from os import execl, getenv
from sys import argv, exception, executable
from textwrap import dedent
import traceback

import gradio as gr
from torch import cuda

from detikzify.webui import BANNER, build_ui, make_light

def is_official_demo():
    return getenv("SPACE_AUTHOR_NAME") == "nllg"

# Hack to temporarily work around memory leak, see:
#   * https://huggingface.co/spaces/nllg/DeTikZify/discussions/2
#   * https://github.com/gradio-app/gradio/issues/8503
def reload_on_oom_hook(func):
    def wrapper(*args, **kwargs):
        if isinstance(exception(), (MemoryError, cuda.OutOfMemoryError)):
            execl(executable, executable, *argv)
        return func(*args, **kwargs)
    return wrapper

if is_official_demo() and not cuda.is_available():
    center = ".gradio-container {text-align: center}"
    with gr.Blocks(css=center, theme=make_light(gr.themes.Soft()), title="DeTikZify") as demo:
        badge = "https://huggingface.co/datasets/huggingface/badges/resolve/main/duplicate-this-space-xl.svg"
        link = "https://huggingface.co/spaces/nllg/DeTikZify?duplicate=true"
        html = f'<a style="display:inline-block" href="{link}"> <img src="{badge}" alt="Duplicate this Space"> </a>'
        message = dedent("""\
        The resources required by our models surpass those provided by Hugging
        Face Spaces' free CPU tier. For full functionality, we suggest
        duplicating this space using a paid private GPU runtime.
        """)
        gr.HTML(f'{BANNER}\n<p>{message}</p>\n{html}')
else:
    use_big_models = cuda.is_available() and cuda.get_device_properties(0).total_memory > 15835660288
    model = f"detikzify-ds-{'7' if use_big_models else '1.3'}b"
    demo = build_ui(lock=is_official_demo(), model=model, light=True).queue()
    traceback.print_exc = reload_on_oom_hook(traceback.print_exc)

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)