examples/scripts/dpo_vlm.py

# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# /// script
# dependencies = [
#     "trl[peft]",
#     "Pillow>=9.4.0",
#     "torchvision",
#     "trackio",
#     "kernels",
# ]
# ///

"""
Without dataset streaming:

```
accelerate launch examples/scripts/dpo_vlm.py \
    --dataset_name HuggingFaceH4/rlaif-v_formatted \
    --model_name_or_path Qwen/Qwen2.5-VL-3B-Instruct \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 32 \
    --dataset_num_proc 32 \
    --output_dir dpo_qwen_2_5_rlaif-v \
    --dtype bfloat16 \
    --use_peft \
    --lora_target_modules all-linear
```

With dataset streaming:

```
accelerate launch examples/scripts/dpo_vlm.py \
    --dataset_name HuggingFaceH4/rlaif-v_formatted \
    --dataset_streaming \
    --model_name_or_path Qwen/Qwen2.5-VL-3B-Instruct \
    --per_device_train_batch_size 2 \
    --max_steps 100 \
    --gradient_accumulation_steps 32 \
    --dataset_num_proc 32 \
    --output_dir dpo_qwen_2_5_rlaif-v \
    --dtype bfloat16 \
    --use_peft \
    --lora_target_modules all-linear
```
"""

import torch
from datasets import load_dataset
from transformers import AutoModelForImageTextToText, AutoProcessor

from trl import (
    DPOConfig,
    DPOTrainer,
    ModelConfig,
    ScriptArguments,
    TrlParser,
    get_kbit_device_map,
    get_peft_config,
    get_quantization_config,
)


if __name__ == "__main__":
    parser = TrlParser((ScriptArguments, DPOConfig, ModelConfig))
    script_args, training_args, model_args = parser.parse_args_and_config()

    ################
    # Model & Processor
    ################
    dtype = model_args.dtype if model_args.dtype in ["auto", None] else getattr(torch, model_args.dtype)

    model_kwargs = dict(
        revision=model_args.model_revision,
        attn_implementation=model_args.attn_implementation,
        dtype=dtype,
    )
    quantization_config = get_quantization_config(model_args)
    if quantization_config is not None:
        # Passing None would not be treated the same as omitting the argument, so we include it only when valid.
        model_kwargs["device_map"] = get_kbit_device_map()
        model_kwargs["quantization_config"] = quantization_config

    model = AutoModelForImageTextToText.from_pretrained(
        model_args.model_name_or_path,
        trust_remote_code=model_args.trust_remote_code,
        **model_kwargs,
    )
    peft_config = get_peft_config(model_args)

    processor = AutoProcessor.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code, do_image_splitting=False
    )

    if script_args.ignore_bias_buffers:
        # torch distributed hack
        model._ddp_params_and_buffers_to_ignore = [
            name for name, buffer in model.named_buffers() if buffer.dtype == torch.bool
        ]

    ################
    # Dataset
    ################
    dataset = load_dataset(
        script_args.dataset_name,
        name=script_args.dataset_config,
        streaming=script_args.dataset_streaming,
    )

    ################
    # Training
    ################
    trainer = DPOTrainer(
        model,
        args=training_args,
        train_dataset=dataset[script_args.dataset_train_split],
        eval_dataset=dataset[script_args.dataset_test_split] if training_args.eval_strategy != "no" else None,
        peft_config=peft_config,
    )

    trainer.train()

    # Save and push to hub
    trainer.save_model(training_args.output_dir)
    if training_args.push_to_hub:
        trainer.push_to_hub(dataset_name=script_args.dataset_name)