app.py

import gradio as gr
import spaces
import torch
from transformers import T5Tokenizer, T5ForConditionalGeneration, AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForCausalLM, AutoModel, pipeline, logging
import languagecodes
import requests, os
import polars as pl 

logging.set_verbosity_error()
favourite_langs = {"German": "de", "Romanian": "ro", "English": "en", "-----": "-----"}
df = pl.read_parquet("isolanguages.parquet")
non_empty_isos = df.slice(1).filter(pl.col("ISO639-1") != "").rows()
# all_langs = languagecodes.iso_languages_byname
all_langs = {iso[0]: (iso[1], iso[2], iso[3]) for iso in non_empty_isos}

# Language options as list, add favourite languages first
options = list(favourite_langs.keys())
options.extend(list(all_langs.keys()))

models = ["Helsinki-NLP", "Helsinki-NLP/opus-mt-tc-bible-big-mul-mul",
          "facebook/nllb-200-distilled-600M", "facebook/nllb-200-distilled-1.3B", "facebook/nllb-200-1.3B", "facebook/nllb-200-3.3B",
          "facebook/mbart-large-50-many-to-many-mmt", "facebook/mbart-large-50-one-to-many-mmt", "facebook/mbart-large-50-many-to-one-mmt",
          "facebook/m2m100_418M", "facebook/m2m100_1.2B",
          "bigscience/mt0-small", "bigscience/mt0-base", "bigscience/mt0-large", "bigscience/mt0-xl",
          "bigscience/bloomz-560m", "bigscience/bloomz-1b1", "bigscience/bloomz-1b7", "bigscience/bloomz-3b",
          "t5-small", "t5-base", "t5-large",
          "google/flan-t5-small", "google/flan-t5-base", "google/flan-t5-large", "google/flan-t5-xl",
          "Argos", "Google",
          "HuggingFaceTB/SmolLM3-3B", "winninghealth/WiNGPT-Babel-2",
          "utter-project/EuroLLM-1.7B", "utter-project/EuroLLM-1.7B-Instruct",
          "Unbabel/Tower-Plus-2B", "Unbabel/TowerInstruct-7B-v0.2", "Unbabel/TowerInstruct-Mistral-7B-v0.2",
          "openGPT-X/Teuken-7B-instruct-commercial-v0.4", "openGPT-X/Teuken-7B-instruct-v0.6"
          ]

def model_to_cuda(model):
    # Move the model to GPU if available
    if torch.cuda.is_available():
        model = model.to('cuda')
        print("CUDA is available! Using GPU.")
    else:
        print("CUDA not available! Using CPU.")
    return model

def HelsinkiNLPAutoTokenizer(sl, tl, input_text): # deprecated
    if model_name == "Helsinki-NLP":
        message_text = f'Translated from {sl} to {tl} with {model_name}.'
        try:
            model_name = f"Helsinki-NLP/opus-mt-{sl}-{tl}"
            tokenizer = AutoTokenizer.from_pretrained(model_name)
            model = model_to_cuda(AutoModelForSeq2SeqLM.from_pretrained(model_name))
        except EnvironmentError:
            try:   
                model_name = f"Helsinki-NLP/opus-tatoeba-{sl}-{tl}"
                tokenizer = AutoTokenizer.from_pretrained(model_name)
                model = model_to_cuda(AutoModelForSeq2SeqLM.from_pretrained(model_name))
                input_ids = tokenizer.encode(prompt, return_tensors="pt")
                output_ids = model.generate(input_ids, max_length=512)
                translated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
                return translated_text, message_text
            except EnvironmentError as error:
                return f"Error finding model: {model_name}! Try other available language combination.", error
                      
class Translators:
    def __init__(self, model_name: str, sl: str, tl: str, input_text: str):
        self.model_name = model_name
        self.sl, self.tl = sl, tl
        self.input_text = input_text
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        
    def google(self):  
        url = os.environ['GCLIENT'] + f'sl={self.sl}&tl={self.tl}&q={self.input_text}'
        response = requests.get(url)
        return response.json()[0][0][0]
    
    @classmethod
    def download_argos_model(cls, from_code, to_code):
        import argostranslate.package
        print('Downloading model', from_code, to_code) 
        # Download and install Argos Translate package
        argostranslate.package.update_package_index()
        available_packages = argostranslate.package.get_available_packages()
        package_to_install = next(
            filter(lambda x: x.from_code == from_code and x.to_code == to_code, available_packages)
        )
        argostranslate.package.install_from_path(package_to_install.download())
    
    def argos(self):
        import argostranslate.translate, argostranslate.package
        try:
            Translators.download_argos_model(self.sl, self.tl) # Download model
            translated_text = argostranslate.translate.translate(self.input_text, self.sl, self.tl) # Translate
        except StopIteration:
            # packages_info = ', '.join(f"{pkg.get_description()}->{str(pkg.links)} {str(pkg.source_languages)}" for pkg in argostranslate.package.get_available_packages())
            packages_info = ', '.join(f"{pkg.from_name} ({pkg.from_code}) -> {pkg.to_name} ({pkg.to_code})" for pkg in argostranslate.package.get_available_packages())
            translated_text = f"No Argos model for {self.sl} to {self.tl}. Try other model or languages combination from the available Argos models: {packages_info}."
        except Exception as error:
            translated_text = error
        return translated_text

    def HelsinkiNLP_mulmul(self):
        try:
            pipe = pipeline("translation", model=self.model_name, device=self.device)                
            iso_dict = {iso[1]: iso[3] for iso in non_empty_isos}
            iso3tl = iso_dict.get(self.tl) # 'deu', 'ron', 'eng', 'fra'
            translation = pipe(f'>>{iso3tl}<< {self.input_text}')
            return translation[0]['translation_text'], f'Translated from {self.sl} to {self.tl} with {self.model_name}.'
        except Exception as error:
            return f"Error translating with model: {self.model_name}! Try other available language combination.", error
    
    def HelsinkiNLP(self):
        try: # Standard bilingual model
            model_name = f"Helsinki-NLP/opus-mt-{self.sl}-{self.tl}"
            pipe = pipeline("translation", model=model_name, device=self.device)
            translation = pipe(self.input_text)
            return translation[0]['translation_text'], f'Translated from {self.sl} to {self.tl} with {model_name}.'
        except EnvironmentError:
            try: # Tatoeba models
                model_name = f"Helsinki-NLP/opus-tatoeba-{self.sl}-{self.tl}"
                pipe = pipeline("translation", model=model_name, device=self.device)
                translation = pipe(self.input_text)
                return translation[0]['translation_text'], f'Translated from {self.sl} to {self.tl} with {model_name}.'
            except EnvironmentError as error:
                self.model_name = "Helsinki-NLP/opus-mt-tc-bible-big-mul-mul" # Last resort: try multi to multi
                return self.HelsinkiNLP_mulmul()
        except KeyError as error:
            return f"Error: Translation direction {self.sl} to {self.tl} is not supported by Helsinki Translation Models", error
        
    def smollm(self):
        tokenizer = AutoTokenizer.from_pretrained(self.model_name)
        model = AutoModelForCausalLM.from_pretrained(self.model_name)
        prompt = f"""Translate the following {self.sl} text to {self.tl}, generating only the translated text and maintaining the original meaning and tone:
        {self.input_text}
        Translation:"""
        inputs = tokenizer(prompt, return_tensors="pt")
        outputs = model.generate(
            inputs.input_ids,
            max_length=len(inputs.input_ids[0]) + 150,
            temperature=0.3,
            do_sample=True
        ) 
        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
        print(response)
        return response.split("Translation:")[-1].strip()

    def flan(self):
        tokenizer = T5Tokenizer.from_pretrained(self.model_name, legacy=False)
        model = T5ForConditionalGeneration.from_pretrained(self.model_name)
        prompt = f"translate {self.sl} to {self.tl}: {self.input_text}"
        input_ids = tokenizer(prompt, return_tensors="pt").input_ids
        outputs = model.generate(input_ids)
        return tokenizer.decode(outputs[0], skip_special_tokens=True).strip()

    def tfive(self):
        tokenizer = T5Tokenizer.from_pretrained(self.model_name)
        model = T5ForConditionalGeneration.from_pretrained(self.model_name, device_map="auto")
        prompt = f"translate {self.sl} to {self.tl}: {self.input_text}"
        input_ids = tokenizer.encode(prompt, return_tensors="pt")
        output_ids = model.generate(input_ids, max_length=512)
        translated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True).strip()
        return translated_text
    
    def mbart_many_to_many(self):
        from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
        model = MBartForConditionalGeneration.from_pretrained(self.model_name)
        tokenizer = MBart50TokenizerFast.from_pretrained(self.model_name)
        # translate source to target
        tokenizer.src_lang = languagecodes.mbart_large_languages[self.sl]
        encoded = tokenizer(self.input_text, return_tensors="pt")
        generated_tokens = model.generate(
            **encoded,
            forced_bos_token_id=tokenizer.lang_code_to_id[languagecodes.mbart_large_languages[self.tl]]
        )
        return tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]
    
    def mbart_one_to_many(self):
        # translate from English
        from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
        model = MBartForConditionalGeneration.from_pretrained(self.model_name)
        tokenizer = MBart50TokenizerFast.from_pretrained(self.model_name, src_lang="en_XX")
        model_inputs = tokenizer(self.input_text, return_tensors="pt")
        langid = languagecodes.mbart_large_languages[self.tl]
        generated_tokens = model.generate(
            **model_inputs,
            forced_bos_token_id=tokenizer.lang_code_to_id[langid]
        )
        return tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]
    
    def mbart_many_to_one(self):
        # translate to English
        from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
        model = MBartForConditionalGeneration.from_pretrained(self.model_name)
        tokenizer = MBart50TokenizerFast.from_pretrained(self.model_name)
        tokenizer.src_lang = languagecodes.mbart_large_languages[self.sl]
        encoded = tokenizer(self.input_text, return_tensors="pt")
        generated_tokens = model.generate(**encoded)
        return tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]
        
    def mtom(self):
        from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer
        model = M2M100ForConditionalGeneration.from_pretrained(self.model_name)
        tokenizer = M2M100Tokenizer.from_pretrained(self.model_name)
        tokenizer.src_lang = self.sl
        encoded = tokenizer(self.input_text, return_tensors="pt")
        generated_tokens = model.generate(**encoded, forced_bos_token_id=tokenizer.get_lang_id(self.tl))
        return tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]

    def bigscience(self):  
        tokenizer = AutoTokenizer.from_pretrained(self.model_name)
        model = AutoModelForSeq2SeqLM.from_pretrained(self.model_name)
        self.input_text = self.input_text if self.input_text.endswith('.') else f'{self.input_text}.'
        inputs = tokenizer.encode(f"Translate to {self.tl}: {self.input_text}", return_tensors="pt")
        outputs = model.generate(inputs)
        translation = tokenizer.decode(outputs[0])
        translation = translation.replace('<pad> ', '').replace('</s>', '')
        return translation
    
    def bloomz(self):  
        tokenizer = AutoTokenizer.from_pretrained(self.model_name)
        model = AutoModelForCausalLM.from_pretrained(self.model_name)
        self.input_text = self.input_text if self.input_text.endswith('.') else f'{self.input_text}.'
        # inputs = tokenizer.encode(f"Translate from {self.sl} to {self.tl}: {self.input_text} Translation:", return_tensors="pt")
        inputs = tokenizer.encode(f"Translate to {self.tl}: {self.input_text}", return_tensors="pt")
        outputs = model.generate(inputs)
        translation = tokenizer.decode(outputs[0])
        translation = translation.replace('<pad> ', '').replace('</s>', '')
        translation = translation.split('Translation:')[-1].strip() if 'Translation:' in translation else translation.strip()
        return translation
    
    def nllb(self):
        tokenizer = AutoTokenizer.from_pretrained(self.model_name, src_lang=self.sl)
        # model = AutoModelForSeq2SeqLM.from_pretrained(self.model_name, device_map="auto", torch_dtype=torch.bfloat16)
        model = AutoModelForSeq2SeqLM.from_pretrained(self.model_name)
        translator = pipeline('translation', model=model, tokenizer=tokenizer, src_lang=self.sl, tgt_lang=self.tl)
        translated_text = translator(self.input_text, max_length=512)
        return translated_text[0]['translation_text']
    
    def wingpt(self):
        model = AutoModelForCausalLM.from_pretrained(
           self.model_name,
           torch_dtype="auto",
           device_map="auto"
        )
        tokenizer = AutoTokenizer.from_pretrained(self.model_name)
        # input_json = '{"input_text": self.input_text}'
        messages = [
           {"role": "system", "content": f"Translate this to {self.tl} language"}, 
           {"role": "user", "content": self.input_text}
        ]
        
        text = tokenizer.apply_chat_template(
           messages,
           tokenize=False,
           add_generation_prompt=True
        )
        model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
        
        generated_ids = model.generate(
           **model_inputs,
           max_new_tokens=512,
           temperature=0.1
        )
        
        generated_ids = [
           output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
        ]
        print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True))
        output = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
        result = output.split('\n')[-1].strip() if '\n' in output else output.strip()
        return result
    
    def eurollm(self):
        tokenizer = AutoTokenizer.from_pretrained(self.model_name)
        model = AutoModelForCausalLM.from_pretrained(self.model_name)  
        prompt = f"{self.sl}: {self.input_text} {self.tl}:"
        inputs = tokenizer(prompt, return_tensors="pt")
        outputs = model.generate(**inputs, max_new_tokens=512)
        output = tokenizer.decode(outputs[0], skip_special_tokens=True)
        print(output)
        # result = output.rsplit(f'{self.tl}:')[-1].strip() if f'{self.tl}:' in output else output.strip()
        result = output.rsplit(f'{self.tl}:')[-1].strip() if '\n' in output or f'{self.tl}:' in output else output.strip()
        return result
    
    def eurollm_instruct(self):
        tokenizer = AutoTokenizer.from_pretrained(self.model_name)
        model = AutoModelForCausalLM.from_pretrained(self.model_name)
        text = f'<|im_start|>system\n<|im_end|>\n<|im_start|>user\nTranslate the following {self.sl} source text to {self.tl}:\n{self.sl}: {self.input_text} \n{self.tl}: <|im_end|>\n<|im_start|>assistant\n'
        inputs = tokenizer(text, return_tensors="pt")
        outputs = model.generate(**inputs, max_new_tokens=512)
        output = tokenizer.decode(outputs[0], skip_special_tokens=True)
        if f'{self.tl}:' in output:
            output = output.rsplit(f'{self.tl}:')[-1].strip().replace('assistant\n', '').strip()
        return output

    def teuken(self):
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        model = AutoModelForCausalLM.from_pretrained(
            self.model_name,
            trust_remote_code=True,
            torch_dtype=torch.bfloat16,
        )
        model = model.to(device).eval()
        tokenizer = AutoTokenizer.from_pretrained(
            self.model_name,
            use_fast=False,
            trust_remote_code=True,
        )
        translation_prompt = f"Translate the following text from {self.sl} into {self.tl}: {self.input_text}"
        messages = [{"role": "User", "content": translation_prompt}]
        prompt_ids = tokenizer.apply_chat_template(messages, chat_template="EN", tokenize=True, add_generation_prompt=False, return_tensors="pt")
        prediction = model.generate(
            prompt_ids.to(model.device),
            max_length=512,
            do_sample=True,
            top_k=50,
            top_p=0.95,
            temperature=0.7,
            num_return_sequences=1,
        )
        translation = tokenizer.decode(prediction[0].tolist())
        return translation

    def unbabel(self):
        pipe = pipeline("text-generation", model=self.model_name, torch_dtype=torch.bfloat16, device_map="auto")
        messages = [{"role": "user",
                     "content": f"Translate the following text from {self.sl} into {self.tl}.\n{self.sl}: {self.input_text}.\n{self.tl}:"}]
        prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
        tokenized_input = pipe.tokenizer(self.input_text, return_tensors="pt")
        num_input_tokens = len(tokenized_input["input_ids"][0])
        max_new_tokens = round(num_input_tokens + 0.25 * num_input_tokens)
        outputs = pipe(prompt, max_new_tokens=max_new_tokens, do_sample=False)
        translated_text = outputs[0]["generated_text"]
        print(f"Input chars: {len(input_text)}", f"Input tokens: {num_input_tokens}", f"max_new_tokens: {max_new_tokens}",
              "Chars to tokens ratio:", round(len(input_text) / num_input_tokens, 2), f"Raw translation: {translated_text}")
        markers = ["<end_of_turn>", "<|im_end|>", "<|im_start|>assistant"] # , "\n" 
        for marker in markers:
            if marker in translated_text:
                translated_text = translated_text.split(marker)[1].strip()
        translated_text = translated_text.replace('Answer:', '', 1).strip() if translated_text.startswith('Answer:') else translated_text
        translated_text = translated_text.split("Translated text:")[0].strip() if "Translated text:" in translated_text else translated_text
        split_translated_text = translated_text.split('\n', translated_text.count('\n'))
        translated_text = '\n'.join(split_translated_text[:input_text.count('\n')+1])
        return translated_text

@spaces.GPU
def translate_text(input_text: str, s_language: str, t_language: str, model_name: str) -> tuple[str, str]:
    """
    Translates the input text from the source language to the target language  using a specified model.

    Parameters:
        input_text (str): The source text to be translated
        s_language (str): The source language of the input text
        t_language (str): The target language in which the input text is translated
        model_name (str): The selected translation model name

    Returns:
        tuple: 
            translated_text(str): The input text translated to the selected target language
            message_text(str):  A descriptive message summarizing the translation process. Example: "Translated from English to German with Helsinki-NLP."
    
    Example:
        >>> translate_text("Hello world", "English", "German", "Helsinki-NLP")
        ("Hallo Welt", "Translated from English to German with Helsinki-NLP.")
    """
    
    sl = all_langs[s_language][0]
    tl = all_langs[t_language][0]
    message_text = f'Translated from {s_language} to {t_language} with {model_name}'
    try:    
        if model_name == "Helsinki-NLP/opus-mt-tc-bible-big-mul-mul":
            translated_text, message_text = Translators(model_name, sl, tl, input_text).HelsinkiNLP_mulmul()
        
        elif model_name == "Helsinki-NLP":
            translated_text, message_text = Translators(model_name, sl, tl, input_text).HelsinkiNLP()
        
        elif model_name == 'Argos':
            translated_text = Translators(model_name, sl, tl, input_text).argos()
    
        elif model_name == 'Google':
            translated_text = Translators(model_name, sl, tl, input_text).google()
    
        elif "m2m" in model_name.lower():
            translated_text = Translators(model_name, sl, tl, input_text).mtom()
        
        elif model_name.startswith('t5'):
            translated_text = Translators(model_name, s_language, t_language, input_text).tfive()
            
        elif 'flan' in model_name.lower():
            translated_text = Translators(model_name, s_language, t_language, input_text).flan()
            
        elif 'mt0' in model_name.lower():
            translated_text = Translators(model_name, s_language, t_language, input_text).bigscience()
    
        elif 'bloomz' in model_name.lower():
            translated_text = Translators(model_name, s_language, t_language, input_text).bloomz()
            
        elif 'nllb' in model_name.lower():
            nnlbsl, nnlbtl = languagecodes.nllb_language_codes[s_language], languagecodes.nllb_language_codes[t_language]
            translated_text = Translators(model_name, nnlbsl, nnlbtl, input_text).nllb()
        
        elif model_name == "facebook/mbart-large-50-many-to-many-mmt":
            translated_text = Translators(model_name, s_language, t_language, input_text).mbart_many_to_many()
    
        elif model_name == "facebook/mbart-large-50-one-to-many-mmt":
            translated_text = Translators(model_name, s_language, t_language, input_text).mbart_one_to_many()
    
        elif model_name == "facebook/mbart-large-50-many-to-one-mmt":
            translated_text = Translators(model_name, s_language, t_language, input_text).mbart_many_to_one()
        
        elif 'teuken' in model_name.lower():
            translated_text = Translators(model_name, s_language, t_language, input_text).teuken()
    
        elif model_name == "utter-project/EuroLLM-1.7B-Instruct":
            translated_text = Translators(model_name, s_language, t_language, input_text).eurollm_instruct()
        
        elif model_name == "utter-project/EuroLLM-1.7B":
            translated_text = Translators(model_name, s_language, t_language, input_text).eurollm()
                
        elif 'Unbabel' in model_name:   
            translated_text = Translators(model_name, s_language, t_language, input_text).unbabel()
            
        elif model_name == "HuggingFaceTB/SmolLM3-3B":
            translated_text = Translators(model_name, s_language, t_language, input_text).smollm()

        elif model_name == "winninghealth/WiNGPT-Babel-2":      
            translated_text = Translators(model_name, s_language, t_language, input_text).wingpt()
            
    except Exception as error:
        translated_text = error
    finally:
        print(input_text, translated_text, message_text)
        return translated_text, message_text

# Function to swap dropdown values
def swap_languages(src_lang, tgt_lang):
    return tgt_lang, src_lang 

def create_interface():
    with gr.Blocks() as interface:
        gr.Markdown("### Machine Text Translation with Gradio API and MCP Server")

        with gr.Row():
            input_text = gr.Textbox(label="Enter text to translate:", placeholder="Type your text here, maximum 512 tokens")
        
        with gr.Row():
            s_language = gr.Dropdown(choices=options, value = options[0], label="Source language", interactive=True)
            t_language = gr.Dropdown(choices=options, value = options[1], label="Target language", interactive=True)
            swap_button = gr.Button("Swap Languages", size="md")
            swap_button.click(fn=swap_languages, inputs=[s_language, t_language], outputs=[s_language, t_language], api_name=False, show_api=False)

        model_name = gr.Dropdown(choices=models, label=f"Select a model. Default is {models[0]}.", value = models[0], interactive=True)
        translate_button = gr.Button("Translate")

        translated_text = gr.Textbox(label="Translated text:", placeholder="Display field for translation", interactive=False, show_copy_button=True)
        message_text = gr.Textbox(label="Messages:", placeholder="Display field for status and error messages", interactive=False,
                                  value=f'Default translation settings: from {s_language.value} to {t_language.value} with {model_name.value}.')
        allmodels = gr.HTML(label="Model links:", value=', '.join([f'<a href="https://huggingface.co/{model}">{model}</a>' for model in models]))

        translate_button.click(
            fn=translate_text, 
            inputs=[input_text, s_language, t_language, model_name], 
            outputs=[translated_text, message_text]
        )

    return interface

interface = create_interface()
if __name__ == "__main__":
    interface.launch(mcp_server=True)
    # interface.queue().launch(server_name="0.0.0.0", show_error=True, server_port=7860, mcp_server=True)