app.py

import gradio as gr
from flores import code_mapping
from functools import lru_cache
import openai  # 用于调用外部API
import os
import spaces
import gradio as gr
from sacremoses import MosesPunctNormalizer
from stopes.pipelines.monolingual.utils.sentence_split import get_split_algo
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from flores import code_mapping
import platform
import torch
import nltk
from functools import lru_cache


print(os.getenv('key'))
code_mapping = dict(sorted(code_mapping.items(), key=lambda item: item[0]))
flores_codes = list(code_mapping.keys())
target_languages = flores_codes  # 简化列表

# 假设openai_client已定义，例如：

device = "cpu" if platform.system() == "Darwin" else "cuda"
MODEL_NAME = "ByteDance-Seed/Seed-X-PPO-7B"

code_mapping = dict(sorted(code_mapping.items(), key=lambda item: item[0]))
flores_codes = list(code_mapping.keys())


def load_model():
    model = AutoModel.from_pretrained(MODEL_NAME).to(device)
    print(f"Model loaded in {device}")
    return model


model = load_model()


# Loading the tokenizer once, because re-loading it takes about 1.5 seconds each time
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)



@lru_cache(maxsize=100)
def translate(text: str, src_lang: str, tgt_lang: str):
    if not src_lang:
        raise gr.Error("The source language is empty! Please choose it in the dropdown list.")
    if not tgt_lang:
        raise gr.Error("The target language is empty! Please choose it in the dropdown list.")
    return _translate(text, src_lang, tgt_lang)

# Only assign GPU if cache not used
@spaces.GPU
def _translate(text: str, src_lang: str, tgt_lang: str):
    src_code = code_mapping[src_lang]
    tgt_code = code_mapping[tgt_lang]
    tokenizer.src_lang = src_code
    tokenizer.tgt_lang = tgt_code

    # normalizing the punctuation first
    text = punct_normalizer.normalize(text)

    paragraphs = text.split("\n")
    translated_paragraphs = []

    for paragraph in paragraphs:
        splitter = get_language_specific_sentence_splitter(src_code)
        sentences = list(splitter(paragraph))
        translated_sentences = []

        for sentence in sentences:
            input_tokens = (
                tokenizer(sentence, return_tensors="pt")
                .input_ids[0]
                .cpu()
                .numpy()
                .tolist()
            )
            translated_chunk = model.generate(
                input_ids=torch.tensor([input_tokens]).to(device),
                forced_bos_token_id=tokenizer.convert_tokens_to_ids(tgt_code),
                max_length=len(input_tokens) + 50,
                num_return_sequences=1,
                num_beams=5,
                no_repeat_ngram_size=4,  # repetition blocking works better if this number is below num_beams  
                renormalize_logits=True,  # recompute token probabilities after banning the repetitions
            )
            translated_chunk = tokenizer.decode(
                translated_chunk[0], skip_special_tokens=True
            )
            translated_sentences.append(translated_chunk)

        translated_paragraph = " ".join(translated_sentences)
        translated_paragraphs.append(translated_paragraph)

    return "\n".join(translated_paragraphs)


# def _translate(text: str, src_lang: str, tgt_lang: str):
#     prompt = f"Translate the following text from {src_lang} to {tgt_lang}. Direct output translation result without any explaination:\n\n{text}"
#     key=os.getenv('key')
#     openai_client = openai.OpenAI(base_url="https://ssapi.cppbear.site/v1", api_key=key)
#     response = openai_client.chat.completions.create(
#         model="tbai.xin-dpsk-deepseek-v3",  # 如gpt-3.5-turbo或其他兼容模型
#         messages=[{"role": "user", "content": prompt}],
#         max_tokens=30240,
#         temperature=0.0
#     )
#     print(response)
#     return response.choices[0].message.content.strip()


description = """
<div style="text-align: center;">
    <img src="https://github.com/user-attachments/assets/c42e675e-497c-4508-8bb9-093ad4d1f216" alt="UNESCO Meta Hugging Face Banner" style="max-width: 800px; width: 100%; margin: 0 auto;">
    <h1 style="color: #0077be; font-size: 3em;">Seed-X, powered by Bytedance</h1>
</div>
We are excited to introduce Seed-X, a powerful series of open-source multilingual translation language models, including an instruction model, a reinforcement learning model, and a reward model. It pushes the boundaries of translation capabilities within 7 billion parameters. We develop Seed-X as an accessible, off-the-shelf tool to support the community in advancing translation research and applications:
"""

examples_inputs = [["Seed-X is indeed a good translation model ","English","Chinese"],]

with gr.Blocks() as demo:
    gr.Markdown(description)
    with gr.Row():
        src_lang = gr.Dropdown(label="Source Language", choices=flores_codes)
        target_lang = gr.Dropdown(label="Target Language", choices=target_languages)
    with gr.Row():
        input_text = gr.Textbox(label="Input Text", lines=6)
    with gr.Row():
        btn = gr.Button("Translate text")
    with gr.Row():
        output = gr.Textbox(label="Output Text", lines=6)
    btn.click(
        translate,
        inputs=[input_text, src_lang, target_lang],
        outputs=output,
    )
    examples = gr.Examples(examples=examples_inputs,inputs=[input_text, src_lang,target_lang], fn=translate, outputs=output, cache_examples=True)
demo.launch()