From 4e2fb4911e3ee46243c72fe50665470886ad6028 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Sun, 16 Feb 2025 18:34:45 -0800 Subject: [PATCH] Update rl_replacements.py --- unsloth/models/rl_replacements.py | 7 ++----- 1 file changed, 2 insertions(+), 5 deletions(-) diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py index 8e930261f6..92b12647cf 100644 --- a/unsloth/models/rl_replacements.py +++ b/unsloth/models/rl_replacements.py @@ -201,9 +201,6 @@ RL_FUNCTIONS["grpo_trainer"].append(grpo_trainer__get_per_token_logps) 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): if function_name != "compute_loss": return function @@ -235,8 +232,8 @@ def grpo_trainer_compute_loss(function_name, function): loss, completion_length, mean_kl = grpo_compute_loss( ref_per_token_logps, per_token_logps, input_ids, completion_mask, self.beta, advantages, ) - global INPUTS - INPUTS = ( + from unsloth_zoo.rl_replacements import RL_REPLACEMENTS + RL_REPLACEMENTS["data"] = ( ref_per_token_logps, per_token_logps, input_ids, completion_mask, self.beta, advantages, loss, completion_length, mean_kl, )