fix token_level_kl

huggingface · Jan 16, 2025 · 475a157 · 475a157
1 parent 57d9a97
commit 475a157
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 5 deletions.
diff --git a/trl/trainer/rloo_config.py b/trl/trainer/rloo_config.py
@@ -93,6 +93,6 @@ class RLOOConfig(OnPolicyConfig):
         metadata={"help": "Whether to normalize advantages"},
     )
     token_level_kl: bool = field(
-        default=True,
+        default=False,
         metadata={"help": "Whether to use token-level KL penalty or sequence-level KL penalty"},
     )
diff --git a/trl/trainer/rloo_trainer.py b/trl/trainer/rloo_trainer.py
@@ -397,13 +397,16 @@ def repeat_generator():
                 # Compute total reward with KL penalty
                 if args.token_level_kl:
                     # Token-level KL penalty: apply KL penalty per token
-                    token_kl_penalty = -args.kl_coef * kl
-                    non_score_reward = token_kl_penalty.sum(1)
+                    kl_reward = -args.kl_coef * kl
+                    # Apply reward at the last non-padded token position for each sequence
+                    eos_indices = sequence_lengths.unsqueeze(1) - 1
+                    last_reward = torch.zeros_like(kl).scatter_(dim=1, index=eos_indices, src=scores.unsqueeze(1))
+                    non_score_reward = kl_reward.sum(1) + last_reward.sum(1)
                 else:
                     # Sequence-level KL penalty: sum KL across tokens first
                     sequence_kl = kl.sum(1)
-                    non_score_reward = -args.kl_coef * sequence_kl
-                rlhf_reward = scores + non_score_reward
+                    non_score_reward = -args.kl_coef * sequence_kl + scores
+                rlhf_reward = non_score_reward
 
                 # vectorized RLOO advantages implementation
                 rlhf_reward = rlhf_reward.reshape(args.rloo_k, -1)