app.py

import spaces
import gradio as gr
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

# Model configuration
model_name = "ai4bharat/IndicBART"

# Load tokenizer and model on CPU
print("Loading IndicBART tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(model_name, do_lower_case=False, use_fast=False, keep_accents=True)

print("Loading IndicBART model on CPU...")
model = AutoModelForSeq2SeqLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="cpu")

# Language mapping
LANGUAGE_CODES = {
    "Assamese": "<2as>",
    "Bengali": "<2bn>", 
    "English": "<2en>",
    "Gujarati": "<2gu>",
    "Hindi": "<2hi>",
    "Kannada": "<2kn>",
    "Malayalam": "<2ml>",
    "Marathi": "<2mr>",
    "Oriya": "<2or>",
    "Punjabi": "<2pa>",
    "Tamil": "<2ta>",
    "Telugu": "<2te>"
}

@spaces.GPU(duration=60)
def generate_response(input_text, source_lang, target_lang, task_type, max_length):
    """Generate response using IndicBART"""
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model_gpu = model.to(device)
    
    # Get language codes
    src_code = LANGUAGE_CODES[source_lang]
    tgt_code = LANGUAGE_CODES[target_lang]
    
    # Format input based on task type
    if task_type == "Translation":
        formatted_input = f"{input_text} </s> {src_code}"
        decoder_start_token = tgt_code
    elif task_type == "Text Completion":
        # For completion, use target language
        formatted_input = f"{input_text} </s> {tgt_code}"
        decoder_start_token = tgt_code
    else:  # Text Generation
        formatted_input = f"{input_text} </s> {src_code}"
        decoder_start_token = tgt_code
    
    # Tokenize input
    inputs = tokenizer(formatted_input, return_tensors="pt", padding=True, truncation=True, max_length=512)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    # Get decoder start token id
    decoder_start_token_id = tokenizer._convert_token_to_id_with_added_voc(decoder_start_token)
    
    # Generate
    with torch.no_grad():
        outputs = model_gpu.generate(
            **inputs,
            decoder_start_token_id=decoder_start_token_id,
            max_length=max_length,
            num_beams=4,
            early_stopping=True,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id,
            use_cache=True
        )
    
    # Decode output
    generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
    
    # Move model back to CPU
    model_gpu.cpu()
    torch.cuda.empty_cache()
    
    return generated_text

# Create Gradio interface
with gr.Blocks(title="IndicBART Multilingual Assistant", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🇮🇳 IndicBART Multilingual Assistant
    
    Experience IndicBART - trained on **11 Indian languages**! Perfect for translation, text completion, and multilingual generation.
    
    **Supported Languages**: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, Telugu, English
    """)
    
    with gr.Row():
        with gr.Column(scale=3):
            input_text = gr.Textbox(
                label="Input Text",
                placeholder="Enter text in any supported language...",
                lines=3
            )
            
            output_text = gr.Textbox(
                label="Generated Output",
                lines=5,
                interactive=False
            )
            
            generate_btn = gr.Button("Generate", variant="primary", size="lg")
        
        with gr.Column(scale=1):
            task_type = gr.Dropdown(
                choices=["Translation", "Text Completion", "Text Generation"],
                value="Translation",
                label="Task Type"
            )
            
            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=50,
                maximum=300,
                value=100,
                step=10,
                label="Max Length"
            )
    
    # Examples
    gr.Markdown("### 💡 Try these examples:")
    
    examples = [
        ["Hello, how are you?", "English", "Hindi", "Translation", 100],
        ["मैं एक छात्र हूं", "Hindi", "English", "Translation", 100],
        ["আমি ভাত খাই", "Bengali", "English", "Translation", 100],
        ["भारत एक", "Hindi", "Hindi", "Text Completion", 150],
        ["The capital of India", "English", "English", "Text Completion", 100]
    ]
    
    gr.Examples(
        examples=examples,
        inputs=[input_text, source_lang, target_lang, task_type, max_length],
        outputs=output_text,
        fn=generate_response
    )
    
    # Connect generate button
    generate_btn.click(
        generate_response,
        inputs=[input_text, source_lang, target_lang, task_type, max_length],
        outputs=output_text
    )

if __name__ == "__main__":
    demo.launch()