fix _get_per_token_logps_and_entropies return type

trl/trainer/grpo_trainer.py

@@ -1031,7 +1031,7 @@ def _get_per_token_logps_and_entropies(
         mm_token_type_ids=None,
         pixel_position_ids=None,
         image_position_ids=None,
-    ) -> dict[str, torch.Tensor | None]:
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
         """Compute log-probs and (optionally) entropies for each token."""
         batch_size = batch_size or input_ids.size(0)  # Chunk inputs into smaller batches to reduce memory peak
         all_logps = []

trl/trainer/rloo_trainer.py

@@ -681,7 +681,7 @@ def _get_per_token_logps_and_entropies(
         token_type_ids=None,
         mm_token_type_ids=None,
         pixel_position_ids=None,
-    ) -> dict[str, torch.Tensor | None]:
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
         """Compute log-probs and (optionally) entropies for each token."""
         batch_size = batch_size or input_ids.size(0)  # Chunk inputs into smaller batches to reduce memory peak
         all_logps = []