model.py

import spaces
import torch
from transformers import AutoTokenizer, BitsAndBytesConfig
from peft import AutoPeftModelForCausalLM 

MODEL_ID = "Playingyoyo/aLLoyM"
model = None
tokenizer = None

def load_model():
    global model, tokenizer
    print(f"Loading Model: {MODEL_ID}...")
    
    quantization_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_use_double_quant=True,
        bnb_4bit_quant_type="nf4"
    )
    
    try:
        model = AutoPeftModelForCausalLM.from_pretrained(
            MODEL_ID,
            quantization_config=quantization_config,
            device_map="auto"
        )
        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
        
        if tokenizer.pad_token_id is None:
            tokenizer.pad_token_id = tokenizer.eos_token_id
            
        model.eval()
        print("Model loaded successfully!")
    except Exception as e:
        print(f"Error loading model: {e}")

@spaces.GPU(duration=120)
def run_inference(user_input, system_instruction, history):
    global model, tokenizer
    
    # historyがNoneなら空リストにする
    if history is None:
        history = []

    if model is None:
        load_model()
        if model is None:
            # エラー時も辞書形式で追加
            history.append({"role": "user", "content": user_input})
            history.append({"role": "assistant", "content": "Error: Model failed to load."})
            return history

    # Alpaca形式プロンプト
    prompt = f"""### Instruction:
{system_instruction}

### Input:
{user_input}

### Output:
"""

    inputs = tokenizer(
        [prompt],
        return_tensors='pt',
        truncation=True,
        max_length=2048
    ).to("cuda")

    with torch.no_grad():
        outputs = model.generate(
            **inputs, 
            max_new_tokens=512,
            use_cache=True,
            do_sample=False,
            pad_token_id=tokenizer.eos_token_id
        )
    
    full_output = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
    
    if "### Output:" in full_output:
        generated_response = full_output.split("### Output:")[1].strip()
    else:
        generated_response = full_output.replace(prompt, "").strip()

    # 【重要】エラーメッセージに従い、辞書形式 (Messages format) で返す
    history.append({"role": "user", "content": user_input})
    history.append({"role": "assistant", "content": generated_response})
    
    return history