fix bug, tested

Author: YeAnbang

.gitignore

@@ -167,3 +167,7 @@ applications/ColossalChat/wandb
 applications/ColossalChat/model
 applications/ColossalChat/eval
 applications/ColossalChat/rollouts
+applications/ColossalChat/*.txt
+applications/ColossalChat/*.db
+applications/ColossalChat/stdin
+applications/ColossalChat/*.zip

applications/ColossalChat/coati/distributed/consumer.py

@@ -132,9 +132,7 @@ def loop(self) -> None:
                             format_acc = raw_batch["format_acc"][:, :, 0]
                             ans_acc = raw_batch["ans_acc"][:, :, 0]
                             response_len = (
-                                raw_batch["response_idx"][:, :, 1]
-                                - raw_batch["response_idx"][:, :, 0]
-                                + 1
+                                raw_batch["response_idx"][:, :, 1] - raw_batch["response_idx"][:, :, 0] + 1
                             ).type(torch.float32)
                             effective_group_mask = None
                             if self.filter_range is not None and self.grpo_config.get("dynamic_batching", True):

applications/ColossalChat/coati/distributed/producer.py

@@ -291,7 +291,7 @@ def loop(self) -> None:
                     reward_model_output = self.reward_model(
                         outputs["input_ids"].view((-1, outputs["input_ids"].size(-1))),
                         gt_answer=gt_answer,
-                        response_idx=outputs["response_idx"],
+                        response_idx=outputs["response_idx"].view((-1, 2)),
                     )
                 outputs["reward"] = (
                     torch.tensor([value[0] for value in reward_model_output])