app.py

import os
import subprocess
import gradio as gr
from transformers import AutoTokenizer
from optimum.intel.openvino import OVModelForCausalLM
from sentence_transformers import SentenceTransformer
import faiss
import numpy as np
import warnings

warnings.filterwarnings(
    "ignore",
    category=DeprecationWarning,
    message="__array__ implementation doesn't accept a copy keyword"
)

# 設定模型 ID 與轉換後存檔路徑（8-bit 量化版）
model_id = "agentica-org/DeepScaleR-1.5B-Preview"
export_path = "exported_model_openvino_int8"

print("Loading model as OpenVINO int8 (8-bit) model...")

if os.path.exists(export_path) and os.listdir(export_path):
    print(f"Found quantized OpenVINO model at '{export_path}', loading it...")
    model = OVModelForCausalLM.from_pretrained(export_path, device_map="auto", use_cache=False)
else:
    print("No quantized model found, exporting and quantizing to OpenVINO int8 now...")
    # 透過 optimum-cli 導出並量化模型（此命令行參數根據你的任務可能需要調整）
    command = [
        "optimum-cli", "export", "openvino",
        "--model", model_id,
        "--task", "text-generation",
        "--weight-format", "int8",
        export_path
    ]
    subprocess.run(command, check=True)
    print(f"Quantized model saved to '{export_path}'.")
    model = OVModelForCausalLM.from_pretrained(export_path, device_map="auto", use_cache=False)

print("Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)

# 載入向量模型（用於將文本轉換為向量）
encoder = SentenceTransformer("all-MiniLM-L6-v2")

# FAQ 知識庫（問題 + 回答）
faq_data = [
    ("What is FAISS?", "FAISS is a library for efficient similarity search and clustering of dense vectors."),
    ("How does FAISS work?", "FAISS uses indexing structures to quickly retrieve the nearest neighbors of a query vector."),
    ("Can FAISS run on GPU?", "Yes, FAISS supports GPU acceleration for faster computation."),
    ("What is OpenVINO?", "OpenVINO is an inference engine optimized for Intel hardware."),
    ("How to fine-tune a model?", "Fine-tuning involves training a model on a specific dataset to adapt it to a particular task."),
    ("What is the best way to optimize inference speed?", "Using quantization and model distillation can significantly improve inference speed.")
]

# 將 FAQ 問題轉換為向量
faq_questions = [q for q, _ in faq_data]
faq_answers = [a for _, a in faq_data]
faq_vectors = np.array(encoder.encode(faq_questions)).astype("float32")

# 建立 FAISS 索引（使用 L2 距離）
d = faq_vectors.shape[1]  # 向量維度
index = faiss.IndexFlatL2(d)
index.add(faq_vectors)

# 對話歷史記錄
history = []

# 查詢函數：先嘗試從 FAQ 中檢索答案，若無匹配則使用 OpenVINO 模型生成回答
def respond(prompt):
    global history
    # 將輸入轉換為向量，並使用 FAISS 查詢最相近的 FAQ 問題
    query_vector = np.array(encoder.encode([prompt])).astype("float32")
    D, I = index.search(query_vector, 1)
    
    if D[0][0] < 1.0:
        response = faq_answers[I[0][0]]
    else:
        # 若 FAQ 無匹配，則使用 OpenVINO 模型生成回答
        messages = [{"role": "system", "content": "Answer the question in English only."}]
        for user_text, assistant_text in history:
            messages.append({"role": "user", "content": user_text})
            messages.append({"role": "assistant", "content": assistant_text})
        messages.append({"role": "user", "content": prompt})
        
        # 將對話訊息組成一個 prompt（以換行分隔）
        chat_prompt = "\n".join([f"{msg['role']}: {msg['content']}" for msg in messages])
        model_inputs = tokenizer(chat_prompt, return_tensors="pt").to(model.device)
        generated_ids = model.generate(
            **model_inputs,
            max_new_tokens=512,
            temperature=0.7,
            top_p=0.9,
            do_sample=True
        )
        response = tokenizer.decode(generated_ids[0], skip_special_tokens=True).strip()
    
    history.append((prompt, response))
    return response

# 清除對話歷史記錄
def clear_history():
    global history
    history = []
    return "History cleared!"

# 建立 Gradio 介面
with gr.Blocks() as demo:
    gr.Markdown("# DeepScaleR-1.5B-Preview (OpenVINO int8) Chatbot with FAISS FAQ")
    
    with gr.Tabs():
        with gr.TabItem("Chat"):
            chat_interface = gr.Interface(
                fn=respond,
                inputs=gr.Textbox(label="Prompt", placeholder="Enter your message..."),
                outputs=gr.Textbox(label="Response", interactive=False),
                api_name="hchat",
                title="DeepScaleR-1.5B-Preview (OpenVINO int8) Chatbot",
                description="This chatbot first searches an FAQ database using FAISS, then uses an OpenVINO 8-bit model to generate a response if no FAQ match is found."
            )
    
    with gr.Row():
        clear_button = gr.Button("🧹 Clear History")
    
    clear_button.click(fn=clear_history, inputs=[], outputs=[])

if __name__ == "__main__":
    print("Launching Gradio app...")
    demo.launch(server_name="0.0.0.0", server_port=7860, share=True)