moved some DPOTrainer args to DPOConfig for future trl release

* only run the remainder of the gpu test suite if one case passes first

* also reduce the test matrix

.github/workflows/base.yml

@@ -40,7 +40,7 @@ jobs:
             cuda_version: 12.4.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
     steps:
       - name: Checkout

.github/workflows/tests-nightly.yml

@@ -82,13 +82,6 @@ jobs:
             num_gpus: 1
             axolotl_extras: mamba-ssm
             nightly_build: "true"
-          - cuda: 121
-            cuda_version: 12.1.1
-            python_version: "3.11"
-            pytorch: 2.3.1
-            num_gpus: 1
-            axolotl_extras: mamba-ssm
-            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"

.github/workflows/tests.yml

@@ -72,13 +72,53 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-  docker-e2e-tests:
+  docker-e2e-tests-1st:
     if: github.repository_owner == 'axolotl-ai-cloud'
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 90
     needs: [pre-commit, pytest]
 
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.4.1
+            num_gpus: 1
+            axolotl_extras:
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+      - name: Install Modal
+        run: |
+          python -m pip install --upgrade pip
+          pip install modal==0.63.64 jinja2
+      - name: Update env vars
+        run: |
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
+          echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
+          echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
+          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
+          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+      - name: Run tests job on Modal
+        run: |
+          modal run cicd.tests
+
+  docker-e2e-tests:
+    if: github.repository_owner == 'axolotl-ai-cloud'
+    # this job needs to be run on self-hosted GPU runners...
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 90
+    needs: [pre-commit, pytest, docker-e2e-tests-1st]
+
     strategy:
       fail-fast: false
       matrix:
@@ -89,18 +129,6 @@ jobs:
             pytorch: 2.3.1
             num_gpus: 1
             axolotl_extras: mamba-ssm
-          - cuda: 121
-            cuda_version: 12.1.1
-            python_version: "3.11"
-            pytorch: 2.3.1
-            num_gpus: 1
-            axolotl_extras: mamba-ssm
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            num_gpus: 1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"

docs/config.qmd

@@ -183,6 +183,8 @@ test_datasets:
 
 # use RL training: 'dpo', 'ipo', 'kto'
 rl:
+# whether to perform weighting if doing DPO training. Boolean.
+dpo_use_weighting:
 
 # The name of the chat template to use for training, following values are supported:
 # - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.

requirements.txt

@@ -43,7 +43,7 @@ s3fs>=2024.5.0
 gcsfs>=2024.5.0
 # adlfs
 
-trl @ git+https://github.com/huggingface/trl.git@31d02cfb795284591a084416b9dcb7bef5d08924
+trl @ git++https://github.com/huggingface/trl.git@5e90682836969310e16ed8aa711dd429f85863b7
 zstandard==0.22.0
 fastcore
 

src/axolotl/core/trainer_builder.py

@@ -48,6 +48,7 @@ from trl import (
 )
 from trl.trainer.utils import RewardDataCollatorWithPadding, pad_to_length
 
+from axolotl.integrations.base import PluginManager
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils import is_comet_available, is_mlflow_available
@@ -1147,6 +1148,12 @@ class TrainerBuilderBase(abc.ABC):
 
     def get_callbacks(self) -> List[TrainerCallback]:
         callbacks = []
+
+        plugin_manager = PluginManager.get_instance()
+        callbacks.extend(
+            plugin_manager.add_callbacks_pre_trainer(cfg=self.cfg, model=self.model)
+        )
+
         if self.cfg.use_wandb:
             callbacks.append(
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
@@ -1173,11 +1180,17 @@ class TrainerBuilderBase(abc.ABC):
 
         return callbacks
 
-    @abstractmethod
     def get_post_trainer_create_callbacks(self, trainer):
         """
         Callbacks added after the trainer is created, usually b/c these need access to the trainer
         """
+        callbacks = []
+
+        plugin_manager = PluginManager.get_instance()
+        callbacks.extend(
+            plugin_manager.add_callbacks_post_trainer(cfg=self.cfg, trainer=trainer)
+        )
+        return callbacks
 
     def hook_pre_create_training_args(self, training_arguments_kwargs):
         # TODO
@@ -1223,7 +1236,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
         if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
             LogPredictionCallback = log_prediction_callback_factory(
                 trainer, self.tokenizer, "wandb"
@@ -1791,7 +1804,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
         return callbacks
 
     def build_training_arguments(self, total_num_steps):
@@ -1877,17 +1890,18 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             # default to saving each epoch if not defined
             training_args_kwargs["save_strategy"] = "epoch"
 
+        training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
+
         if self.cfg.rl_beta:
             training_args_kwargs["beta"] = self.cfg.rl_beta
         if self.cfg.orpo_alpha:
             # trl does some odd mapping of alpha to beta to reuse the beta parameter ???
             training_args_kwargs["beta"] = self.cfg.orpo_alpha
 
-        training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
-        training_args_cls = AxolotlDPOConfig
         if self.cfg.rpo_alpha is not None:
             training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
 
+        training_args_cls = None
         if self.cfg.rl == "simpo":
             training_args_cls = AxolotlCPOConfig
             training_args_kwargs["loss_type"] = "simpo"
@@ -1896,13 +1910,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.cpo_alpha is not None:
                 training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
 
-        if self.cfg.rl == "orpo":
+        elif self.cfg.rl == "orpo":
             training_args_cls = AxolotlORPOConfig
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
-        if self.cfg.rl == "kto":
+        elif self.cfg.rl == "kto":
             training_args_cls = AxolotlKTOConfig
 
             training_args_kwargs["desirable_weight"] = (
@@ -1912,11 +1926,32 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
                 self.cfg.kto_undesirable_weight or 1.0
             )
 
-            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
+        else:
+            training_args_cls = AxolotlDPOConfig
+
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+
+            training_args_kwargs["max_target_length"] = None
+            if self.cfg.max_prompt_len is not None:
+                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
+
+            if self.cfg.dpo_use_weighting is not None:
+                training_args_kwargs["use_weighting"] = self.cfg.dpo_use_weighting
+
+            if self.cfg.rl == "ipo":
+                training_args_kwargs["loss_type"] = "ipo"
+                if self.cfg.dpo_label_smoothing:
+                    training_args_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
+            
+            if self.cfg.precompute_ref_log_probs is not None:
+                training_args_kwargs["precompute_ref_log_probs"] = self.cfg.precompute_ref_log_probs
+
+            training_args_kwargs["generate_during_eval"] = self.cfg.use_wandb
+
         training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
             output_dir=self.cfg.output_dir,
             per_device_train_batch_size=self.cfg.micro_batch_size,
@@ -1936,27 +1971,16 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
     def build(self, total_num_steps):
         training_args = self.build_training_arguments(total_num_steps)
         dpo_trainer_kwargs = {}
-        if self.cfg.rl == "ipo":
-            dpo_trainer_kwargs["loss_type"] = "ipo"
-            if self.cfg.dpo_label_smoothing:
-                dpo_trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
+        
         if self.eval_dataset:
             dpo_trainer_kwargs["eval_dataset"] = self.eval_dataset
         if self.cfg.adapter and self.peft_config:
             dpo_trainer_kwargs["peft_config"] = self.peft_config
-        if self.cfg.precompute_ref_log_probs is not None:
-            dpo_trainer_kwargs[
-                "precompute_ref_log_probs"
-            ] = self.cfg.precompute_ref_log_probs
+  
         if self.cfg.rl in ["dpo", "ipo"]:
             trainer_cls = AxolotlDPOTrainer
             trainer_cls_args = [self.model, self.model_ref]
 
-            # these aren't used for the ORPO trainer
-            dpo_trainer_kwargs["max_length"] = self.cfg.sequence_len
-            dpo_trainer_kwargs["max_target_length"] = None
-            dpo_trainer_kwargs["max_prompt_length"] = self.cfg.sequence_len
-            dpo_trainer_kwargs["generate_during_eval"] = self.cfg.use_wandb
         elif self.cfg.rl == "orpo":
             trainer_cls = AxolotlORPOTrainer
             trainer_cls_args = [self.model]
@@ -2000,11 +2024,11 @@ class HFPPOTrainerBuilder(TrainerBuilderBase):
     """
 
     def get_callbacks(self):
-        callbacks = []
+        callbacks = super().get_callbacks()
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
         return callbacks
 
     def build(self, total_num_steps):

src/axolotl/integrations/base.py

@@ -18,9 +18,10 @@ Plugins can be used to integrate third-party models, modify the training process
 
 To create a new plugin, you need to inherit from the BasePlugin class and implement the required methods.
 """
+import collections
 import importlib
 import logging
-from typing import List
+from typing import OrderedDict
 
 
 class BasePlugin:
@@ -47,7 +48,7 @@ class BasePlugin:
         Initializes the BasePlugin.
         """
 
-    def register(self, cfg):
+    def register(self, cfg):  # pylint: disable=unused-argument
         """
         Registers the plugin with the given configuration.
 
@@ -63,7 +64,7 @@ class BasePlugin:
         Returns a pydantic model for the plugin's input arguments.
         """
 
-    def pre_model_load(self, cfg):
+    def pre_model_load(self, cfg):  # pylint: disable=unused-argument
         """
         Performs actions before the model is loaded.
 
@@ -74,7 +75,7 @@ class BasePlugin:
         None
         """
 
-    def post_model_load(self, cfg, model):
+    def post_model_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions after the model is loaded.
 
@@ -86,7 +87,7 @@ class BasePlugin:
         None
         """
 
-    def pre_lora_load(self, cfg, model):
+    def pre_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions before LoRA weights are loaded.
 
@@ -98,7 +99,7 @@ class BasePlugin:
         None
         """
 
-    def post_lora_load(self, cfg, model):
+    def post_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions after LoRA weights are loaded.
 
@@ -110,7 +111,7 @@ class BasePlugin:
         None
         """
 
-    def create_optimizer(self, cfg, trainer):
+    def create_optimizer(self, cfg, trainer):  # pylint: disable=unused-argument
         """
         Creates and returns an optimizer for training.
 
@@ -122,7 +123,9 @@ class BasePlugin:
         object: The created optimizer.
         """
 
-    def create_lr_scheduler(self, cfg, trainer, optimizer):
+    def create_lr_scheduler(
+        self, cfg, trainer, optimizer
+    ):  # pylint: disable=unused-argument
         """
         Creates and returns a learning rate scheduler.
 
@@ -135,7 +138,7 @@ class BasePlugin:
         object: The created learning rate scheduler.
         """
 
-    def add_callbacks_pre_trainer(self, cfg, model):
+    def add_callbacks_pre_trainer(self, cfg, model):  # pylint: disable=unused-argument
         """
         Adds callbacks to the trainer before training.
 
@@ -146,8 +149,11 @@ class BasePlugin:
         Returns:
         List[callable]: A list of callback functions to be added to the TrainingArgs
         """
+        return []
 
-    def add_callbacks_post_trainer(self, cfg, trainer):
+    def add_callbacks_post_trainer(
+        self, cfg, trainer
+    ):  # pylint: disable=unused-argument
         """
         Adds callbacks to the trainer after training.
 
@@ -158,8 +164,9 @@ class BasePlugin:
         Returns:
         List[callable]: A list of callback functions to be added to the TrainingArgs
         """
+        return []
 
-    def post_train(self, cfg, model):
+    def post_train(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions after training is complete.
 
@@ -171,7 +178,7 @@ class BasePlugin:
         None
         """
 
-    def post_train_unload(self, cfg):
+    def post_train_unload(self, cfg):  # pylint: disable=unused-argument
         """
         Performs actions after training is complete and the model is unloaded.
 
@@ -227,7 +234,7 @@ class PluginManager:
     pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.
     """
 
-    plugins: List[BasePlugin] = []
+    plugins: OrderedDict[str, BasePlugin] = collections.OrderedDict()
 
     _instance = None
 
@@ -237,7 +244,7 @@ class PluginManager:
         """
         if cls._instance is None:
             cls._instance = super(PluginManager, cls).__new__(cls)
-            cls._instance.plugins: List[BasePlugin] = []
+            cls._instance.plugins = collections.OrderedDict()
         return cls._instance
 
     @staticmethod
@@ -265,7 +272,7 @@ class PluginManager:
         """
         try:
             plugin = load_plugin(plugin_name)
-            self.plugins.append(plugin)
+            self.plugins[plugin_name] = plugin
         except ImportError:
             logging.error(f"Failed to load plugin: {plugin_name}")
 
@@ -277,7 +284,7 @@ class PluginManager:
         list[str]: A list of Pydantic classes for all registered plugins' input arguments.'
         """
         input_args = []
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             input_args_from_plugin = plugin.get_input_args()
             if input_args_from_plugin is not None:
                 input_args.append(input_args_from_plugin)
@@ -293,7 +300,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.pre_model_load(cfg)
 
     def post_model_load(self, cfg, model):
@@ -307,7 +314,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.post_model_load(cfg, model)
 
     def pre_lora_load(self, cfg, model):
@@ -321,7 +328,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.pre_lora_load(cfg, model)
 
     def post_lora_load(self, cfg, model):
@@ -335,7 +342,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.post_lora_load(cfg, model)
 
     def create_optimizer(self, cfg, trainer):
@@ -349,7 +356,7 @@ class PluginManager:
         Returns:
         object: The created optimizer, or None if none was found.
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             optimizer = plugin.create_optimizer(cfg, trainer)
             if optimizer is not None:
                 return optimizer
@@ -367,7 +374,7 @@ class PluginManager:
         Returns:
         object: The created learning rate scheduler, or None if none was found.
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             scheduler = plugin.create_lr_scheduler(cfg, trainer, optimizer)
             if scheduler is not None:
                 return scheduler
@@ -385,7 +392,7 @@ class PluginManager:
         List[callable]: A list of callback functions to be added to the TrainingArgs.
         """
         callbacks = []
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             callbacks.extend(plugin.add_callbacks_pre_trainer(cfg, model))
         return callbacks
 
@@ -401,7 +408,7 @@ class PluginManager:
         List[callable]: A list of callback functions to be added to the TrainingArgs.
         """
         callbacks = []
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             callbacks.extend(plugin.add_callbacks_post_trainer(cfg, trainer))
         return callbacks
 
@@ -416,5 +423,5 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.post_train_unload(cfg)

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -588,6 +588,9 @@ class AxolotlInputConfig(
 
     rl: Optional[RLType] = None
     reward_model: Optional[bool] = None
+    dpo_use_weighting: Optional[
+        bool
+    ] = None  # whether to use weighting in DPO trainer. If none, default is false in the trainer.
 
     datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore
     test_datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore

src/axolotl/utils/data/sft.py

@@ -2,9 +2,11 @@
 
 import functools
 import logging
+import time
 from pathlib import Path
 from typing import List, Optional, Tuple, Union
 
+import requests
 from datasets import (
     Dataset,
     DatasetDict,
@@ -53,6 +55,28 @@ from axolotl.utils.trainer import (
 LOG = logging.getLogger("axolotl")
 
 
+def retry_on_request_exceptions(max_retries=3, delay=1):
+    def decorator(func):
+        @functools.wraps(func)
+        def wrapper(*args, **kwargs):  # pylint: disable=inconsistent-return-statements
+            for attempt in range(max_retries):
+                try:
+                    return func(*args, **kwargs)
+                except (
+                    requests.exceptions.ReadTimeout,
+                    requests.exceptions.ConnectionError,
+                ) as exc:
+                    if attempt < max_retries - 1:
+                        time.sleep(delay)
+                    else:
+                        raise exc
+
+        return wrapper
+
+    return decorator
+
+
+@retry_on_request_exceptions(max_retries=3, delay=5)
 def prepare_dataset(cfg, tokenizer, processor=None):
     prompters = []
     if not cfg.pretraining_dataset:

test.yml

@@ -0,0 +1,59 @@
+base_model: JackFram/llama-68m
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: arcee-ai/distilabel-intel-orca-dpo-pairs-binarized
+    type: chatml.ultra
+    split: train
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/out
+
+sequence_len: 2048
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+early_stopping_patience:
+resume_from_checkpoint:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+rl: dpo
+dpo_use_weighting: true
+
+warmup_steps: 10
+evals_per_epoch: 2
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: <|end_of_text|>

test2.yml

@@ -0,0 +1,43 @@
+base_model: JackFram/llama-68m
+
+load_in_8bit: true
+
+datasets:
+  - path: arcee-ai/distilabel-intel-orca-dpo-pairs-binarized
+    type: chatml.ultra
+    split: train
+output_dir: ./outputs/lora-out
+
+sequence_len: 1024
+
+adapter: lora
+lora_r: 64
+lora_alpha: 32
+lora_dropout: 0.1
+lora_target_linear: true
+
+rl: dpo
+dpo_use_weighting: true
+
+wandb_project: check_dpotrainer
+wandb_entity: axolotl-ai
+wandb_watch:
+wandb_name: baseline/dpo_base/dpo_use_weighting
+wandb_log_model:
+
+
+num_epochs: 1
+micro_batch_size: 4
+gradient_accumulation_steps: 1
+learning_rate: 0.00001
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+max_steps": 20
+save_steps: 10
+warmup_steps: 5
+gradient_checkpointing: True
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+#special_tokens:
+#  pad_token: <|end_of_text|>
+

tests/e2e/test_dpo.py

@@ -115,6 +115,51 @@ class TestDPOLlamaLora(unittest.TestCase):
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "checkpoint-20/adapter_model.safetensors").exists()
 
+    @with_temp_dir
+    def test_dpo_use_weighting(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 64,
+                "lora_alpha": 32,
+                "lora_dropout": 0.1,
+                "lora_target_linear": True,
+                "special_tokens": {},
+                "rl": "dpo",
+                "dpo_use_weighting": True,
+                "datasets": [
+                    {
+                        "path": "arcee-ai/distilabel-intel-orca-dpo-pairs-binarized",
+                        "type": "chatml.ultra",
+                        "split": "train",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "paged_adamw_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "warmup_steps": 5,
+                "gradient_checkpointing": True,
+                "gradient_checkpointing_kwargs": {"use_reentrant": True},
+            }
+        )
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "checkpoint-20/adapter_model.safetensors").exists()
+
     @pytest.mark.skip("kto_pair no longer supported in trl")
     @with_temp_dir
     def test_kto_pair_lora(self, temp_dir):