examples/scripts/sft_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",
#     "trackio",
#     "kernels",
# ]
# ///

"""
pip install pillow

# Tested on 8x H100 GPUs
accelerate launch \
    --config_file examples/accelerate_configs/deepspeed_zero3.yaml \
    examples/scripts/sft_vlm.py \
    --dataset_name HuggingFaceH4/llava-instruct-mix-vsft \
    --model_name_or_path llava-hf/llava-1.5-7b-hf \
    --gradient_accumulation_steps 8 \
    --output_dir LLaVA-1.5-7B-SFT \
    --dtype bfloat16

For LLaVA-NeXT, use:
    --model_name_or_path llava-hf/llava-v1.6-mistral-7b-hf

For meta-llama/Llama-3.2-11B-Vision-Instruct, use:
    --model_name_or_path meta-llama/Llama-3.2-11B-Vision-Instruct

accelerate launch \
    --config_file examples/accelerate_configs/deepspeed_zero3.yaml \
    examples/scripts/sft_vlm.py \
    --dataset_name HuggingFaceH4/llava-instruct-mix-vsft \
    --model_name_or_path HuggingFaceTB/SmolVLM-Instruct \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --output_dir SmolVLM-SFT \
    --dtype bfloat16 \
    --use_peft \
    --lora_target_modules down_proj, o_proj, k_proj, q_proj, gate_proj, up_proj, v_proj
"""

import torch
from datasets import load_dataset
from transformers import AutoModelForImageTextToText

from trl import (
    ModelConfig,
    ScriptArguments,
    SFTConfig,
    SFTTrainer,
    TrlParser,
    get_kbit_device_map,
    get_peft_config,
    get_quantization_config,
)


if __name__ == "__main__":
    parser = TrlParser((ScriptArguments, SFTConfig, ModelConfig))
    script_args, training_args, model_args = parser.parse_args_and_config()
    training_args.max_length = None

    ################
    # Model
    ################
    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
    )

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

    ################
    # Training
    ################
    trainer = SFTTrainer(
        model=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=get_peft_config(model_args),
    )

    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)