From d8a29af1cbb3db86fa5c8c5b29ef0e27d176d4ff Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 21 Jun 2025 22:14:21 -0700
Subject: [PATCH] Fix DAPO, TRL 0.19.0

---
 unsloth/models/rl.py | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index 69785be206..8ebb5c824b 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -471,6 +471,7 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
         "num_generations"             : 8,
         "top_k"                       : None,
         "vllm_mode"                   : "colocate",
+        "generation_kwargs"           : {},
     }
     for k, v in replacements.items():
         x = f"{k}( = [^,\n]{{1,}})?,\n"
@@ -510,6 +511,8 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
 
     # Check for loss_type = dr_grpo and scale_rewards for GRPO
     if "loss_type" in call_args and "scale_rewards" in call_args:
+        # See https://github.com/huggingface/trl/issues/3130#issuecomment-2746947835
+        # DAPO uses per token loss so BNPO loss used
         check_dr_grpo = \
         "if loss_type.lower() == 'dr_grpo':\n"\
         "    loss_type = 'dr_grpo'\n"\
@@ -519,13 +522,16 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
         "    if scale_rewards == None:\n"\
         "        scale_rewards = True\n"\
         "    elif scale_rewards == True:\n"\
-        "        print('The Dr GRPO paper recommends setting `scale_rewards` to False! Will override. Set it to `None` to force False.')\n"\
+        "        print('Unsloth: The Dr GRPO paper recommends setting `scale_rewards` to False! Will override. Set it to `None` to force False.')\n"\
         "        scale_rewards = False\n"\
         "elif loss_type.lower() == 'dapo':\n"\
-        "    print('The DAPO paper recommends `mask_truncated_completions = True`')\n"\
-        "    print('The DAPO paper recommends `epsilon_high = 0.28`')\n"\
+        "    print('Unsloth: The DAPO paper recommends `mask_truncated_completions = True`')\n"\
+        "    print('Unsloth: The DAPO paper recommends `epsilon_high = 0.28`')\n"\
+        "    print('Unsloth: The DAPO paper recommends setting `beta = 0.0` to remove the KL term')\n"\
         "    mask_truncated_completions = True\n"\
         "    epsilon_high = 0.28\n"\
+        "    beta = 0.0\n"\
+        "    loss_type = 'bnpo'"
         "\n"
         extra_args += check_dr_grpo
     pass