# app.py (为 Hugging Face ZeroGPU 修改)
import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
import spaces
# --------------------------------------------------------------------------
# 1. 配置 (在应用启动时执行)
# --------------------------------------------------------------------------
# !! 重要的模型 ID (从 HF Hub 加载)
model_id = "AIDC-AI/Marco-MT-Algharb"
# --- ZeroGPU 修改 1:
# 在启动时 *只* 定义全局变量为 None
# 大模型将在第一个请求到来时被加载
# ---
model = None
tokenizer = None
generation_config = None
print("ZeroGPU 启动脚本开始...")
print(f"准备从 {model_id} 加载 Tokenizer...")
# Tokenizer 很小, 可以在启动时加载
# ★★★ 提醒: 这仍然需要你已在 Space settings 中设置 HF_TOKEN 密钥 ★★★
try:
tokenizer = AutoTokenizer.from_pretrained(
model_id,
trust_remote_code=True
)
print("Tokenizer 加载成功!")
# --- ZeroGPU 修改 2:
# Tokenizer 加载成功后, *立即* 定义 GenerationConfig
# (这解决了你之前关于 Qwen3 停止 token 的问题)
# ---
im_end_id = tokenizer.convert_tokens_to_ids("<|im_end|>")
eot_id = tokenizer.eos_token_id
print(f"设置停止 IDs: <|im_end|_id={im_end_id}, <|endoftext|_id={eot_id}")
generation_config = GenerationConfig(
do_sample=False,
max_new_tokens=512,
eos_token_id=[im_end_id, eot_id],
pad_token_id=eot_id
)
print("GenerationConfig 配置成功。")
except Exception as e:
print(f"Tokenizer 加载失败: {e}")
print("!! 严重错误: 如果这是 Gated Repo 问题, 请确保 HF_TOKEN 密钥已设置并重启 Space。")
# 语言代码到全名的映射 (保持不变)
source_lang_name_map = {
"en": "english",
"ja": "japanese",
"cs": "czech",
"de": "german",
}
target_lang_name_map = {
"zh": "chinese",
"ko": "korean",
"ja": "japanese",
"ar": "arabic",
"cs": "czech",
"ru": "russian",
"uk": "ukraine",
"et": "estonian",
"bho": "bhojpuri",
"sr_latin": "serbian",
"de": "german",
}
# --------------------------------------------------------------------------
# 2. 定义核心翻译函数 (修改版)
# --------------------------------------------------------------------------
@spaces.GPU
def translate(source_text, source_lang_code, target_lang_code):
"""
接收用户输入并返回翻译结果
(ZeroGPU: 在首次调用时加载模型)
"""
global model # ★★★ 关键: 引用全局 'model' 变量
# --- ZeroGPU 修改 3: 首次调用时加载模型 ---
if model is None:
if tokenizer is None:
return "错误：Tokenizer 未能成功加载，无法继续。请检查启动日志。"
print("--- 首次请求 ---")
print("检测到模型未加载。正在加载模型到 ZeroGPU (Nvidia H200)...")
try:
# 这一步会触发 ZeroGPU 分配 H200
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto", # 'auto' 将会检测到 H200
trust_remote_code=True
)
model.eval()
print("模型已成功加载到 GPU!")
except Exception as e:
print(f"在首次加载时模型失败: {e}")
return f"错误：模型在加载到 GPU 时失败: {e}"
# -----------------------------------------
# (从这里开始, 代码与之前版本相同)
# 简单的输入验证
if not source_text or not source_text.strip():
return ""
source_language_name = source_lang_name_map.get(source_lang_code, "the source language")
target_language_name = target_lang_name_map.get(target_lang_code, "the target language")
prompt = (
f"Human: Please translate the following text into {target_language_name}: \n"
f"{source_text}<|im_end|>\n"
f"Assistant:"
)
try:
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
outputs = model.generate(
**inputs,
generation_config=generation_config
)
input_length = inputs.input_ids.shape[1]
generated_ids = outputs[0][input_length:]
generated_text = tokenizer.decode(generated_ids, skip_special_tokens=True).strip()
return generated_text
except Exception as e:
print(f"翻译过程中出错: {e}")
return f"翻译时发生错误: {e}"
# --------------------------------------------------------------------------
# 3. 创建并配置 Gradio 界面 (这部分保持不变)
# --------------------------------------------------------------------------
# <--- 定义自定义 CSS 样式 --->
css = """
/* ... 你的所有 CSS 样式 ... */
.gradio-textbox {
min-height: 300px !important;
}
"""
# <--- 修复: choices 定义 ---
source_lang_choices = [(name.capitalize(), code) for code, name in source_lang_name_map.items()]
target_lang_choices = [(name.capitalize(), code) for code, name in target_lang_name_map.items()]
# <--- 使用 gr.Blocks 并保持主题 --->
with gr.Blocks(
theme=gr.themes.Soft(primary_hue="amber", secondary_hue="amber"),
css=css,
) as demo:
gr.HTML(
"""
"""
)
# --- 标题 ---
gr.HTML(f"""
""")
# --- 翻译器主界面 (两栏布局) ---
with gr.Row(variant="panel", equal_height=True):
# --- 左侧输入卡片 ---
with gr.Group():
source_lang_dd = gr.Dropdown(
choices=source_lang_choices,
value="en",
label="源语言 (Source Language)"
)
source_text_tb = gr.Textbox(
lines=10,
label="源文本 (Source Text)",
placeholder="Enter text to translate here...",
elem_classes=["gradio-textbox"]
)
# --- 右侧输出卡片 ---
with gr.Group():
target_lang_dd = gr.Dropdown(
choices=target_lang_choices,
value="zh",
label="目标语言 (Target Language)"
)
output_text_tb = gr.Textbox(
lines=10,
label="翻译结果 (Translation)",
interactive=False,
elem_classes=["gradio-textbox"]
)
# --- 按钮行 ---
with gr.Row():
clear_btn = gr.ClearButton(
value="清除 (Clear)",
components=[source_text_tb, output_text_tb, source_lang_dd, target_lang_dd]
)
submit_btn = gr.Button("翻译 (Submit)", variant="primary", scale=1)
# --- 示例 ---
example_list = [
["The quick brown fox jumps over the lazy dog.", "en", "zh"],
["The sunset painted the sky with brilliant shades of orange and purple.", "en", "ko"],
["The ancient ruins stand as a silent testament to the rise and fall of a great civilization.", "en", "ja"],
]
gr.Examples(
examples=example_list,
inputs=[source_text_tb, source_lang_dd, target_lang_dd]
)
# --- 支持的语向卡片 ---
gr.HTML(f"""
Supported Language Pairs
From English (en):
- en2zh
- en2ja
- en2ko
- en2ar
- en2et
- en2sr_latin
- en2ru
- en2uk
- en2cs
- en2bho
From Czech (cs):
From Japanese (ja):
""")
# --- 设定按钮的点击逻辑 ---
submit_btn.click(
fn=translate,
inputs=[source_text_tb, source_lang_dd, target_lang_dd],
outputs=[output_text_tb],
api_name="translate"
)
# 启动应用
if __name__ == "__main__":
demo.launch() 
