demo_runtime/weights.py

from __future__ import annotations

import os
from pathlib import Path
from typing import Dict, Iterable, List


APP_TMP = Path("/tmp/bila-space-demo")


def _writable_data_dir() -> Path:
    data = Path("/data")
    if data.exists() and os.access(data, os.W_OK):
        return data / "bila-space-demo"
    return APP_TMP


def configure_runtime_cache() -> Path:
    base = _writable_data_dir()
    hf_home = Path(os.environ.get("HF_HOME", str(base / "hf-home")))
    torch_home = Path(os.environ.get("TORCH_HOME", str(base / "torch-home")))
    gradio_tmp = Path(os.environ.get("GRADIO_TEMP_DIR", str(base / "gradio-tmp")))

    os.environ.setdefault("HF_HOME", str(hf_home))
    os.environ.setdefault("TORCH_HOME", str(torch_home))
    os.environ.setdefault("GRADIO_TEMP_DIR", str(gradio_tmp))

    for path in (hf_home, torch_home, gradio_tmp):
        path.mkdir(parents=True, exist_ok=True)
    return base


def _allow_patterns_for_model(model_cfg: Dict) -> List[str]:
    patterns = []
    for rel_path in model_cfg["weights"].values():
        if rel_path.endswith((".pth", ".bin", ".safetensors", ".json")):
            patterns.append(rel_path)
        else:
            patterns.append(rel_path.rstrip("/") + "/**")
    metric_file = model_cfg.get("evidence", {}).get("metric_file")
    if metric_file:
        patterns.append(metric_file)
    return patterns


def resolve_model_root(model_key: str, model_cfg: Dict) -> Path:
    local_root = os.environ.get("BILA_MODEL_ROOT")
    if local_root:
        return Path(local_root).expanduser().resolve()

    repo_id = os.environ.get("BILA_WEIGHTS_REPO")
    if not repo_id:
        raise RuntimeError(
            "Set BILA_WEIGHTS_REPO to the Hugging Face model repo containing demo weights, "
            "or set BILA_MODEL_ROOT to a local directory with the same layout."
        )

    from huggingface_hub import snapshot_download

    cache_dir = Path(os.environ.get("BILA_MODEL_CACHE", str(_writable_data_dir() / "hf-cache")))
    cache_dir.mkdir(parents=True, exist_ok=True)
    return Path(
        snapshot_download(
            repo_id=repo_id,
            repo_type=os.environ.get("BILA_WEIGHTS_REPO_TYPE", "model"),
            cache_dir=str(cache_dir),
            allow_patterns=_allow_patterns_for_model(model_cfg),
            token=os.environ.get("HF_TOKEN"),
        )
    )


def require_paths(root: Path, rel_paths: Iterable[str]) -> Dict[str, Path]:
    resolved = {}
    missing = []
    for rel_path in rel_paths:
        path = root / rel_path
        resolved[rel_path] = path
        if not path.exists():
            missing.append(str(path))
    if missing:
        raise FileNotFoundError("Missing required weight paths:\n" + "\n".join(missing))
    return resolved