app.py

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr
from snac import SNAC

def redistribute_codes(row):
    """
    Convert a sequence of token codes into an audio waveform using SNAC.
    The code assumes each 7 tokens represent one group of instructions.
    """
    row_length = row.size(0)
    new_length = (row_length // 7) * 7
    trimmed_row = row[:new_length]
    code_list = [t - 128266 for t in trimmed_row]
    
    layer_1, layer_2, layer_3 = [], [], []
    
    for i in range((len(code_list) + 1) // 7):
        layer_1.append(code_list[7 * i][None])
        layer_2.append(code_list[7 * i + 1][None] - 4096)
        layer_3.append(code_list[7 * i + 2][None] - (2 * 4096))
        layer_3.append(code_list[7 * i + 3][None] - (3 * 4096))
        layer_2.append(code_list[7 * i + 4][None] - (4 * 4096))
        layer_3.append(code_list[7 * i + 5][None] - (5 * 4096))
        layer_3.append(code_list[7 * i + 6][None] - (6 * 4096))
    
    with torch.no_grad():
        codes = [
            torch.concat(layer_1),
            torch.concat(layer_2),
            torch.concat(layer_3)
        ]
        for i in range(len(codes)):
            codes[i][codes[i] < 0] = 0
            codes[i] = codes[i][None]
        
        audio_hat = snac_model.decode(codes)
        return audio_hat.cpu()[0, 0]

# Load the SNAC model (shared by all)
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to("cuda")

# Load all the single-speaker language models
models = {
    "Luna": {
        "tokenizer": AutoTokenizer.from_pretrained('prithivMLmods/Llama-3B-Mono-Luna'),
        "model": AutoModelForCausalLM.from_pretrained('prithivMLmods/Llama-3B-Mono-Luna', torch_dtype=torch.bfloat16).cuda()
    },
    "Ceylia": {
        "tokenizer": AutoTokenizer.from_pretrained('prithivMLmods/Llama-3B-Mono-Ceylia'),
        "model": AutoModelForCausalLM.from_pretrained('prithivMLmods/Llama-3B-Mono-Ceylia', torch_dtype=torch.bfloat16).cuda()
    },
    "Cooper": {
        "tokenizer": AutoTokenizer.from_pretrained('prithivMLmods/Llama-3B-Mono-Cooper'),
        "model": AutoModelForCausalLM.from_pretrained('prithivMLmods/Llama-3B-Mono-Cooper', torch_dtype=torch.bfloat16).cuda()
    },
    "Jim": {
        "tokenizer": AutoTokenizer.from_pretrained('prithivMLmods/Llama-3B-Mono-Jim'),
        "model": AutoModelForCausalLM.from_pretrained('prithivMLmods/Llama-3B-Mono-Jim', torch_dtype=torch.bfloat16).cuda()
    },
}

def generate_audio(text, temperature, top_p, max_new_tokens, model_name):
    """
    Given input text and model parameters, generate speech audio using the chosen model.
    """
    # Retrieve the chosen tokenizer and model
    chosen = models[model_name]
    tokenizer = chosen["tokenizer"]
    model = chosen["model"]
    
    prompt = f'<custom_token_3><|begin_of_text|>{text}<|eot_id|><custom_token_4><custom_token_5><custom_token_1>'
    input_ids = tokenizer(prompt, add_special_tokens=False, return_tensors='pt').to('cuda')
    
    with torch.no_grad():
        generated_ids = model.generate(
            **input_ids,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=temperature,
            top_p=top_p,
            repetition_penalty=1.1,
            num_return_sequences=1,
            eos_token_id=128258,
        )
    
    row = generated_ids[0, input_ids['input_ids'].shape[1]:]
    y_tensor = redistribute_codes(row)
    y_np = y_tensor.detach().cpu().numpy()
    return (24000, y_np)

# Example texts with emotion tokens
example_texts = [
    ["Hi, my name is Alex. <laugh> It's a wonderful day! <chuckle> I love coding."],
    ["I woke up feeling sleepy. <yawn> I need coffee! <sniffle> But I'm ready to work."],
    ["Oh no, I forgot my keys! <groan> <uhm> Maybe I'll try again later. <sigh>"],
    ["This is amazing! <gasp> Really, it's fantastic. <giggles>"]
]

# Gradio Interface
with gr.Blocks() as demo:
    # Sidebar for model selection
    with gr.Sidebar():
        gr.Markdown("# Choose Model")
        model_choice = gr.Dropdown(choices=list(models.keys()), value="Luna", label="Model")
    
    gr.Markdown("# Single Speaker Audio Generation")
    gr.Markdown("Generate speech audio using one of the single-speaker models. Use the examples below to see how emotion tokens like `<laugh>`, `<chuckle>`, `<sigh>`, etc. can be incorporated.")
    
    with gr.Row():
        text_input = gr.Textbox(lines=4, label="Input Text")
    
    # Examples with emotion tokens
    gr.Examples(
        examples=example_texts,
        inputs=text_input,
        label="Emotion Examples",
        cache_examples=False
    )
    
    with gr.Row():
        temp_slider = gr.Slider(minimum=0.1, maximum=2.0, step=0.1, value=0.9, label="Temperature")
        top_p_slider = gr.Slider(minimum=0.1, maximum=1.0, step=0.05, value=0.8, label="Top-p")
        tokens_slider = gr.Slider(minimum=100, maximum=3500, step=50, value=1200, label="Max New Tokens")
    
    output_audio = gr.Audio(type="numpy", label="Generated Audio")
    generate_button = gr.Button("Generate Audio")
    
    # Pass the selected model name along with other parameters
    generate_button.click(
        fn=generate_audio,
        inputs=[text_input, temp_slider, top_p_slider, tokens_slider, model_choice],
        outputs=output_audio
    )

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