From 91db85048897dffdd97bcb6a8dfec96615789ccd Mon Sep 17 00:00:00 2001
From: pluesclues <136766175+pluesclues@users.noreply.github.com>
Date: Tue, 4 Nov 2025 10:29:27 -0500
Subject: [PATCH] Detach logits before returning from function (#3554)

---
 unsloth/models/rl_replacements.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 281e3a71df..6584e2a426 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -510,7 +510,7 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
 
             os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0"
             # logits = logits[:, :-1, :]  # (B, L-1, V), exclude the last logit: it corresponds to the next token pred
-            return logits, entropies  # logps, entropies
+            return logits.detach(), entropies  # logps, entropies
             # input_ids = input_ids[:, -logits_to_keep:]
             # For transformers<=4.48, logits_to_keep argument isn't supported, so here we drop logits ourselves.
             # See https://github.com/huggingface/trl/issues/2770