dpo_vlm.py

[liuchaohu]

System Info

trl env

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder
• My own task or dataset (give details below)

Reproduction

# Copyright 2025 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.

"""
accelerate launch examples/scripts/dpo_vlm.py \
    --dataset_name HuggingFaceH4/rlaif-v_formatted \
    --model_name_or_path HuggingFaceM4/idefics2-8b \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 32 \
    --dataset_num_proc 32 \
    --output_dir dpo_idefics_rlaif-v \
    --bf16 \
    --torch_dtype bfloat16 \
    --gradient_checkpointing \
    --use_peft \
    --lora_target_modules=all-linear
"""
import os
import torch
from datasets import load_dataset, features
from transformers import AutoModelForVision2Seq, AutoProcessor

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

from dpo_trainer import DPOTrainer


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

    ################
    # Model & Tokenizer
    ################
    torch_dtype = (
        model_args.torch_dtype if model_args.torch_dtype in ["auto", None] else getattr(torch, model_args.torch_dtype)
    )
    quantization_config = get_quantization_config(model_args)

    model_kwargs = dict(
        revision=model_args.model_revision,
        attn_implementation=model_args.attn_implementation,
        torch_dtype=torch_dtype,
        device_map=get_kbit_device_map() if quantization_config is not None else None,
        quantization_config=quantization_config,
        low_cpu_mem_usage=True,
    )
    model = AutoModelForVision2Seq.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)
    if peft_config is None:
        ref_model = AutoModelForVision2Seq.from_pretrained(
            model_args.model_name_or_path,
            trust_remote_code=model_args.trust_remote_code,
            **model_kwargs,
        )
    else:
        ref_model = None
    processor = AutoProcessor.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code, do_image_splitting=False
    )
    tokenizer = processor.tokenizer

    # Set up the chat template
    if model.config.model_type == "idefics2":
        pass  # the processor already has a valid chat template
    elif model.config.model_type == "paligemma":
        processor.chat_template = """{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}<|im_start|>{% if message['role'] == 'user' %}USER: {% else %}ASSISTANT: {% endif %}{% for item in message['content'] if item['type'] == 'text' %}{{ item['text'] }}<|im_end|>{% endfor %}{% if message['role'] == 'user' %} {% else %}{{eos_token}}{% endif %}{% endfor %}{% if add_generation_prompt %}ASSISTANT: {% endif %}"""
    elif model.config.model_type == "llava":
        processor.chat_template = """{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{% if message['role'] == 'user' %}USER: {% else %}ASSISTANT: {% endif %}{% for item in message['content'] %}{% if item['type'] == 'text' %}{{ item['text'] }}{% elif item['type'] == 'image' %}<image>{% endif %}{% endfor %}{% if message['role'] == 'user' %} {% else %}{{eos_token}}{% endif %}{% endfor %}{% if add_generation_prompt %}ASSISTANT: {% endif %}"""

    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    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)

    def format(examples):
        """
        Convert prompt from "xxx" to [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "xxx"}]}]
        and chosen and rejected from "xxx" to [{"role": "assistant", "content": [{"type": "text", "text": "xxx"}]}].
        Images are wrapped in a list.
        """
        output = {"images": [], "prompt": [], "chosen": [], "rejected": []}
        for image, question, chosen, rejected in zip(examples["image"], examples["question"], examples["chosen"], examples["rejected"]):
            prompt = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": question}]}]
            chosen = [{"role": "assistant", "content": [{"type": "text", "text": chosen}]}]
            rejected = [{"role": "assistant", "content": [{"type": "text", "text": rejected}]}]
            output["images"].append([image])
            output["prompt"].append(prompt)
            output["chosen"].append(chosen)
            output["rejected"].append(rejected)
        return output


    dataset = load_dataset("data/openbmb/RLAIF-V-Dataset", split="train", num_proc=os.cpu_count())
    dataset = dataset.select(range(100))
    cols = dataset.column_names
    print(os.cpu_count())
    dataset = dataset.map(format, batched=True, writer_batch_size=4, batch_size=4, remove_columns=cols, num_proc=24)
    f = dataset.features
    f["images"] = features.Sequence(features.Image(decode=True))  # to avoid bytes
    dataset = dataset.cast(f)
    dataset = dataset.train_test_split(test_size=0.05)



    ################
    # Training
    ################
    trainer = DPOTrainer(
        model,
        ref_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,
        processing_class=processor,
        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)

{
  "version": "0.2.0",
  "configurations": [
    {
      "name": "Python: Debug with Arguments",
      "type": "python",
      "request": "launch",
      "program": "${file}", 
      "console": "integratedTerminal",
      "args": [
        "--dataset_name", "../data/openbmb/RLAIF-V-Dataset",
        "--model_name_or_path", "../data/llava-hf/llava-v1.6-vicuna-7b-hf",
        "--per_device_train_batch_size", "1",
        "--gradient_accumulation_steps", "1",
        "--output_dir", "dpo_idefics_rlaif-v",
        "--bf16",
        "--torch_dtype", "bfloat16",
        "--learning_rate", "1e-5",
        "--rpo_alpha", "0.1",
        "--gradient_checkpointing",
        "--use_peft",
        "--lora_target_modules=all-linear",
        "--dataset_num_proc", "1",
        "--attn_implementation", "flash_attention_2",
        "--logging_steps", "1",
        "--output_dir", "results/debug",
      ]
    }
  ]
}

The model is llava-v1.6-vicuna-7b-hf, and the dataset is RLAIF-V-Dataset.

In the dpo_trainer.py, self.train_dataset includes

Dataset({
    features: ['images', 'prompt_input_ids', 'pixel_values', 'chosen_input_ids', 'rejected_input_ids', 'image_sizes'],
    num_rows: 100
})

However, when the program runs to DataCollatorForPreference, pixel_values disappears.
This leads to the fact that in subsequent training, the model cannot receive pixel_values, but it can receive image_sizes, which is very strange.

Expected behavior

I hope to know why the dataset itself contains pixel_values ​​but why it disappears during data_collect

Checklist

• I have checked that my issue isn't already filed (see open issues)
• I have included my system information
• Any code provided is minimal, complete, and reproducible (more on MREs)
• Any code provided is properly formatted in code blocks, (no screenshot, more on code blocks)
• Any traceback provided is complete

[liuchaohu]

I know the reason why pixel_values disappears.
We should run the code the param "--remove_unused_columns false", otherwise pixel_values will be eliminated.