moved some DPOTrainer args to DPOConfig for future trl release

README.md

@@ -121,7 +121,7 @@ Features:
 
 Get started with Axolotl in just a few steps! This quickstart guide will walk you through setting up and running a basic fine-tuning task.
 
-**Requirements**: Python >=3.10 and Pytorch >=2.1.1.
+**Requirements**: Nvidia GPU (Ampere architecture or newer for `bf16` and Flash Attention), Python >=3.10 and PyTorch >=2.3.1.
 
 ```bash
 git clone https://github.com/axolotl-ai-cloud/axolotl
@@ -383,7 +383,7 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
         - typescript
       type: ... # unimplemented custom format
 
-      # fastchat conversation (deprecation soon, use chat_template)
+      # fastchat conversation (deprecation soon, use chat_template https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html#chat_template)
       # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
     - path: ...
       type: sharegpt
@@ -562,8 +562,7 @@ plugins:
   - axolotl.integrations.liger.LigerPlugin
 liger_rope: true
 liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
+liger_swiglu: true
 liger_fused_linear_cross_entropy: true
 ```
 

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.

examples/llama-3/fft-8b-liger-fsdp.yaml

@@ -4,26 +4,32 @@ plugins:
   - axolotl.integrations.liger.LigerPlugin
 liger_rope: true
 liger_rms_norm: true
-liger_glu_activation: true
+liger_swiglu: true
 liger_fused_linear_cross_entropy: true
 
 strict: false
 
+chat_template: llama3
 datasets:
-    - path: tatsu-lab/alpaca
-      type: alpaca
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+
 dataset_prepared_path: last_run_prepared
-val_set_size: 0
+val_set_size: 0.02
 output_dir: ./outputs/out
 
 sequence_len: 4096
 sample_packing: true
 pad_to_sequence_len: true
 
-wandb_project: check_liger_hf_GA_llama_fix-3
-wandb_entity: axolotl-ai
+wandb_project:
+wandb_entity:
 wandb_watch:
-wandb_name: pr/fix333-tr4.46.1
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4

requirements.txt

@@ -1,7 +1,7 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
 peft==0.13.2
-transformers==4.46.1
+transformers==4.46.0
 tokenizers>=0.20.1
 bitsandbytes==0.44.1
 accelerate==1.0.1
@@ -33,8 +33,8 @@ gradio==3.50.2
 tensorboard
 python-dotenv==1.0.1
 autoawq>=0.2.5
-triton>=3.1.0
-liger-kernel==0.3.1
+triton>=2.3.0
+liger-kernel==0.3.0
 
 mamba-ssm==1.2.0.post1
 
@@ -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

@@ -1890,17 +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"
@@ -1909,13 +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"] = (
@@ -1925,11 +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,
@@ -1949,27 +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]

src/axolotl/integrations/liger/__init__.py

@@ -18,23 +18,20 @@ Module for the Plugin for LIGER integraton with Axolotl.
 Liger Kernel is the collection of Triton-native kernels for LLM Training.
 It is designed to be performant, correct, and light-weight.
 """
-import inspect
 import logging
 import sys
+from functools import partial
 
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
-from liger_kernel.transformers.monkey_patch import MODEL_TYPE_TO_APPLY_LIGER_FN
+from liger_kernel.transformers.geglu import LigerGEGLUMLP
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
 
 from axolotl.integrations.base import BasePlugin
 
-from ...utils.distributed import zero_only
 from .args import LigerArgs  # pylint: disable=unused-import. # noqa: F401
 
-LOG = logging.getLogger("axolotl.integrations.liger")
-
 
 class LigerPlugin(BasePlugin):
     """
@@ -45,31 +42,59 @@ class LigerPlugin(BasePlugin):
         return "axolotl.integrations.liger.LigerArgs"
 
     def pre_model_load(self, cfg):
-        if cfg.model_config_type in MODEL_TYPE_TO_APPLY_LIGER_FN:
-            apply_liger_fn = MODEL_TYPE_TO_APPLY_LIGER_FN[cfg.model_config_type]
-            liger_fn_sig = inspect.signature(apply_liger_fn)
-            kwargs = {}
-            if "rope" in liger_fn_sig.parameters:
-                kwargs["rope"] = cfg.liger_rope
-            if "cross_entropy" in liger_fn_sig.parameters:
-                kwargs["cross_entropy"] = cfg.liger_cross_entropy
-            if "fused_linear_cross_entropy" in liger_fn_sig.parameters:
-                kwargs[
-                    "fused_linear_cross_entropy"
-                ] = cfg.liger_fused_linear_cross_entropy
-            if "rms_norm" in liger_fn_sig.parameters:
-                kwargs["rms_norm"] = cfg.liger_rms_norm
-            if "layer_norm" in liger_fn_sig.parameters:
-                kwargs["layer_norm"] = cfg.liger_layer_norm
-            if "geglu" in liger_fn_sig.parameters:
-                kwargs["geglu"] = cfg.liger_glu_activation
-            elif "swiglu" in liger_fn_sig.parameters:
-                kwargs["swiglu"] = cfg.liger_glu_activation
-            with zero_only():
-                LOG.info(
-                    f"Applying LIGER to {cfg.model_config_type} with kwargs: {kwargs}"
+        if cfg.model_config_type == "llama":
+            from liger_kernel.transformers.model.llama import (
+                lce_forward as llama_lce_forward,
+            )
+            from transformers.models.llama import modeling_llama
+
+            if cfg.liger_rope:
+                modeling_llama.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_llama.LlamaRMSNorm = LigerRMSNorm
+            if cfg.liger_swiglu:
+                modeling_llama.LlamaMLP = LigerSwiGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_llama.CrossEntropyLoss = LigerCrossEntropyLoss
+            elif cfg.liger_fused_linear_cross_entropy:
+                modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
+
+        elif cfg.model_config_type == "mistral":
+            from liger_kernel.transformers.model.mistral import (
+                lce_forward as mistral_lce_forward,
+            )
+            from transformers.models.mistral import modeling_mistral
+
+            if cfg.liger_rope:
+                modeling_mistral.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_mistral.MistralRMSNorm = LigerRMSNorm
+            if cfg.liger_swiglu:
+                modeling_mistral.MistralMLP = LigerSwiGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_mistral.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                modeling_mistral.MistralForCausalLM.forward = mistral_lce_forward
+
+        elif cfg.model_config_type == "gemma":
+            from liger_kernel.transformers.model.gemma import (
+                lce_forward as gemma_lce_forward,
+            )
+            from transformers.models.gemma import modeling_gemma
+
+            if cfg.liger_rope:
+                modeling_gemma.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_gemma.GemmaRMSNorm = partial(
+                    LigerRMSNorm, offset=1.0, init_fn="zeros", casting_mode="gemma"
                 )
-            apply_liger_fn(**kwargs)
+            if cfg.liger_swiglu:
+                modeling_gemma.GemmaMLP = LigerGEGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_gemma.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                modeling_gemma.GemmaForCausalLM.forward = gemma_lce_forward
+
         elif cfg.model_config_type == "jamba":
             from transformers.models.jamba import modeling_jamba
 
@@ -79,12 +104,30 @@ class LigerPlugin(BasePlugin):
                 modeling_jamba.apply_rotary_pos_emb = liger_rotary_pos_emb
             if cfg.liger_rms_norm:
                 modeling_jamba.JambaRMSNorm = LigerRMSNorm
-            if cfg.liger_glu_activation:
+            if cfg.liger_swiglu:
                 modeling_jamba.JambaMLP = LigerSwiGLUMLP
             if cfg.liger_cross_entropy:
                 modeling_jamba.CrossEntropyLoss = LigerCrossEntropyLoss
             if cfg.liger_fused_linear_cross_entropy:
                 modeling_jamba.JambaForCausalLM.forward = jamba_lce_forward
+
+        elif cfg.model_config_type == "qwen2":
+            from liger_kernel.transformers.model.qwen2 import (
+                lce_forward as qwen2_lce_forward,
+            )
+            from transformers.models.qwen2 import modeling_qwen2
+
+            if cfg.liger_rope:
+                modeling_qwen2.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_qwen2.Qwen2RMSNorm = LigerRMSNorm
+            if cfg.liger_swiglu:
+                modeling_qwen2.Qwen2MLP = LigerSwiGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_qwen2.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                modeling_qwen2.Qwen2ForCausalLM.forward = qwen2_lce_forward
+
         elif cfg.model_config_type == "deepseek_v2":
             from accelerate import init_empty_weights
             from transformers import AutoModelForCausalLM
@@ -103,9 +146,44 @@ class LigerPlugin(BasePlugin):
                 logging.warning("Fused liger_rope is not supported for DeepseekV2.")
             if cfg.liger_rms_norm:
                 modeling_mod.DeepseekV2RMSNorm = LigerRMSNorm
-            if cfg.liger_glu_activation:
+            if cfg.liger_swiglu:
                 modeling_mod.DeepseekV2MLP.forward = LigerSwiGLUMLP.forward
             if cfg.liger_cross_entropy:
                 modeling_mod.CrossEntropyLoss = LigerCrossEntropyLoss
             if cfg.liger_fused_linear_cross_entropy:
                 modeling_mod.DeepseekV2ForCausalLM.forward = deepseekv2_lce_forward
+
+        elif cfg.model_config_type == "gemma2":
+            from transformers.models.gemma2 import modeling_gemma2
+
+            if cfg.liger_rope:
+                modeling_gemma2.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_gemma2.Gemma2RMSNorm = partial(
+                    LigerRMSNorm, offset=1.0, init_fn="zeros", casting_mode="gemma"
+                )
+            if cfg.liger_swiglu:
+                modeling_gemma2.Gemma2MLP = LigerGEGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_gemma2.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                logging.warning(
+                    "Fused linear cross entropy is not supported for Gemma 2."
+                )
+
+        elif cfg.model_config_type == "phi3":
+            from liger_kernel.transformers.model.phi3 import (
+                lce_forward as phi3_lce_forward,
+            )
+            from transformers.models.phi3 import modeling_phi3
+
+            if cfg.liger_rope:
+                modeling_phi3.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_phi3.Phi3RMSNorm = LigerRMSNorm
+            if cfg.liger_swiglu:
+                modeling_phi3.Phi3MLP = LigerSwiGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_phi3.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward

src/axolotl/integrations/liger/args.py

@@ -15,12 +15,9 @@
 """
 Module for handling LIGER input arguments.
 """
-import logging
 from typing import Optional
 
-from pydantic import BaseModel, model_validator
-
-LOG = logging.getLogger("axolotl.integrations.liger.args")
+from pydantic import BaseModel
 
 
 class LigerArgs(BaseModel):
@@ -30,24 +27,6 @@ class LigerArgs(BaseModel):
 
     liger_rope: Optional[bool] = None
     liger_rms_norm: Optional[bool] = None
-    liger_layer_norm: Optional[bool] = None
     liger_swiglu: Optional[bool] = None
-    liger_glu_activation: Optional[bool] = None
     liger_cross_entropy: Optional[bool] = None
     liger_fused_linear_cross_entropy: Optional[bool] = None
-
-    @model_validator(mode="before")
-    @classmethod
-    def check_deprecated_swiglu(cls, data):
-        if data.get("liger_swiglu") is not None:
-            if data.get("liger_glu_activation") is not None:
-                raise ValueError(
-                    "You cannot have both `liger_swiglu` and `liger_glu_activation` set."
-                )
-
-            LOG.warning(
-                "The 'liger_swiglu' argument is deprecated and will be removed in a future release. "
-                "Please use 'liger_glu_activation' instead."
-            )
-            data["liger_glu_activation"] = data.pop("liger_swiglu")
-        return data

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

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):