diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index b91c808710..8e930261f6 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -202,6 +202,7 @@ grpo_compute_loss = RL_REPLACEMENTS["grpo_compute_loss"]
 RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(grpo_compute_loss))
 
 global INPUTS
+INPUTS = None
 
 # Edit _get_per_token_logps to handle mixed precision
 def grpo_trainer_compute_loss(function_name, function):