Tool Outputs in Attention Mask for Log Probabilities #7
Description
From my understanding, the attention mask that is passed into https://github.com/Time-Search/TimeSearch-R/blob/main/time_r1/trainer/grpo_trainer_env.py#L874 (def _get_per_token_logps(self, model, input_ids, attention_mask, multimodal_inputs, logits_to_keep, batch_size=None) -> torch.Tensor:
is defined on line 1163
completion_mask = compute_tool_response_mask(completion_ids) * completion_mask which masks out tool outputs.
Why are the tool outputs not included in the attention mask for computing the log probabilities? Shouldn't they just be masked out for the loss?