File size: 5,227 Bytes
1fa3c6c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
# 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",
#     "math-verify",
#     "latex2sympy2_extended",
#     "torchvision",
#     "trackio",
#     "kernels",
# ]
# ///

"""

pip install math_verify



# For Qwen/Qwen2.5-VL-3B-Instruct

accelerate launch \

    --config_file examples/accelerate_configs/deepspeed_zero3.yaml \

    examples/scripts/rloo_vlm.py \

    --model_name_or_path Qwen/Qwen2.5-VL-3B-Instruct \

    --output_dir rloo-Qwen2.5-VL-3B-Instruct \

    --learning_rate 1e-5 \

    --dtype bfloat16 \

    --max_completion_length 1024 \

    --use_vllm \

    --vllm_mode colocate \

    --use_peft \

    --lora_target_modules "q_proj", "v_proj" \

    --log_completions



# For HuggingFaceTB/SmolVLM2-2.2B-Instruct

pip install num2words==0.5.14



accelerate launch \

    --config_file examples/accelerate_configs/deepspeed_zero3.yaml \

    examples/scripts/rloo_vlm.py \

    --model_name_or_path HuggingFaceTB/SmolVLM2-2.2B-Instruct \

    --output_dir rloo-SmolVLM2-2.2B-Instruct \

    --learning_rate 1e-5 \

    --dtype bfloat16 \

    --max_completion_length 1024 \

    --use_peft \

    --lora_target_modules "q_proj", "v_proj" \

    --log_completions \

    --per_device_train_batch_size 1 \

    --gradient_accumulation_steps 2 \

    --num_generations 2



"""

import torch
from datasets import load_dataset

from trl import (
    ModelConfig,
    RLOOConfig,
    RLOOTrainer,
    ScriptArguments,
    TrlParser,
    get_kbit_device_map,
    get_peft_config,
    get_quantization_config,
)
from trl.rewards import accuracy_reward, think_format_reward


if __name__ == "__main__":
    parser = TrlParser((ScriptArguments, RLOOConfig, ModelConfig))
    script_args, training_args, model_args = parser.parse_args_and_config()
    ################
    # Model
    ################
    dtype = model_args.dtype if model_args.dtype in ["auto", None] else getattr(torch, model_args.dtype)
    training_args.model_init_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.
        training_args.model_init_kwargs["device_map"] = get_kbit_device_map()
        training_args.model_init_kwargs["quantization_config"] = quantization_config

    ################
    # Dataset
    ################
    dataset = load_dataset("lmms-lab/multimodal-open-r1-8k-verified", split="train")
    dataset = dataset.train_test_split(test_size=100, seed=42)

    SYSTEM_PROMPT = (
        "A conversation between user and assistant. The user asks a question, and the assistant solves it. The "
        "assistant first thinks about the reasoning process in the mind and then provides the user with the answer. "
        "The reasoning process and answer are enclosed within <think></think> tags, i.e., <think>\nThis is my "
        "reasoning.\n</think>\nThis is my answer."
    )

    def make_conversation(example):
        prompt = [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": example["problem"]},
        ]
        return {"prompt": prompt}

    dataset = dataset.map(make_conversation)

    # Filter have big images
    def filter_big_images(example):
        image = example["image"]
        return image.size[0] < 512 and image.size[1] < 512

    dataset = dataset.filter(filter_big_images)

    def convert_to_rgb(example):
        image = example["image"]
        if image.mode != "RGB":
            image = image.convert("RGB")
        example["image"] = image
        return example

    dataset = dataset.map(convert_to_rgb)

    train_dataset = dataset["train"]
    eval_dataset = dataset["test"] if training_args.eval_strategy != "no" else None

    ################
    # Training
    ################
    trainer = RLOOTrainer(
        model=model_args.model_name_or_path,
        args=training_args,
        reward_funcs=[think_format_reward, accuracy_reward],
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        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)