From 3e75f210869904878d5e1e58f4d8f473ee3804cb Mon Sep 17 00:00:00 2001
From: simpissa <91391384+simpissa@users.noreply.github.com>
Date: Sat, 21 Jun 2025 23:58:06 -0500
Subject: [PATCH] Fix for grpo_compute_loss_slow (#2702)

* slice last logit

* move slicing
---
 unsloth/models/rl_replacements.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 1fe63bb44c..96e256b653 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -337,6 +337,8 @@ def grpo_trainer_compute_loss(function_name, function):
             old_hidden_states = None
         input_ids = input_ids[:, -logits_to_keep:]
         if per_token_logps is not None:
+            ref_per_token_logps = ref_per_token_logps[:, :-1, :]
+            per_token_logps = per_token_logps[:, :-1, :]
             loss, completion_length, mean_kl = grpo_compute_loss_slow(
                 ref_per_token_logps, per_token_logps, old_hidden_states, input_ids, completion_mask, self.beta, advantages,
                 loss_type = self.args.loss_type,