app.py

import gradio as gr
import torch
import os

# You must use the exact same model name as your repo
MODEL_ID = "nineninesix/Kani-TTS-370m"

@spaces.GPU
def generate_speech(text: str, model_choice: str, speaker_display: str):
    if not text.strip():
        return "Please enter text for speech generation.", None
    
    try:
        device = "cuda" if torch.cuda.is_available() else "cpu"
        print(f"Using device: {device}")
        
        # --- This is the key part to load a specific model ---
        if model_choice not in MODELS:
            return f"Model '{model_choice}' not found.", None
            
        selected_model = MODELS[model_choice]
        
        # --- This part handles speakers ---
        cfg = selected_model[1]  # Model config
        speaker_map = cfg.get('speaker_id', {}) if cfg is not None else {}
        if speaker_display and speaker_map:
            speaker_id = speaker_map.get(speaker_display)
        else:
            speaker_id = None
            
        print(f"Generating speech with {model_choice}...")
        
        # --- Use the specific part of the model for generation ---
        model_to_generate = selected_model[0]
        audio, _, time_report = model_to_generate.run_model(
            text=text, 
            speaker_id=speaker_id,
            temperature=0.7,
            repetition_penalty=1.2,
            max_tokens=1024
        )
        
        sample_rate = 22050
        print("Speech generation completed!")
        
        return (sample_rate, audio), time_report

def load_models():
    global MODELS
    if not MODELS:
        print("Loading models into GPU memory...")
        from transformers import AutoModel
        model_path = MODEL_ID
        
        # Load both the main model and its config
        model = AutoModel.from_pretrained(model_path, trust_remote_code=True)
        config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
        
        MODELS = {
            "Kani TTS 370M": (model, config)
        }
        
        print(f"Models loaded. Available speakers: {list(config.speaker_id.keys()) if config.speaker_id else []}")
        return MODELS

# --- Gradio interface setup ---
MODELS = load_models()

with gr.Blocks(title="😻 KaniTTS - Text to Speech") as demo:
    gr.Markdown("# 😻 KaniTTS: Fast and Expressive Speech Generation Model")
    
    model_dropdown = gr.Dropdown(
        choices=list(MODELS.keys()),
        value=list(MODELS.keys())[0],
        label="Selected Model"
    )
    
    # --- Speaker selector (populated on model load) ---
    all_speakers = list(MODELS[list(MODELS.keys())[0]][1].speaker_id.keys()) if MODELS and MODELS[list(MODELS.keys())[0]][1] and MODELS[list(MODELS.keys())[0]][1].speaker_id else []
    speaker_dropdown = gr.Dropdown(
        choices=all_speakers,
        value=None,
        label="Speaker",
        visible=True,
        allow_custom_value=True
    )
    
    text_input = gr.Textbox(label="Text", lines=5)
    
    generate_btn = gr.Button("Generate Speech", variant="primary")
    
    audio_output = gr.Audio(label="Generated Audio", type="numpy")
    
    # --- Event handlers ---
    model_dropdown.change(
        fn=lambda choice: gr.update(choices=list(MODELS[choice][1].speaker_id.keys()), value=None, visible=True) if MODELS and MODELS[choice][1].speaker_id else gr.update(visible=False),
        inputs=[model_dropdown],
        outputs=[speaker_dropdown]
    )
    
    generate_btn.click(
        fn=generate_speech,
        inputs=[text_input, model_dropdown, speaker_dropdown],
        outputs=[audio_output]
    )
    
    # --- This is the API enabling line ---
    demo.queue().launch(show_api=True)