"""Hugging Face Space entry point (Gradio SDK on ZeroGPU hardware).

ZeroGPU rules followed here: ``spaces`` is imported before torch, the NLI cross-encoder is
placed on CUDA at module level (ZeroGPU emulates CUDA outside GPU functions), and only the
NLI scoring call runs inside ``@spaces.GPU``. Off ZeroGPU the decorator does nothing and
everything runs on CPU, so the same file works locally:

    uv run --extra demo --extra nli python space/app.py
"""

from __future__ import annotations

import os
import sys
from pathlib import Path

import spaces  # must be imported before torch on ZeroGPU

HERE = Path(__file__).resolve().parent
# In the Space, src/ sits next to this file; in the repository it is one level up.
ROOT = HERE if (HERE / "src").is_dir() else HERE.parent
sys.path.insert(0, str(ROOT / "src"))
os.chdir(ROOT)  # configs/default.yaml uses paths relative to the project root

import torch
from sentence_transformers import CrossEncoder

from graphtrace.app.gradio_app import CSS, DemoData, build_demo, make_theme
from graphtrace.config import load_project
from graphtrace.verification import NLIVerifier

cfg = load_project("configs/default.yaml")
MODEL = cfg.extraction.nli_model
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

nli = NLIVerifier(MODEL, encoder=CrossEncoder(MODEL, device=DEVICE))
_cpu_nli: NLIVerifier | None = nli if DEVICE == "cpu" else None


@spaces.GPU(duration=20)
def gpu_verify(claim: str, texts: list[str]):
    return nli.verify(claim, texts)


def cpu_verify(claim: str, texts: list[str]):
    """Fallback when the GPU is unavailable (quota reached, allocation failed)."""
    global _cpu_nli
    if _cpu_nli is None:
        _cpu_nli = NLIVerifier(MODEL, encoder=CrossEncoder(MODEL, device="cpu"))
    return _cpu_nli.verify(claim, texts)


demo = build_demo(DemoData.load(cfg), nli, verify=gpu_verify, fallback=None if DEVICE == "cpu" else cpu_verify)
demo.queue(default_concurrency_limit=4, max_size=32)

if __name__ == "__main__":
    demo.launch(theme=make_theme(), css=CSS, server_port=int(os.environ.get("PORT", "7860")))