app.py

import os, json, torch, gradio as gr
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import LoraConfig, PeftModel, get_peft_model
from trl import SFTTrainer, SFTConfig

# ==== 基本配置（可改小模型/步数）====
BASE_MODEL = os.getenv("BASE_MODEL", "Qwen/Qwen2.5-0.5B-Instruct")
ADAPTER_DIR = os.getenv("ADAPTER_DIR", "lora_adapter")
TRAIN_PATH = os.getenv("TRAIN_PATH", "data/sft_train.jsonl")
VAL_PATH   = os.getenv("VAL_PATH", "data/sft_val.jsonl")

# ==== 懒加载：先占位，按钮点了再真正下载 ====
_tokenizer = None
_base_model = None
_gen_model = None  # 推理用（可能带LoRA）

def load_base(load_in_4bit=None):
    global _tokenizer, _base_model
    if _tokenizer is None:
        _tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, use_fast=True)
    if _base_model is None:
        use_4bit = torch.cuda.is_available() if load_in_4bit is None else load_in_4bit
        if use_4bit:
            bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")
            _base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, quantization_config=bnb, device_map="auto")
        else:
            _base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.float32, device_map="cpu")
    return _tokenizer, _base_model


def train_qlora(max_steps=500, lora_r=16, lora_alpha=32, lora_dropout=0.05, per_device_bs=1, grad_accum=8):
    # 准备数据
    if not os.path.exists(TRAIN_PATH):
        return f"[Error] Train file not found: {TRAIN_PATH}"
    if not os.path.exists(VAL_PATH):
        return f"[Error] Val file not found: {VAL_PATH}"

    train_ds = load_dataset("json", data_files=TRAIN_PATH)["train"]
    val_ds   = load_dataset("json", data_files=VAL_PATH)["train"]

    tok, base = load_base(load_in_4bit=True)

    # LoRA 配置
    peft_cfg = LoraConfig(
        r=lora_r, lora_alpha=lora_alpha, lora_dropout=lora_dropout,
        target_modules=["q_proj","k_proj","v_proj","o_proj","up_proj","down_proj"],
        task_type="CAUSAL_LM"
    )
    model = get_peft_model(base, peft_cfg)

    # 训练配置（TRL）
    sft_cfg = SFTConfig(
        output_dir=ADAPTER_DIR,
        max_steps=int(max_steps),
        per_device_train_batch_size=per_device_bs,
        gradient_accumulation_steps=grad_accum,
        learning_rate=2e-4,
        bf16=torch.cuda.is_available(),
        logging_steps=20,
        save_steps=200,
        packing=False
    )
    trainer = SFTTrainer(
        model=model,
        tokenizer=tok,
        train_dataset=train_ds,
        eval_dataset=val_ds,
        args=sft_cfg
    )
    trainer.train()
    trainer.save_model(ADAPTER_DIR)
    return f"✅ Trained LoRA saved to: {ADAPTER_DIR}"

def load_for_infer(adapter_dir=ADAPTER_DIR):
    global _gen_model
    tok, base = load_base(load_in_4bit=True)
    if adapter_dir and os.path.isdir(adapter_dir):
        _gen_model = PeftModel.from_pretrained(base, adapter_dir)
    else:
        _gen_model = base
    return "✅ Model ready (with LoRA)" if adapter_dir and os.path.isdir(adapter_dir) else "✅ Model ready (base only)"

def generate(prompt, max_new_tokens=200, adapter_dir=ADAPTER_DIR):
    if _gen_model is None:
        load_for_infer(adapter_dir)
    tok, _ = load_base(load_in_4bit=True)
    inputs = tok(prompt, return_tensors="pt").to(_gen_model.device)
    with torch.no_grad():
        out = _gen_model.generate(**inputs, max_new_tokens=int(max_new_tokens), do_sample=True, temperature=0.8)
    return tok.decode(out[0], skip_special_tokens=True)

# ==== Gradio UI ====
with gr.Blocks(title="WeChat Style QLoRA (Minimal)") as demo:
    gr.Markdown("## WeChat Style QLoRA — Minimal Demo  \n"
                "使用 QLoRA 在私有 JSONL 上做最小监督微调（SFT）并进行推理。  \n"
                "**建议流程**：先用 CPU Basic 启动验证 → 切到 ZeroGPU/T4 训练 30–60 分钟 → 保存 LoRA → 返回 CPU 测试推理。")

    with gr.Tab("Train (QLoRA)"):
        gr.Markdown("**请先把 `data/sft_train.jsonl` 和 `data/sft_val.jsonl` 上传到本 Space 的 `data/` 目录。**")
        ms = gr.Number(value=500, label="max_steps")
        r  = gr.Number(value=16, label="lora_r")
        a  = gr.Number(value=32, label="lora_alpha")
        d  = gr.Number(value=0.05, label="lora_dropout")
        bsz= gr.Number(value=1, label="per_device_train_batch_size")
        gas= gr.Number(value=8, label="gradient_accumulation_steps")
        train_btn = gr.Button("Start Training (GPU/ZeroGPU)")
        train_log = gr.Textbox(label="Training Log", interactive=False)
        train_btn.click(fn=train_qlora, inputs=[ms,r,a,d,bsz,gas], outputs=train_log)

    with gr.Tab("Inference"):
        gr.Markdown("默认会尝试加载 `lora_adapter/`。若还没训练，可直接用基础模型。")
        adapter = gr.Textbox(value=ADAPTER_DIR, label="LoRA adapter dir")
        load_btn = gr.Button("Load (with/without LoRA)")
        load_log = gr.Textbox(label="Status", interactive=False)
        load_btn.click(fn=load_for_infer, inputs=[adapter], outputs=load_log)

        prompt = gr.Textbox(lines=6, label="Prompt")
        gen_tokens = gr.Slider(32, 512, value=200, step=8, label="max_new_tokens")
        gen_btn = gr.Button("Generate")
        output = gr.Textbox(lines=12, label="Output")
        gen_btn.click(fn=generate, inputs=[prompt, gen_tokens, adapter], outputs=output)

    gr.Markdown("> 提示：训练时请先在 **Settings → Hardware** 切到 **ZeroGPU/T4**；完成后切回 **CPU Basic** 并停止 Space 以节省费用。")

if __name__ == "__main__":
    demo.queue().launch()