app/_cuda_bootstrap.py

"""Make Windows able to find the CUDA / cuDNN DLLs that CTranslate2 needs.

The single most common failure when running faster-whisper / WhisperX on Windows is:

    RuntimeError: Library cudnn_ops64_9.dll is not found ...
    (or: Could not locate cudnn_ops_infer64_8.dll)

It happens because CTranslate2 loads cuDNN/cuBLAS by name via the OS loader, but the
DLLs live inside pip wheels (``nvidia-cudnn-cu12`` / ``nvidia-cublas-cu12``) or inside
``torch/lib`` (the cu128 wheel bundles them) — directories that are NOT on the default
DLL search path.

Importing this module **first** — before importing ``ctranslate2``, ``faster_whisper``
or ``whisperx`` — registers those directories with the OS loader using
``os.add_dll_directory`` (and prepends them to PATH as a belt-and-suspenders fallback).

Usage::

    from app import _cuda_bootstrap          # noqa: F401  (must come first)
    from faster_whisper import WhisperModel   # now finds cuDNN

This is a no-op on non-Windows platforms.
"""
from __future__ import annotations

import importlib.util
import os
import sys

# cuDNN must come last so its directory is searched first (add_dll_directory is LIFO-ish
# in practice via PATH ordering); the order here is the order we *append* candidates.
_NVIDIA_PKGS = (
    "nvidia.cuda_runtime",
    "nvidia.cuda_nvrtc",
    "nvidia.cublas",
    "nvidia.cudnn",
)


def _add_dll_dir(path: str) -> bool:
    if not path or not os.path.isdir(path):
        return False
    try:
        os.add_dll_directory(path)  # type: ignore[attr-defined]  (Windows only)
    except (OSError, AttributeError):
        return False
    # Some libraries resolve via LoadLibrary(name) which consults PATH, not the
    # add_dll_directory list — so prepend to PATH too.
    if path not in os.environ.get("PATH", ""):
        os.environ["PATH"] = path + os.pathsep + os.environ.get("PATH", "")
    return True


def _candidate_dirs() -> list[str]:
    dirs: list[str] = []
    for pkg in _NVIDIA_PKGS:
        try:
            spec = importlib.util.find_spec(pkg)
        except (ImportError, ValueError, ModuleNotFoundError):
            spec = None
        if spec and spec.submodule_search_locations:
            base = list(spec.submodule_search_locations)[0]
            # Windows wheels put DLLs in /bin, Linux in /lib — add both, harmless if absent.
            dirs.append(os.path.join(base, "bin"))
            dirs.append(os.path.join(base, "lib"))
    # torch's bundled CUDA libs (the cu128 wheel ships cudnn/cublas DLLs here).
    try:
        import torch  # noqa: WPS433 (local import on purpose)

        dirs.append(os.path.join(os.path.dirname(torch.__file__), "lib"))
    except Exception:  # noqa: BLE001 — torch may not be importable yet; ignore.
        pass
    return dirs


def setup_cuda_dll_path() -> list[str]:
    """Register CUDA/cuDNN DLL directories with the Windows loader. Returns dirs added."""
    if sys.platform != "win32":
        return []
    added: list[str] = []
    for d in _candidate_dirs():
        if _add_dll_dir(d):
            added.append(d)
    return added


# Run on import. Safe to import multiple times.
ADDED_DLL_DIRS = setup_cuda_dll_path()