app.py

import spaces
import gradio as gr
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from IndicTransToolkit.processor import IndicProcessor

# Model configurations
INDIC_EN_MODEL = "ai4bharat/indictrans2-indic-en-1B"
EN_INDIC_MODEL = "ai4bharat/indictrans2-en-indic-1B"

print("Loading IndicTrans2 models...")
# Load tokenizers
indic_en_tokenizer = AutoTokenizer.from_pretrained(INDIC_EN_MODEL, trust_remote_code=True)
en_indic_tokenizer = AutoTokenizer.from_pretrained(EN_INDIC_MODEL, trust_remote_code=True)

# Load models on CPU
indic_en_model = AutoModelForSeq2SeqLM.from_pretrained(
    INDIC_EN_MODEL,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="cpu"
)

en_indic_model = AutoModelForSeq2SeqLM.from_pretrained(
    EN_INDIC_MODEL,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="cpu"
)

# Initialize IndicProcessor (CRUCIAL for proper preprocessing)
ip = IndicProcessor(inference=True)

# Language mappings (exact codes from official documentation)
LANGUAGE_CODES = {
    "Assamese": "asm_Beng",
    "Bengali": "ben_Beng", 
    "Bodo": "brx_Deva",
    "Dogri": "doi_Deva",
    "Gujarati": "guj_Gujr",
    "Hindi": "hin_Deva",
    "Kannada": "kan_Knda",
    "Kashmiri (Arabic)": "kas_Arab",
    "Kashmiri (Devanagari)": "kas_Deva",
    "Konkani": "gom_Deva",
    "Maithili": "mai_Deva",
    "Malayalam": "mal_Mlym",
    "Manipuri (Bengali)": "mni_Beng",
    "Manipuri (Meitei)": "mni_Mtei",
    "Marathi": "mar_Deva",
    "Nepali": "npi_Deva",
    "Odia": "ory_Orya",
    "Punjabi": "pan_Guru",
    "Sanskrit": "san_Deva",
    "Santali": "sat_Olck",
    "Sindhi (Arabic)": "snd_Arab",
    "Sindhi (Devanagari)": "snd_Deva",
    "Tamil": "tam_Taml",
    "Telugu": "tel_Telu",
    "Urdu": "urd_Arab",
    "English": "eng_Latn"
}

@spaces.GPU(duration=120)
def translate_text(input_text, source_lang, target_lang, max_length):
    """Translate using IndicTrans2 with proper preprocessing"""
    
    if not input_text.strip():
        return "Please enter text to translate."
    
    try:
        device = "cuda" if torch.cuda.is_available() else "cpu"
        
        # Get language codes
        src_code = LANGUAGE_CODES[source_lang]
        tgt_code = LANGUAGE_CODES[target_lang]
        
        # Determine direction and select appropriate model/tokenizer
        if source_lang == "English" and target_lang != "English":
            # English to Indic
            model_gpu = en_indic_model.to(device)
            tokenizer = en_indic_tokenizer
            direction = "en_to_indic"
        elif source_lang != "English" and target_lang == "English":
            # Indic to English
            model_gpu = indic_en_model.to(device)
            tokenizer = indic_en_tokenizer
            direction = "indic_to_en"
        else:
            return "Please select English as either source or target language (not both)."
        
        # CRUCIAL: Use IndicProcessor for proper preprocessing
        input_sentences = [input_text.strip()]
        
        # Preprocess using IndicProcessor (this handles the proper formatting)
        batch = ip.preprocess_batch(
            input_sentences,
            src_lang=src_code,
            tgt_lang=tgt_code,
        )
        
        # Tokenize the preprocessed batch
        inputs = tokenizer(
            batch,
            truncation=True,
            padding="longest",
            return_tensors="pt",
            return_attention_mask=True,
        ).to(device)
        
        # Generate translation
        with torch.no_grad():
            generated_tokens = model_gpu.generate(
                **inputs,
                use_cache=True,
                min_length=0,
                max_length=max_length,
                num_beams=5,
                num_return_sequences=1,
                early_stopping=True,
                pad_token_id=tokenizer.pad_token_id,
                eos_token_id=tokenizer.eos_token_id
            )
        
        # Decode generated tokens
        generated_tokens = tokenizer.batch_decode(
            generated_tokens,
            skip_special_tokens=True,
            clean_up_tokenization_spaces=True,
        )
        
        # CRUCIAL: Postprocess using IndicProcessor
        translations = ip.postprocess_batch(generated_tokens, lang=tgt_code)
        
        # Move model back to CPU
        model_gpu.cpu()
        torch.cuda.empty_cache()
        
        return translations[0] if translations else "Translation failed."
        
    except Exception as e:
        if 'model_gpu' in locals():
            model_gpu.cpu()
        torch.cuda.empty_cache()
        return f"Error during translation: {str(e)}"

# Create Gradio interface
with gr.Blocks(title="IndicTrans2 Official Translator", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🇮🇳 IndicTrans2 - Official AI4Bharat Translator
    
    High-quality neural machine translation between English and 22 Indian languages.
    Uses official IndicTransToolkit for proper preprocessing.
    
    **Supported Languages**: Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Kannada, Kashmiri, 
    Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, 
    Sindhi, Tamil, Telugu, Urdu.
    """)
    
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(
                label="Input Text",
                placeholder="Enter text to translate...",
                lines=5
            )
            
            with gr.Row():
                source_lang = gr.Dropdown(
                    choices=list(LANGUAGE_CODES.keys()),
                    value="English",
                    label="Source Language"
                )
                
                target_lang = gr.Dropdown(
                    choices=list(LANGUAGE_CODES.keys()),
                    value="Hindi",
                    label="Target Language"
                )
            
            max_length = gr.Slider(
                minimum=32,
                maximum=256,
                value=128,
                step=16,
                label="Max Output Length"
            )
            
            translate_btn = gr.Button("Translate", variant="primary", size="lg")
        
        with gr.Column():
            output_text = gr.Textbox(
                label="Translation",
                lines=5,
                interactive=False
            )
            
            clear_btn = gr.Button("Clear", variant="secondary")
    
    # Examples from official documentation
    gr.Markdown("### 💡 Official Examples:")
    
    examples = [
        ["When I was young, I used to go to the park every day.", "English", "Hindi", 128],
        ["We watched a new movie last week, which was very inspiring.", "English", "Bengali", 128],
        ["जब मैं छोटा था, मैं हर रोज़ पार्क जाता था।", "Hindi", "English", 128],
        ["हमने पिछले सप्ताह एक नई फिल्म देखी जो कि बहुत प्रेरणादायक थी।", "Hindi", "English", 128],
        ["Technology is changing our world rapidly.", "English", "Tamil", 128]
    ]
    
    gr.Examples(
        examples=examples,
        inputs=[input_text, source_lang, target_lang, max_length],
        outputs=output_text,
        fn=translate_text
    )
    
    # Event handlers
    def clear_all():
        return "", ""
    
    translate_btn.click(
        translate_text,
        inputs=[input_text, source_lang, target_lang, max_length],
        outputs=output_text
    )
    
    clear_btn.click(
        clear_all,
        outputs=[input_text, output_text]
    )

if __name__ == "__main__":
    demo.launch(
        share=True,
        show_error=True
    )