app.py

import gradio as gr
import json
import os
import torch
import pandas as pd
from transformers import (
    AutoTokenizer, 
    AutoModelForCausalLM, 
    TrainingArguments,
    BitsAndBytesConfig
)
from peft import LoraConfig, get_peft_model, TaskType, PeftModel
from trl import SFTTrainer
from datasets import Dataset

# --- КОНФИГУРАЦИЯ ---
MODEL_ID = "Maincode/Maincoder-1B"
OUTPUT_DIR = "mandre_qlora_adapter"
JSON_FILE_NAME = "train_data.json"

# Глобальные переменные для чата
chat_model = None
chat_tokenizer = None

# ==========================================
# ЧАСТЬ 1: ГЕНЕРАТОР ДАТАСЕТА
# ==========================================

def generate_json_dataset(files):
    if not files:
        return None, "❌ Ошибка: Вы не загрузили файлы."

    data_entries = []
    
    for file_item in files:
        if isinstance(file_item, str):
            file_path = file_item
        elif hasattr(file_item, 'name'):
            file_path = file_item.name
        else:
            continue

        try:
            with open(file_path, 'r', encoding='utf-8') as f:
                content = f.read()
        except Exception as e:
            print(f"Skipping file {file_path}: {e}")
            continue 
            
        filename = os.path.basename(file_path)
        
        instruction = f"Analyze the code/text in file '{filename}' from the MandreLib project."
        text = f"### Instruction:\n{instruction}\n\n### Response:\n{content}<|endoftext|>"
        
        data_entries.append({"text": text})

    if not data_entries:
        return None, "❌ Не удалось прочитать ни один текстовый файл."

    try:
        with open(JSON_FILE_NAME, 'w', encoding='utf-8') as f:
            json.dump(data_entries, f, indent=4, ensure_ascii=False)
        abs_path = os.path.abspath(JSON_FILE_NAME)
        return abs_path, f"✅ Готово! Обработано файлов: {len(data_entries)}. Файл {JSON_FILE_NAME} создан."
    except Exception as e:
        return None, f"❌ Ошибка записи JSON: {e}"

# ==========================================
# ЧАСТЬ 2: ОБУЧЕНИЕ (ИСПРАВЛЕНО)
# ==========================================

def train_mandre_ai(file_obj, epochs, lr):
    if file_obj is None:
        if os.path.exists(JSON_FILE_NAME):
            json_path = JSON_FILE_NAME
            yield f"⚠️ Файл не передан, используем {JSON_FILE_NAME} из прошлой генерации."
        else:
            yield "❌ Ошибка: Нет файла с данными!"
            return
    else:
        json_path = file_obj.name if hasattr(file_obj, 'name') else file_obj

    yield f"🚀 Старт обучения {MODEL_ID}..."
    
    try:
        # 1. Загрузка данных
        with open(json_path, 'r', encoding='utf-8') as f:
            data = json.load(f)
        dataset = Dataset.from_pandas(pd.DataFrame(data))
        
        yield f"📊 Данные: {len(dataset)} строк. Загрузка токенизатора..."

        # 2. Токенизатор (FIX: use_fast=False чтобы избежать ошибки Rust)
        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=False, trust_remote_code=True)
        tokenizer.pad_token = tokenizer.eos_token

        # 3. LoRA Config
        peft_config = LoraConfig(
            task_type=TaskType.CAUSAL_LM,
            inference_mode=False,
            r=16,
            lora_alpha=32,
            lora_dropout=0.05,
            target_modules=['q_proj', 'v_proj', 'k_proj', 'o_proj']
        )

        # 4. Аргументы
        training_args = TrainingArguments(
            output_dir=OUTPUT_DIR,
            num_train_epochs=float(epochs),
            per_device_train_batch_size=2,
            gradient_accumulation_steps=4,
            learning_rate=float(lr),
            weight_decay=0.01,
            use_cpu=True,           
            no_cuda=True,           
            fp16=False,             
            logging_steps=1,
            save_total_limit=1,
            push_to_hub=False,
            report_to="none"
        )

        yield "📥 Загрузка модели (Maincoder-1B)..."
        
        # 5. Загрузка модели
        model = AutoModelForCausalLM.from_pretrained(MODEL_ID, trust_remote_code=True)
        model.gradient_checkpointing_enable() 
        model.enable_input_require_grads()

        yield "🧠 Инициализация тренера..."

        # 6. Trainer
        trainer = SFTTrainer(
            model=model,
            args=training_args,
            train_dataset=dataset,
            dataset_text_field="text",
            peft_config=peft_config,
            tokenizer=tokenizer,
            max_seq_length=1024
        )

        yield "🔥 ОБУЧЕНИЕ ЗАПУЩЕНО! Ждите завершения..."
        
        trainer.train()
        
        yield "💾 Сохранение..."
        trainer.model.save_pretrained(OUTPUT_DIR)
        tokenizer.save_pretrained(OUTPUT_DIR)
        
        yield f"✅ УСПЕХ! Адаптер в папке '{OUTPUT_DIR}'. Можно чатиться."

    except Exception as e:
        import traceback
        yield f"❌ ОШИБКА:\n{traceback.format_exc()}"

# ==========================================
# ЧАСТЬ 3: ЧАТ
# ==========================================

def load_chat_model():
    global chat_model, chat_tokenizer
    if chat_model is not None: return "Уже загружено"

    try:
        # FIX: use_fast=False и здесь
        chat_tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=False, trust_remote_code=True)
        
        if os.path.exists(os.path.join(OUTPUT_DIR, "adapter_config.json")):
            base = AutoModelForCausalLM.from_pretrained(MODEL_ID, trust_remote_code=True)
            chat_model = PeftModel.from_pretrained(base, OUTPUT_DIR)
            return f"✅ Адаптер QLoRA загружен!"
        else:
            chat_model = AutoModelForCausalLM.from_pretrained(MODEL_ID, trust_remote_code=True)
            return "⚠️ Адаптер не найден. Работает чистая модель."
            
    except Exception as e:
        return f"Ошибка: {e}"

def generate_answer(prompt, history):
    if not chat_model: 
        status = load_chat_model()
        if "Ошибка" in status: return status
    
    formatted_prompt = f"### Instruction:\n{prompt}\n\n### Response:\n"
    inputs = chat_tokenizer(formatted_prompt, return_tensors="pt")
    
    outputs = chat_model.generate(
        **inputs, 
        max_new_tokens=300,
        do_sample=True,
        temperature=0.6,
        top_p=0.95
    )
    
    response = chat_tokenizer.decode(outputs[0], skip_special_tokens=True)
    if "### Response:" in response:
        return response.split("### Response:")[-1].strip()
    return response

# ==========================================
# ИНТЕРФЕЙС
# ==========================================

with gr.Blocks(title="MandreAI Fix") as demo:
    gr.Markdown("# 🦎 MandreAI 1B (CPU Fix)")
    
    with gr.Tabs():
        with gr.Tab("1. Датасет"):
            files_input = gr.File(file_count="multiple", label="Исходные файлы")
            btn_gen = gr.Button("Создать JSON", variant="primary")
            json_output = gr.File(label="Готовый датасет")
            status_output = gr.Textbox(label="Статус")
            btn_gen.click(generate_json_dataset, inputs=[files_input], outputs=[json_output, status_output])

        with gr.Tab("2. Обучение"):
            with gr.Row():
                train_file_input = gr.File(label="train_data.json")
                epochs = gr.Number(value=3, label="Эпохи", precision=0)
                lr = gr.Number(value=2e-4, label="LR")
            
            btn_train = gr.Button("ЗАПУСТИТЬ ОБУЧЕНИЕ", variant="stop")
            log_output = gr.Textbox(label="Лог", lines=10)
            
            btn_train.click(train_mandre_ai, inputs=[train_file_input, epochs, lr], outputs=[log_output])

        with gr.Tab("3. Чат"):
            chatbot = gr.Chatbot(label="MandreAI")
            msg_input = gr.Textbox(label="Вопрос")
            btn_send = gr.Button("Отправить")
            btn_send.click(generate_answer, [msg_input, chatbot], chatbot)

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