2.sft.py

from unsloth import FastVisionModel # FastLanguageModel for LLMs
import torch
from datasets import load_dataset
from unsloth import is_bf16_supported
from unsloth.trainer import UnslothVisionDataCollator
from trl import SFTTrainer, SFTConfig
from transformers import TextStreamer
import datetime
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")

model, tokenizer = FastVisionModel.from_pretrained(
    model_name = "/home/rzhong/project/unsloth/model_pretrain_20250301_113944",
    load_in_4bit = False, # Use 4bit to reduce memory use. False for 16bit LoRA.
                          # 模型已经选了4bit量化后的，这里还需不需要再以4bit加载？建议实验一下 | 应该是一个意思
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for long context
    max_seq_length = 2048, # unsloth支持4x的上下文微调，如果原模型支持8192的上下文，这里只需要设置为2048
    dtype = torch.bfloat16, # A100支持bfloat16，可以减少显存占用。默认为None，也可以选择torch.float16
)

model = FastVisionModel.get_peft_model(
    model,
    finetune_vision_layers     = True, # False if not finetuning vision layers
    finetune_language_layers   = True, # False if not finetuning language layers
    finetune_attention_modules = True, # False if not finetuning attention layers
    finetune_mlp_modules       = True, # False if not finetuning MLP layers
    
    r = 16,           # The larger, the higher the accuracy, but might overfit
    lora_alpha = 16,  # Recommended alpha == r at least
    lora_dropout = 0,
    bias = "none",
    random_state = 3407,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
    # target_modules = "all-linear", # Optional now! Can specify a list if needed
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
)

dataset = load_dataset("/home/share/rzhong/dataset/google-landmark/dataset_4/dataset_file", split = "train")
print(dataset)

instruction = "描述这张图片。"
# instruction = "Write the LaTeX representation for this image."

def convert_to_conversation(sample):
    conversation = [
        { "role": "user",
          "content" : [
            {"type" : "text",  "text"  : instruction},
            {"type" : "image", "image" : sample["image"]} ]
        },
        { "role" : "assistant",
          "content" : [
            {"type" : "text",  "text"  : sample["text"]} ]
        },
    ]
    return { "messages" : conversation }
pass

converted_dataset = [convert_to_conversation(sample) for sample in dataset]

print(converted_dataset[0])

FastVisionModel.for_training(model) # Enable for training!

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    data_collator = UnslothVisionDataCollator(model, tokenizer), # Must use!
    train_dataset = converted_dataset,
    args = SFTConfig(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 4, # 原来是4。可以增加，相当于提高batch size，但不会影响内存消耗。增加会使loss曲线更平滑
        warmup_steps = 5,
        # max_steps = None,
        num_train_epochs = 10, # Set this instead of max_steps for full training runs
        learning_rate = 5e-5, # 2e-4 1e-4 5e-5 2e-5
        fp16 = not is_bf16_supported(),
        bf16 = is_bf16_supported(),
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = f"outputs_pretrain_sft_{timestamp}",
        report_to = "none",     # For Weights and Biases

        # You MUST put the below items for vision finetuning:
        remove_unused_columns = False,
        dataset_text_field = "",
        dataset_kwargs = {"skip_prepare_dataset": True},
        dataset_num_proc = 4,
        max_seq_length = 2048,
    ),
)

trainer_stats = trainer.train()

model.save_pretrained(f"lora_model_pretrain_sft_{timestamp}")  # Local saving
tokenizer.save_pretrained(f"lora_model_pretrain_sft_{timestamp}")