unused

Author: jacklanchantin

src/fairseq2/recipes/lm/_online_finetune/_grpo.py

@@ -73,76 +73,6 @@ class GRPOBatch:
     prompt_lengths: list[int]
     rewards: torch.Tensor
 
-def clip_outputs_at_think_token(rollouts, tokenizer, think_tokens, answer_len=64):
-    """
-    Clip token_ids and logprobs at the </think> token sequence start,
-    and recompute the text from clipped tokens.
-
-    Args:
-        rollouts: List of rollout objects
-        tokenizer: Tokenizer instance
-        think_tokens: List of token IDs for </think>
-        answer_len: Number of tokens to keep after </think>
-
-    Returns:
-        List of modified rollout objects
-    """
-    ret = []
-    for rollout in rollouts:
-        clipped_outputs = []
-
-        for output in rollout.outputs:
-            # Find the position where </think> tokens start
-            think_token_len = len(think_tokens)
-            clip_index = None
-
-            # Search for the think tokens sequence in token_ids
-            for i in range(len(output.token_ids) - think_token_len + 1):
-                if output.token_ids[i:i + think_token_len] == think_tokens:
-                    clip_index = i + answer_len
-                    break
-
-            if clip_index is not None:
-                # Clip token_ids and logprobs
-                clipped_token_ids = output.token_ids[:clip_index]
-                clipped_logprobs = output.logprobs[:clip_index]
-
-                # Recompute text from clipped tokens
-                clipped_text = tokenizer.decode(clipped_token_ids)
-
-                # Recalculate cumulative_logprob from clipped logprobs
-                cumulative_logprob = 0.0
-                for logprob_dict in clipped_logprobs:
-                    # Get the first token's logprob (the selected token)
-                    first_token_id = list(logprob_dict.keys())[0]
-                    cumulative_logprob += logprob_dict[first_token_id].logprob
-
-                # Create new CompletionOutput with clipped data
-                clipped_output = type(output)(
-                    index=output.index,
-                    text=clipped_text,
-                    token_ids=clipped_token_ids,
-                    cumulative_logprob=cumulative_logprob,
-                    logprobs=clipped_logprobs,
-                    finish_reason=output.finish_reason,
-                    stop_reason=output.stop_reason
-                )
-                clipped_outputs.append(clipped_output)
-            else:
-                # If </think> not found, keep original output
-                clipped_outputs.append(output)
-
-        # *** FIX: Create new rollout object with clipped outputs ***
-        clipped_rollout = type(rollout)(
-            outputs=clipped_outputs,
-            # Copy other attributes from original rollout
-            **{k: v for k, v in vars(rollout).items() if k != 'outputs'}
-        )
-        ret.append(clipped_rollout)
-
-    return ret
-
-
 def clip_outputs_after_think_token(rollouts, tokenizer, think_tokens, num_tokens):
     """
     Clip token_ids and logprobs to keep only num_tokens after the </think> token sequence ends,