_part_length = batch.batch['responses'].shape[-1] response_mask = batch.batch['loss_mask'][:, -response_part_length:] else: Loading
