from datasets import load_dataset
from unsloth import FastLanguageModel
import torch
# import os
# os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'

# 加载 jsonl 文件
dataset = load_dataset("json", data_files="dataset/test_dataset.jsonl", split="train")

# 转换成 ChatML 格式的字符串字段
# example 相当于jsonl中的每一行
def to_chatml(example):
    messages = example["messages"]
    chat = ""
    for m in messages:
        # 将原始内容封装为一句话
        chat += f"<|im_start|>{m['role']}\n{m['content']}<|im_end|>\n"
    return {"text": chat.strip()}

# 添加 `text` 字段
dataset = dataset.map(to_chatml)

# print("\n", dataset[0])


# 加载预训练模型
model, tokenizer = FastLanguageModel.from_pretrained(
    # model_name = "unsloth/Qwen3-1.7B-unsloth-bnb-4bit",
    # model_name = "unsloth/Qwen3-8B-unsloth-bnb-4bit",
    model_name = "deepseek-ai/DeepSeek-V2-Lite",
    # model_name = "unsloth/Qwen3-4B-unsloth-bnb-4bit",
    trust_remote_code=True,        # ✅ 允許遠端自訂程式碼, deepseek 系列模型使用
    max_seq_length = 2048,   # Context length - can be longer, but uses more memory
    load_in_4bit = True,     # 4bit uses much less memory , 启用QLoRA
    load_in_8bit = False,    # A bit more accurate, uses 2x memory
    full_finetuning = False, # We have full finetuning now!
    # token = "hf_...",      # use one if using gated models
)

model = FastLanguageModel.get_peft_model(
    model,
    r = 16,           # Choose any number > 0! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 16,  # Best to choose alpha = rank or rank*2
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
    random_state = 3407,
    use_rslora = False,   # We support rank stabilized LoRA
    loftq_config = None,  # And LoftQ
)


from trl import SFTTrainer, SFTConfig
trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    eval_dataset = None, # Can set up evaluation!
    args = SFTConfig(
        dataset_text_field = "text", # 要和 dataset中定义的字段统一
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 4, # Use GA to mimic batch size!
        warmup_steps = 5,
        # num_train_epochs = 1, # Set this for 1 full training run.
        max_steps = 30,
        learning_rate = 2e-4, # Reduce to 2e-5 for long training runs
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "linear",
        seed = 3407,
        report_to = "none", # Use this for WandB etc
    ),
)

trainer.train()


messages = [
    {"role" : "user", "content" : "请介绍一下昊天"}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize = False,
    add_generation_prompt = True, # Must add for generation
    enable_thinking = False, # Disable thinking
)

from transformers import TextStreamer
_ = model.generate(
    **tokenizer(text, return_tensors = "pt").to("cuda"),
    max_new_tokens = 256, # Increase for longer outputs!
    temperature = 0.7, top_p = 0.8, top_k = 20, # For non thinking
    streamer = TextStreamer(tokenizer, skip_prompt = True),
)

# model.cpu()

model.save_pretrained_gguf(
    "DeepSeek-V2-Lite", 
    tokenizer, 
    # quantization_method="q4_k_m",  # 或 "q8_0"  # 量化模式--默认 q8_0, 可选f16, "q4_k_m", "q8_0", "q5_k_m",
    maximum_memory_usage=0.7  # 限制使用 GPU 显存为总容量的 50%
)