app.py

import gradio as gr
import torch
import os
# optimum.onnxruntime 에서 __version__ import 제거
from transformers import AutoTokenizer, __version__ as transformers_version
from optimum.onnxruntime import ORTModelForCausalLM
# import optimum # optimum 자체의 버전 확인 시도 (선택적)

# --- Configuration ---
MODEL_ID = "onnx-community/gemma-3-1b-it-ONNX-GQA"
ONNX_FILE_NAME = None

print(f"Using Transformers version: {transformers_version}")
# try:
#     print(f"Using Optimum version: {optimum.__version__}") # 다른 방법으로 버전 확인 시도
# except AttributeError:
#     print("Could not determine Optimum version automatically.")
print(f"Using Gradio version: {gr.__version__}")

# --- Device Selection ---
try:
    if torch.cuda.is_available():
        device = "cuda:0"
        provider = "CUDAExecutionProvider"
        print("Attempting to use GPU (CUDA).")
    else:
        device = "cpu"
        provider = "CPUExecutionProvider"
        print("Using CPU.")
except Exception as e:
    print(f"Device detection error: {e}. Defaulting to CPU.")
    device = "cpu"
    provider = "CPUExecutionProvider"

# --- Model and Tokenizer Loading ---
model = None
tokenizer = None
model_loaded_successfully = False

print(f"Attempting to load model: {MODEL_ID}")
print(f"Using device: {device}, Execution Provider: {provider}")

try:
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
    print("Tokenizer loaded successfully.")

    # ONNX 모델 로드 시도
    model = ORTModelForCausalLM.from_pretrained(
        MODEL_ID,
        provider=provider,
        use_cache=True,
    )
    print(f"ONNX Model '{MODEL_ID}' loaded successfully with provider '{provider}'.")
    model_loaded_successfully = True

except ValueError as ve:
    # 모델 타입 미지원 오류 처리
    print(f"!!!!!!!!!!!!!! CRITICAL MODEL LOADING ERROR (ValueError) !!!!!!!!!!!!!!")
    print(f"Model: {MODEL_ID}")
    print(f"Error message: {ve}")
    print("This likely means the installed 'transformers' library version does NOT support the 'gemma3_text' architecture.")
    print("Ensure 'requirements.txt' specifies a recent version (e.g., transformers>=4.41.0) and the Space has been rebuilt/restarted.")
    print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
    model_loaded_successfully = False

except Exception as e:
    # 다른 예외 처리
    print(f"!!!!!!!!!!!!!! UNEXPECTED MODEL LOADING ERROR !!!!!!!!!!!!!!")
    print(f"Model: {MODEL_ID}")
    print(f"Error type: {type(e).__name__}")
    print(f"Error message: {e}")
    print("Check Space resources (memory limits), network connection, or other dependencies.")
    print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
    model_loaded_successfully = False

# --- Chat Function ---
def chat_function(message: str, history: list):
    if not model_loaded_successfully or model is None or tokenizer is None:
        return "Error: The AI model is not loaded. Please check the application logs."

    try:
        # 채팅 기록 변환
        chat_messages = [{"role": "system", "content": "You are a helpful AI assistant."}]
        for user_msg, model_msg in history:
            if user_msg: chat_messages.append({"role": "user", "content": user_msg})
            if model_msg: chat_messages.append({"role": "model", "content": model_msg})
        if message: chat_messages.append({"role": "user", "content": message})

        # 프롬프트 생성
        prompt = ""
        try:
             prompt = tokenizer.apply_chat_template(chat_messages, tokenize=False, add_generation_prompt=True)
        except Exception as template_error:
             print(f"Warning: Failed to apply chat template ({template_error}). Using manual prompt construction.")
             prompt_parts = ["<start_of_turn>system\nYou are a helpful AI assistant.<end_of_turn>"]
             for user_msg, model_msg in history:
                 if user_msg: prompt_parts.append(f"<start_of_turn>user\n{user_msg}<end_of_turn>")
                 if model_msg: prompt_parts.append(f"<start_of_turn>model\n{model_msg}<end_of_turn>")
             if message: prompt_parts.append(f"<start_of_turn>user\n{message}<end_of_turn>")
             prompt_parts.append("<start_of_turn>model")
             prompt = "\n".join(prompt_parts)

        # 입력 토큰화
        inputs = tokenizer(prompt, return_tensors="pt").to(device)

        # 응답 생성
        print("Generating response...")
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=512,
                do_sample=True,
                temperature=0.7,
                top_k=50,
                top_p=0.9,
                pad_token_id=tokenizer.eos_token_id
            )
        print("Generation complete.")

        # 디코딩
        input_token_len = inputs['input_ids'].shape[1]
        generated_tokens = outputs[0][input_token_len:]
        response = tokenizer.decode(generated_tokens, skip_special_tokens=True)
        response = response.replace("<end_of_turn>", "").strip()
        if not response:
            print("Warning: Generated empty response.")
            response = "Sorry, I couldn't generate a response for that."
        return response

    except Exception as e:
        print(f"!!!!!!!!!!!!!! Error during generation !!!!!!!!!!!!!!")
        print(f"Error type: {type(e).__name__}")
        print(f"Error message: {e}")
        print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
        return f"Sorry, an error occurred during response generation. Please check logs."

# --- Gradio Interface ---
print("Creating Gradio Interface...")
iface = gr.ChatInterface(
    fn=chat_function,
    title="AI Assistant (Gemma 3 1B ONNX-GQA)",
    description=f"Chat with {MODEL_ID}. Model loaded: {model_loaded_successfully}",
    chatbot=gr.Chatbot(height=600, type="messages", bubble_full_width=False),
    theme=gr.themes.Soft(),
    examples=[["Hello!"], ["Write a poem about the internet."]]
)

# --- Launch App ---
if __name__ == "__main__":
    print("Launching Gradio App...")
    iface.launch()