Bug fixes

pyproject.toml

@@ -55,11 +55,11 @@ huggingfacenotorch = [
     "hf_transfer",
     "diffusers",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,<=4.57.2",
-    "trl>=0.18.2,!=0.19.0,<=0.23.0",
+    "trl>=0.18.2,!=0.19.0,<=0.24.0",
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2025.10.13",
+    "unsloth_zoo>=2025.11.1",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -489,7 +489,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2025.10.13",
+    "unsloth_zoo>=2025.11.1",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,<=4.57.2",
@@ -507,7 +507,7 @@ colab-new = [
 ]
 colab-no-deps = [
     "accelerate>=0.34.1",
-    "trl>=0.18.2,!=0.19.0,<=0.23.0",
+    "trl>=0.18.2,!=0.19.0,<=0.24.0",
     "peft>=0.7.1",
     "xformers ; ('linux' in sys_platform or sys_platform == 'win32') and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
     "bitsandbytes>=0.45.5,!=0.46.0,!=0.48.0",

unsloth/__init__.py

@@ -29,7 +29,6 @@
 # their code at import time. If they're imported first, the original (slower,
 # more memory-intensive) implementations will be used instead of Unsloth's
 # optimized versions, potentially causing OOM errors or slower training.
-
 if already_imported:
     # stacklevel=2 makes warning point to user's import line rather than this library code,
     # showing them exactly where to fix the import order in their script
@@ -40,7 +39,7 @@
         f"Please restructure your imports with 'import unsloth' at the top of your file.",
         stacklevel = 2,
     )
-pass
+del already_imported, critical_modules
 
 # Unsloth currently does not work on multi GPU setups - sadly we are a 2 brother team so
 # enabling it will require much more work, so we have to prioritize. Please understand!
@@ -57,14 +56,12 @@
 # Log Unsloth is being used
 os.environ["UNSLOTH_IS_PRESENT"] = "1"
 
-import importlib.util
-from pathlib import Path
 from importlib.metadata import version as importlib_version
 from importlib.metadata import PackageNotFoundError
 # Check for unsloth_zoo
 try:
     unsloth_zoo_version = importlib_version("unsloth_zoo")
-    if Version(unsloth_zoo_version) < Version("2025.10.13"):
+    if Version(unsloth_zoo_version) < Version("2025.11.1"):
         print(
             "Unsloth: Please update Unsloth and Unsloth-Zoo to the latest version!\n"\
             "Do this via `pip install --upgrade --force-reinstall --no-cache-dir --no-deps unsloth unsloth_zoo`"
@@ -82,7 +79,7 @@
     raise ImportError(f"Unsloth: Please install unsloth_zoo via `pip install unsloth_zoo` then retry!")
 except:
     raise
-pass
+del PackageNotFoundError, importlib_version
 
 # Try importing PyTorch and check version
 try:
@@ -92,8 +89,8 @@
         "Unsloth: Pytorch is not installed. Go to https://pytorch.org/.\n"\
         "We have some installation instructions on our Github page."
     )
-except Exception as exception:
-    raise exception
+except:
+    raise
 pass
 
 from unsloth_zoo.device_type import (
@@ -130,7 +127,7 @@ def is_bf16_supported(including_emulation = False):
     else:
         def is_bf16_supported(): return SUPPORTS_BFLOAT16
         torch.cuda.is_bf16_supported = is_bf16_supported
-    pass
+    del major_version, minor_version
 elif DEVICE_TYPE == "hip":
     SUPPORTS_BFLOAT16 = torch.cuda.is_bf16_supported()
 elif DEVICE_TYPE == "xpu":
@@ -179,6 +176,8 @@ def is_bf16_supported(): return SUPPORTS_BFLOAT16
                 latest_cuda = np.argsort([float(find_number.search(x).group(1)) for x in possible_cudas])[::-1][0]
                 latest_cuda = possible_cudas[latest_cuda]
                 os.system(f"ldconfig /usr/local/{latest_cuda}")
+                del find_number, latest_cuda
+            del possible_cudas, find_cuda
         pass
 
         importlib.reload(bnb)
@@ -200,7 +199,7 @@ def is_bf16_supported(): return SUPPORTS_BFLOAT16
                 "Also try `sudo ldconfig /usr/local/cuda-xx.x` - find the latest cuda version.\n"\
                 "Unsloth will still run for now, but maybe it might crash - let's hope it works!"
             )
-    pass
+    del libcuda_dirs
 elif DEVICE_TYPE == "hip":
     # NO-OP for rocm device
     pass

unsloth/models/_utils.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2025.10.12"
+__version__ = "2025.11.1"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",
@@ -235,6 +235,9 @@ def filter(self, x): return not (self.text in x.getMessage())
 # No label_names provided for model class
 from transformers.trainer import logger as transformers_trainer_logger
 transformers_trainer_logger.addFilter(HideLoggingMessage("No label_names"))
+
+# The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config.
+transformers_trainer_logger.addFilter(HideLoggingMessage("The tokenizer has new"))
 del transformers_trainer_logger
 
 # Using the default loss: `ForCausalLMLoss`.

unsloth/models/rl.py

@@ -752,11 +752,11 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
         "        scale_rewards = False\n"\
         "elif loss_type.lower() == 'dapo':\n"\
         "    if mask_truncated_completions != True:\n"\
-        "        print('Unsloth: The DAPO paper recommends `mask_truncated_completions = True`')\n"\
+        "        print('Unsloth: The DAPO paper recommends `mask_truncated_completions = True` - we will set it.')\n"\
         "    if epsilon_high != 0.28:\n"\
-        "        print('Unsloth: The DAPO paper recommends `epsilon_high = 0.28`')\n"\
+        "        print('Unsloth: The DAPO paper recommends `epsilon_high = 0.28` - we will set it.')\n"\
         "    if beta != 0.0:\n"\
-        "        print('Unsloth: The DAPO paper recommends setting `beta = 0.0` to remove the KL term')\n"\
+        "        print('Unsloth: The DAPO paper recommends setting `beta = 0.0` to remove the KL term - we will set it.')\n"\
         "    mask_truncated_completions = True\n"\
         "    epsilon_high = 0.28\n"\
         "    beta = 0.0\n"\
@@ -1085,10 +1085,13 @@ def patch_functions(RLTrainer, trainer_file, RLTrainer_name, all_imports, import
         )
 
         # Replace self.llm.generate and self.llm.chat
-        lora_name = trainer_file + "_lora_model"
+        if "CUDA_VISIBLE_DEVICES" in os.environ:
+            lora_name = trainer_file + "_lora_model_' + " + "(os.environ.get('CUDA_VISIBLE_DEVICES', '0').replace(',',''))"
+        else:
+            lora_name = trainer_file + "_lora_model'"
         source = re.sub(
             r"(self\.llm\.(?:generate|chat)\([^\)]{1,})\)",
-            r"\1, lora_request = self.model.load_lora('" + lora_name + r"', load_tensors = True))",
+            r"\1, lora_request = self.model.load_lora('" + lora_name + r", load_tensors = True))",
             source
         )
         # Prefer using unsloth's sampling params and fallback to trl's if not found

unsloth/models/rl_replacements.py

@@ -216,17 +216,17 @@ def grpo_trainer__generate_and_score_completions(function_name, function):
     replacement_lines = """
         batch_size = self.args.per_device_train_batch_size if mode == "train" else self.args.per_device_eval_batch_size
         try:
-            #TRL 0.23.1 and below path
+            # TRL 0.23.1 and below path
             if not has_images:
                 # Left pad prompt before calculation old and ref hidden states
                 prompt_completion_ids = left_pack_padding(prompt_completion_ids, self.processing_class.pad_token_id)
             self.model.for_training()
-        except: 
-            #TRL 0.24.0 and below path
+        except:
+            # TRL 0.24.0 and below path
             if images is None:
                 # Left pad prompt before calculation old and ref hidden states
                 prompt_completion_ids = left_pack_padding(prompt_completion_ids, self.processing_class.pad_token_id)
-            self.model.for_training()"""
+        self.model.for_training()"""
 
     function = function.replace(line_to_replace, replacement_lines)
 

unsloth/models/vision.py

@@ -91,6 +91,7 @@
     "qwen2_5_vl",
     "gemma3",
     "mistral3",
+    "qwen3_vl",
 ]
 VLLM_NON_LORA_VLM = [
     "mllama",