# %pip install gradio diffusers

import os
from huggingface_hub import login
login(token=os.getenv("HUGGINGFACEHUB_API_KEY"))

import gradio as gr
import numpy as np
import random

# import spaces #[uncomment to use ZeroGPU]
from diffusers import DiffusionPipeline
import torch

def get_device_type(idx):
    if torch.cuda.is_available():
        return f"cuda:{idx}" if torch.cuda.device_count() >= idx else "cuda", torch.float16
    elif torch.backends.mps.is_available():
        return "mps", torch.float16
    else:
        return "cpu", torch.float32

device0, torch_dtype = get_device_type(0)
device1, torch_dtype = get_device_type(1)

text_generation_model_name="meta-llama/Llama-3.2-3B-Instruct" 

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
tokenizer = AutoTokenizer.from_pretrained(text_generation_model_name)
model = AutoModelForCausalLM.from_pretrained(text_generation_model_name).to(device0, dtype=torch.float16)

text_to_speech_model_name = "suno/bark"
# Only PyTorch models and some Diffusers pipelines have .to().
text_to_speech_pipeline = pipeline("text-to-speech", model=text_to_speech_model_name, device=device0)

text_to_image_model_name = "stabilityai/sdxl-turbo"
text_to_image_pipeline = DiffusionPipeline.from_pretrained(text_to_image_model_name, torch_dtype=torch_dtype).to(device1)

# from diffusers import StableDiffusionPipeline
# text_to_image_model_name = "sd-legacy/stable-diffusion-v1-5"
# text_to_image_pipeline = StableDiffusionPipeline.from_pretrained(text_to_image_model_name, torch_dtype=torch_dtype).to(device1)

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024

system_prompt = (
    "You are a helpful AI assistant that generates a story for kids based on the input provided. "
    "The story should be engaging and creative. "
    "Here is the input: {input} "
    "Please respond with the story."
)

history = []

def generate_text(message):
    global history
    sys_prompt = system_prompt.replace("{input}", message)
    if not history or history[0].get("role") != "system":
        history = [{"role": "system", "content": sys_prompt}] + history
    else:
        history[0]["content"] = sys_prompt

    history.append({"role": "user", "content": message})

    # 1. Build prompt from history using chat template
    prompt = tokenizer.apply_chat_template(history, tokenize=False)
    inputs = tokenizer(prompt, return_tensors="pt").to(device0)
    outputs = model.generate(**inputs, max_new_tokens=128)
    decoded = tokenizer.decode(outputs[0], skip_special_tokens=False)

    if "<|start_header_id|>assistant<|end_header_id|>" in decoded:
        analysis_response = decoded.split("<|start_header_id|>assistant<|end_header_id|>")[-1]
        analysis_response = analysis_response.replace("<|eot_id|>", "").strip()
    elif "<|im_start|>assistant" in decoded:
        # This works for most chat templates that append the assistant's reply at the end
        analysis_response = decoded.split("<|im_start|>assistant")[-1]
        analysis_response = analysis_response.replace("<|im_end|>", "").strip()
    else:
        # Fallback: just return the decoded output
        analysis_response = decoded.strip()

    return analysis_response

def generate_audio(text):
    tts_result = text_to_speech_pipeline(text)

    # example: [[ 0.00073422 0.00038968 0.00035801 ... -0.01280548 -0.0147996 -0.01798675]]
    audio = tts_result["audio"]  # This is a numpy array (wav)

    audio_array = np.array(audio, dtype=np.float32).flatten()

    sample_rate = 22050  # or your actual sample rate

    # gr.Gradio expects tuple[int, np.ndarray]
    return (sample_rate, audio_array)


# @spaces.GPU #[uncomment to use ZeroGPU]
def generate_image(
    prompt,
    negative_prompt,
    guidance_scale,
    num_inference_steps,
    width,
    height,
    seed,
    randomize_seed,
    progress=gr.Progress(track_tqdm=True),
):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    generator = torch.Generator().manual_seed(seed)

    image = text_to_image_pipeline(
        prompt=prompt,
        negative_prompt=negative_prompt,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        width=width,
        height=height,
        generator=generator,
    ).images[0]

    return image, seed

def generate_all(
    prompt,
    negative_prompt,
    guidance_scale,
    num_inference_steps,
    width,
    height,
    seed,
    randomize_seed,
    progress=gr.Progress(track_tqdm=True),
):
    # Generate text from the prompt
    story = generate_text(prompt)

    # Generate audio from the text
    audio = generate_audio(story)
    
    # Generate image from the text
    image, seed = generate_image(
        story,
        negative_prompt,
        guidance_scale,
        num_inference_steps,
        width,
        height,
        seed,
        randomize_seed,
        progress=progress,
    )
    
    return story, audio, image, seed


examples = [
    "sky",
    "sea",
]

css = """
#col-container {
    margin: 0 auto;
    max-width: 640px;
}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown("# Story Generator (text & audio on cuda0, image on cuda1)")

        with gr.Row():
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt",
                container=False,
            )

            run_button = gr.Button("Run", scale=0, variant="primary")

        story = gr.Text(label="Story", show_label=False)

        audio = gr.Audio(label="Audio", show_label=False)

        image = gr.Image(label="Image", show_label=False)

        with gr.Accordion("Advanced Settings", open=False):
            negative_prompt = gr.Text(
                label="Negative prompt",
                max_lines=1,
                placeholder="Enter a negative prompt",
                visible=False,
            )

            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )

            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)

            with gr.Row():
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,  # Replace with defaults that work for your model
                )

                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,  # Replace with defaults that work for your model
                )

            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance scale",
                    minimum=0.0,
                    maximum=10.0,
                    step=0.1,
                    value=0.0,  # Replace with defaults that work for your model
                )

                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=2,  # Replace with defaults that work for your model
                )

        gr.Examples(examples=examples, inputs=[prompt])
    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=generate_all,
        inputs=[
            prompt,
            negative_prompt,
            guidance_scale,
            num_inference_steps,
            width,
            height,
            seed,
            randomize_seed,
        ],
        outputs=[story, audio, image, seed],
    )

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