Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,12 +1,47 @@
|
|
| 1 |
# app.py
|
| 2 |
|
| 3 |
import os
|
|
|
|
|
|
|
| 4 |
import gradio as gr
|
| 5 |
from typing import List, Dict
|
| 6 |
from huggingface_hub import hf_hub_download
|
| 7 |
-
from llama_cpp import Llama
|
| 8 |
|
| 9 |
-
# ---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
|
| 11 |
# 您指定的模型資訊
|
| 12 |
MODEL_NAME = "Qwen3-0.6B-Q8_0.gguf"
|
|
@@ -16,39 +51,36 @@ MODEL_REPO = "Qwen/Qwen3-0.6B-GGUF"
|
|
| 16 |
DEFAULT_SYSTEM_MESSAGE = "You are a friendly and helpful assistant. Please answer the user's questions concisely and accurately."
|
| 17 |
|
| 18 |
# 步驟 1: 下載 GGUF 模型
|
| 19 |
-
# 模型會被下載到 ~/.cache/huggingface/hub/ 或指定的快取目錄
|
| 20 |
try:
|
| 21 |
print(f"嘗試從 {MODEL_REPO} 下載 {MODEL_NAME}...")
|
| 22 |
model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_NAME)
|
| 23 |
print(f"模型下載完成,路徑: {model_path}")
|
| 24 |
except Exception as e:
|
| 25 |
-
print(f"
|
| 26 |
-
#
|
| 27 |
-
# 這裡可以選擇性地退出或使用本地路徑作為備用(如果存在)。
|
| 28 |
-
exit(1)
|
| 29 |
|
| 30 |
|
| 31 |
-
# ---
|
| 32 |
|
| 33 |
# 步驟 2: 初始化 Llama.cpp 實例
|
| 34 |
-
# n_gpu_layers=0 表示不使用 GPU (CPU 推論),如果環境支援 CUDA/cuBLAS,可以設定為 >0
|
| 35 |
try:
|
| 36 |
print("正在初始化 Llama.cpp 實例...")
|
| 37 |
llm = Llama(
|
| 38 |
model_path=model_path,
|
| 39 |
n_ctx=4096, # 上下文長度
|
| 40 |
n_batch=512, # 批次大小
|
| 41 |
-
|
|
|
|
| 42 |
n_gpu_layers=0, # CPU 推論
|
| 43 |
verbose=False # 關閉內部日誌輸出
|
| 44 |
)
|
| 45 |
print("Llama.cpp 模型加載成功。")
|
| 46 |
except Exception as e:
|
| 47 |
print(f"**錯誤**:Llama.cpp 實例初始化失敗。錯誤訊息: {e}")
|
| 48 |
-
exit(1)
|
| 49 |
|
| 50 |
|
| 51 |
-
# ---
|
| 52 |
|
| 53 |
def llama_inference(
|
| 54 |
message: str,
|
|
@@ -60,32 +92,25 @@ def llama_inference(
|
|
| 60 |
) -> str:
|
| 61 |
"""
|
| 62 |
使用 Llama.cpp 實例執行推論並返回回應。
|
| 63 |
-
|
| 64 |
-
:param message: 當前的使用者輸入。
|
| 65 |
-
:param chat_history: Gradio 傳遞的聊天歷史記錄 (list of [user, bot] pairs)。
|
| 66 |
-
:return: LLM 的回應文字。
|
| 67 |
"""
|
| 68 |
|
| 69 |
-
#
|
| 70 |
messages = [{"role": "system", "content": system_message}]
|
| 71 |
|
| 72 |
for human, assistant in chat_history:
|
| 73 |
-
# 歷史對話
|
| 74 |
messages.append({"role": "user", "content": human})
|
| 75 |
messages.append({"role": "assistant", "content": assistant})
|
| 76 |
|
| 77 |
-
# 當前訊息
|
| 78 |
messages.append({"role": "user", "content": message})
|
| 79 |
|
| 80 |
|
| 81 |
try:
|
| 82 |
-
# 呼叫 Llama.cpp 的 create_chat_completion 介面
|
| 83 |
response = llm.create_chat_completion(
|
| 84 |
messages=messages,
|
| 85 |
max_tokens=max_tokens,
|
| 86 |
temperature=temperature,
|
| 87 |
top_p=top_p,
|
| 88 |
-
# stream=False 是預設值
|
| 89 |
)
|
| 90 |
|
| 91 |
# 解析回應
|
|
@@ -100,23 +125,16 @@ def llama_inference(
|
|
| 100 |
return f"❌ 伺服器錯誤 (Llama.cpp 推論失敗): {e}"
|
| 101 |
|
| 102 |
|
| 103 |
-
# ---
|
| 104 |
|
| 105 |
-
# 定義 Gradio 聊天函式 (用於更新介面)
|
| 106 |
def chat_interface(message: str, history: List[List[str]]):
|
| 107 |
"""Gradio 介面調用函式。"""
|
| 108 |
|
| 109 |
-
# 這裡可以固定或從另一個��入元件獲取參數,為了簡化,使用硬編碼值
|
| 110 |
response = llama_inference(
|
| 111 |
message=message,
|
| 112 |
chat_history=history,
|
| 113 |
-
system_message=DEFAULT_SYSTEM_MESSAGE,
|
| 114 |
-
max_tokens=4096,
|
| 115 |
-
temperature=0.7,
|
| 116 |
-
top_p=0.95
|
| 117 |
)
|
| 118 |
|
| 119 |
-
# Gradio 聊天介面要求回傳回應文字
|
| 120 |
return response
|
| 121 |
|
| 122 |
|
|
@@ -129,30 +147,22 @@ with gr.Blocks(title="Qwen3-0.6B-GGUF 聊天機器人") as demo:
|
|
| 129 |
"""
|
| 130 |
)
|
| 131 |
|
| 132 |
-
# 聊天元件
|
| 133 |
chatbot = gr.Chatbot(
|
| 134 |
label="聊天記錄",
|
| 135 |
height=500
|
| 136 |
)
|
| 137 |
|
| 138 |
-
# 聊天輸入元件
|
| 139 |
chat_input = gr.Textbox(
|
| 140 |
show_label=False,
|
| 141 |
placeholder="請輸入你的問題...",
|
| 142 |
container=False
|
| 143 |
)
|
| 144 |
|
| 145 |
-
# 綁定聊天邏輯
|
| 146 |
-
# submit 觸發事件:
|
| 147 |
-
# - fn: 要執行的 Python 函式 (chat_interface)
|
| 148 |
-
# - inputs: 函式接收的輸入 ([Textbox 的內容, Chatbot 的歷史])
|
| 149 |
-
# - outputs: 函式輸出的結果 (Chatbot 的新歷史)
|
| 150 |
chat_input.submit(
|
| 151 |
fn=chat_interface,
|
| 152 |
inputs=[chat_input, chatbot],
|
| 153 |
outputs=chatbot
|
| 154 |
).then(
|
| 155 |
-
# 清空輸入框
|
| 156 |
fn=lambda: "",
|
| 157 |
inputs=None,
|
| 158 |
outputs=chat_input,
|
|
@@ -161,6 +171,4 @@ with gr.Blocks(title="Qwen3-0.6B-GGUF 聊天機器人") as demo:
|
|
| 161 |
|
| 162 |
# 啟動應用程式
|
| 163 |
if __name__ == "__main__":
|
| 164 |
-
# 在 Gradio Space 中,會使用 gunicorn 或類似服務來運行,但如果要在本地測試,可以使用以下命令:
|
| 165 |
-
# python app.py
|
| 166 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
|
| 1 |
# app.py
|
| 2 |
|
| 3 |
import os
|
| 4 |
+
import sys
|
| 5 |
+
import subprocess
|
| 6 |
import gradio as gr
|
| 7 |
from typing import List, Dict
|
| 8 |
from huggingface_hub import hf_hub_download
|
|
|
|
| 9 |
|
| 10 |
+
# --- 0. 內嵌安裝 llama-cpp-python ---
|
| 11 |
+
# 警告:這是一個非標準且可能失敗的解決方案。
|
| 12 |
+
# 建議在 Gradio Space 中使用 requirements.txt 來安裝依賴。
|
| 13 |
+
try:
|
| 14 |
+
print("--- 嘗試動態安裝 llama-cpp-python ---")
|
| 15 |
+
# 執行 pip install 命令
|
| 16 |
+
# 使用 sys.executable 確保使用當前的 Python 解譯器
|
| 17 |
+
subprocess.check_call([
|
| 18 |
+
sys.executable,
|
| 19 |
+
"-m",
|
| 20 |
+
"pip",
|
| 21 |
+
"install",
|
| 22 |
+
"llama-cpp-python",
|
| 23 |
+
"--upgrade" # 確保是最新版本
|
| 24 |
+
])
|
| 25 |
+
print("llama-cpp-python 安裝/更新成功。")
|
| 26 |
+
except subprocess.CalledProcessError as e:
|
| 27 |
+
print(f"**致命錯誤**:llama-cpp-python 安裝失敗。請檢查環境權限或系統依賴。錯誤訊息: {e}")
|
| 28 |
+
# 由於安裝失敗,我們不能繼續執行
|
| 29 |
+
sys.exit(1)
|
| 30 |
+
except Exception as e:
|
| 31 |
+
print(f"**致命錯誤**:發生未知錯誤。錯誤訊息: {e}")
|
| 32 |
+
sys.exit(1)
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
# --- 1. 引入 llama_cpp ---
|
| 36 |
+
# 必須在嘗試安裝之後才能引入
|
| 37 |
+
try:
|
| 38 |
+
from llama_cpp import Llama
|
| 39 |
+
except ImportError:
|
| 40 |
+
print("**致命錯誤**:即使嘗試安裝,仍然無法引入 llama_cpp。請檢查 pip 安裝日誌。")
|
| 41 |
+
sys.exit(1)
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
# --- 2. 模型設定與下載 ---
|
| 45 |
|
| 46 |
# 您指定的模型資訊
|
| 47 |
MODEL_NAME = "Qwen3-0.6B-Q8_0.gguf"
|
|
|
|
| 51 |
DEFAULT_SYSTEM_MESSAGE = "You are a friendly and helpful assistant. Please answer the user's questions concisely and accurately."
|
| 52 |
|
| 53 |
# 步驟 1: 下載 GGUF 模型
|
|
|
|
| 54 |
try:
|
| 55 |
print(f"嘗試從 {MODEL_REPO} 下載 {MODEL_NAME}...")
|
| 56 |
model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_NAME)
|
| 57 |
print(f"模型下載完成,路徑: {model_path}")
|
| 58 |
except Exception as e:
|
| 59 |
+
print(f"**錯誤**:無法下載模型。錯誤訊息: {e}")
|
| 60 |
+
sys.exit(1) # 無法下載模型則退出
|
|
|
|
|
|
|
| 61 |
|
| 62 |
|
| 63 |
+
# --- 3. Llama.cpp 初始化 ---
|
| 64 |
|
| 65 |
# 步驟 2: 初始化 Llama.cpp 實例
|
|
|
|
| 66 |
try:
|
| 67 |
print("正在初始化 Llama.cpp 實例...")
|
| 68 |
llm = Llama(
|
| 69 |
model_path=model_path,
|
| 70 |
n_ctx=4096, # 上下文長度
|
| 71 |
n_batch=512, # 批次大小
|
| 72 |
+
# 為了 Gradio Space 穩定性,使用少量 CPU 核心
|
| 73 |
+
n_threads=os.cpu_count() // 2 or 1,
|
| 74 |
n_gpu_layers=0, # CPU 推論
|
| 75 |
verbose=False # 關閉內部日誌輸出
|
| 76 |
)
|
| 77 |
print("Llama.cpp 模型加載成功。")
|
| 78 |
except Exception as e:
|
| 79 |
print(f"**錯誤**:Llama.cpp 實例初始化失敗。錯誤訊息: {e}")
|
| 80 |
+
sys.exit(1)
|
| 81 |
|
| 82 |
|
| 83 |
+
# --- 4. 推論核心函式 ---
|
| 84 |
|
| 85 |
def llama_inference(
|
| 86 |
message: str,
|
|
|
|
| 92 |
) -> str:
|
| 93 |
"""
|
| 94 |
使用 Llama.cpp 實例執行推論並返回回應。
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
"""
|
| 96 |
|
| 97 |
+
# 格式化訊息列表,包含系統提示和聊天歷史
|
| 98 |
messages = [{"role": "system", "content": system_message}]
|
| 99 |
|
| 100 |
for human, assistant in chat_history:
|
|
|
|
| 101 |
messages.append({"role": "user", "content": human})
|
| 102 |
messages.append({"role": "assistant", "content": assistant})
|
| 103 |
|
|
|
|
| 104 |
messages.append({"role": "user", "content": message})
|
| 105 |
|
| 106 |
|
| 107 |
try:
|
| 108 |
+
# 呼叫 Llama.cpp 的 create_chat_completion 介面
|
| 109 |
response = llm.create_chat_completion(
|
| 110 |
messages=messages,
|
| 111 |
max_tokens=max_tokens,
|
| 112 |
temperature=temperature,
|
| 113 |
top_p=top_p,
|
|
|
|
| 114 |
)
|
| 115 |
|
| 116 |
# 解析回應
|
|
|
|
| 125 |
return f"❌ 伺服器錯誤 (Llama.cpp 推論失敗): {e}"
|
| 126 |
|
| 127 |
|
| 128 |
+
# --- 5. Gradio 介面設定 ---
|
| 129 |
|
|
|
|
| 130 |
def chat_interface(message: str, history: List[List[str]]):
|
| 131 |
"""Gradio 介面調用函式。"""
|
| 132 |
|
|
|
|
| 133 |
response = llama_inference(
|
| 134 |
message=message,
|
| 135 |
chat_history=history,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 136 |
)
|
| 137 |
|
|
|
|
| 138 |
return response
|
| 139 |
|
| 140 |
|
|
|
|
| 147 |
"""
|
| 148 |
)
|
| 149 |
|
|
|
|
| 150 |
chatbot = gr.Chatbot(
|
| 151 |
label="聊天記錄",
|
| 152 |
height=500
|
| 153 |
)
|
| 154 |
|
|
|
|
| 155 |
chat_input = gr.Textbox(
|
| 156 |
show_label=False,
|
| 157 |
placeholder="請輸入你的問題...",
|
| 158 |
container=False
|
| 159 |
)
|
| 160 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 161 |
chat_input.submit(
|
| 162 |
fn=chat_interface,
|
| 163 |
inputs=[chat_input, chatbot],
|
| 164 |
outputs=chatbot
|
| 165 |
).then(
|
|
|
|
| 166 |
fn=lambda: "",
|
| 167 |
inputs=None,
|
| 168 |
outputs=chat_input,
|
|
|
|
| 171 |
|
| 172 |
# 啟動應用程式
|
| 173 |
if __name__ == "__main__":
|
|
|
|
|
|
|
| 174 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|