VoiceStudio/backend/core/cudnn8.py

"""cuDNN 8 side-load, and — the point of this module — a *probe* for it.

CTranslate2 (pinned at 4.4.0; the engine under WhisperX and faster-whisper)
links against **cuDNN 8**, while PyTorch 2.8+ ships cuDNN 9. The two coexist:
the Rust bootstrap side-loads ``nvidia-cudnn-cu12==8.9.7.29`` into a
``cudnn8_compat/`` directory beside the venv's real site-packages (#827/#869),
and we ``ctypes``-preload those libraries at startup so CTranslate2's own
``LoadLibrary``/``dlopen`` finds them already resident in the process.

When that side-load is missing, CTranslate2 does not raise. It prints

    Could not locate cudnn_ops_infer64_8.dll. Please make sure it is in your
    library path!

and calls ``__fastfail`` — killing the **whole backend process** with
``0xC0000409`` (STATUS_STACK_BUFFER_OVERRUN, surfaced as exit code
``-1073740791``). Python never gets a frame, so there is no traceback, no
fallback, and nothing for the crash notice to classify. The desktop shell
restarts the backend, the user retries, and it dies again — #1371.

The bug was never that the preload could fail. It is that we **computed the
answer and threw it away**: the old inline preload in ``main.py`` silently
``pass``-ed on a missing directory and on every ``OSError``, then handed
control to a native library that treats the same condition as fatal. So this
module keeps the outcome and exposes :func:`ctranslate2_cudnn_status`, letting
``asr_backend`` route around a doomed engine *before* calling into it —
exactly how the CTranslate2 exec-stack failure (#692) is already handled.

Deliberately stdlib-only: ``engines/_asr_sidecar/main.py`` runs in a child
process with a clean import path and must not drag in the heavy ``services``
package to get this.
"""
from __future__ import annotations

import glob
import logging
import os
import sys

logger = logging.getLogger(__name__)

# The library CTranslate2 names in its own failure message. Probing the exact
# one the error cites keeps the diagnosis honest — and because the probe runs
# in the same process, through the same OS loader, a failure here is the same
# failure CTranslate2 is about to hit.
_SENTINEL_LIB = (
    "cudnn_ops_infer64_8.dll" if sys.platform == "win32" else "libcudnn_ops_infer.so.8"
)
_LIB_GLOB = "cudnn*64_8.dll" if sys.platform == "win32" else "libcudnn*.so.8"

_preloaded: bool = False
_status: tuple[bool, str] | None = None
# Handles from os.add_dll_directory (Windows). Held for the process lifetime —
# dropping one un-registers its directory. See preload().
_dll_dir_cookies: list = []


def _project_root() -> str:
    # backend/core/cudnn8.py → backend/core → backend → <project root>
    return os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))


def compat_dirs() -> list[str]:
    """Every plausible ``cudnn8_compat`` location, most specific first.

    The old inline preload hardcoded ``<project root>/.venv``. That is right
    for the managed desktop install and wrong everywhere else — a differently
    named venv, ``uv run`` from a checkout, Docker, or a system Python all
    resolve to a directory that does not exist, so the preload silently did
    nothing and the process died later anyway. ``sys.prefix`` is where the
    *running* interpreter actually lives, so it is checked too.
    """
    pyver = f"python{sys.version_info.major}.{sys.version_info.minor}"
    if sys.platform == "win32":
        tail = (("Lib", "site-packages"), ("nvidia", "cudnn", "bin"))
    else:
        tail = (("lib", pyver, "site-packages"), ("nvidia", "cudnn", "lib"))
    site_tail, lib_tail = tail

    roots = [os.path.join(_project_root(), ".venv"), sys.prefix]
    out: list[str] = []
    for root in roots:
        path = os.path.join(root, *site_tail, "cudnn8_compat", *lib_tail)
        if path not in out:
            out.append(path)
    return out


def preload() -> None:
    """Load the side-loaded cuDNN 8 libraries into this process. Idempotent.

    Best-effort by design — a host with no side-load is not an error here, it
    is a fact :func:`ctranslate2_cudnn_status` reports later.
    """
    global _preloaded
    if _preloaded:
        return
    _preloaded = True
    if sys.platform == "darwin":  # no CUDA on macOS
        return

    import ctypes

    mode = 0 if sys.platform == "win32" else ctypes.RTLD_GLOBAL
    for lib_dir in compat_dirs():
        if not os.path.isdir(lib_dir):
            continue
        if sys.platform == "win32":
            # Make the directory searchable by *name* as well, so a CTranslate2
            # LoadLibrary for a dependent library we did not preload explicitly
            # still resolves instead of aborting the process.
            #
            # The returned cookie must be KEPT: it is a context manager whose
            # close/__del__ removes the directory again, so discarding it makes
            # the call a silent no-op — the exact failure this module exists to
            # prevent, reintroduced one line lower (CodeRabbit, #1401). Hold it
            # for the life of the process.
            try:
                _dll_dir_cookies.append(os.add_dll_directory(lib_dir))
            except (OSError, AttributeError) as e:
                logger.debug("cuDNN 8 DLL directory not added (%s): %s", lib_dir, e)
        for so in sorted(glob.glob(os.path.join(lib_dir, _LIB_GLOB))):
            try:
                ctypes.CDLL(so, mode=mode)
            except OSError as e:
                logger.debug("cuDNN 8 preload skipped %s: %s", so, e)
        return  # first directory that exists wins


def _torch_wants_cudnn8() -> tuple[bool, str]:
    """Whether CTranslate2 will reach for cuDNN 8 *on this host at all*.

    cuDNN is a CUDA library: with no CUDA device, CTranslate2 runs on the CPU
    and never touches it, so a missing side-load is harmless and must not
    disqualify the engine. ROCm is excluded for the same reason and one more —
    ``torch.cuda.is_available()`` is True on a HIP build, but CTranslate2 has
    no ROCm backend, so it is CPU-only there regardless. Without this branch a
    Windows/CUDA fix would silently downgrade every ROCm user's ASR engine.
    Mirrors ``bootstrap.rs::classify_cuda_probe``, which likewise declines to
    install cuDNN 8 on HIP hosts (#124).
    """
    try:
        import torch
    except Exception as e:  # noqa: BLE001 — no torch: nothing will run anyway
        return False, f"torch unavailable ({type(e).__name__})"
    if getattr(getattr(torch, "version", None), "hip", None):
        return False, "ROCm build — CTranslate2 has no ROCm backend, runs on CPU"
    try:
        if not torch.cuda.is_available():
            return False, "no CUDA device — CTranslate2 runs on CPU"
    except Exception as e:  # noqa: BLE001
        return False, f"CUDA probe failed ({type(e).__name__})"
    return True, "CUDA device present"


def ctranslate2_cudnn_status() -> tuple[bool, str]:
    """``(usable, reason)`` — can CTranslate2 load cuDNN 8 in this process?

    Cached: the answer cannot change within a process, and it is consulted on
    every ASR engine selection.

    Conservative on purpose. A false negative costs a user WhisperX's forced
    alignment (and so lip-sync accuracy), so this returns True whenever cuDNN 8
    is not actually required, and only reports False for the one condition it
    can prove: the library the failure message names will not load, here, now,
    through the same loader CTranslate2 is about to use.
    """
    global _status
    if _status is not None:
        return _status
    _status = _compute_status()
    if not _status[0]:
        logger.warning("CTranslate2 ASR engines unavailable: %s", _status[1])
    return _status


def _try_load(name: str) -> None:
    """Load a shared library by *bare name*, raising OSError if it will not.

    Deliberately its own seam: this runs in the same process and through the
    same OS loader CTranslate2 is about to use, which is what makes the probe
    faithful — and it gives tests a way to simulate a missing library without
    replacing ``ctypes.CDLL`` process-wide (which would break torch's own load).
    """
    import ctypes

    ctypes.CDLL(name)


def _compute_status() -> tuple[bool, str]:
    # No separate macOS branch: there is no CUDA there, so `_torch_wants_cudnn8`
    # already answers "not required". One gate, testable on any platform.
    needed, why = _torch_wants_cudnn8()
    if not needed:
        return True, f"cuDNN 8 not required ({why})"

    preload()
    try:
        _try_load(_SENTINEL_LIB)
    except OSError as e:
        return False, (
            f"CUDA is active but {_SENTINEL_LIB} cannot be loaded ({e}). "
            f"WhisperX and faster-whisper are CTranslate2, which requires "
            f"cuDNN 8; loading it is not optional and its absence aborts the "
            f"backend process outright rather than raising (#1371). Reinstall "
            f"the compat libraries with "
            f"`uv pip install --target <venv>/{'Lib/site-packages' if sys.platform == 'win32' else 'lib/pythonX.Y/site-packages'}/cudnn8_compat "
            f"nvidia-cudnn-cu12==8.9.7.29`, or pin a non-CTranslate2 engine "
            f"with OMNIVOICE_ASR_BACKEND=pytorch-whisper."
        )
    return True, "cuDNN 8 loadable"


def reset_cache_for_tests() -> None:
    """Clear the memoised probe. Tests only."""
    global _status, _preloaded
    _status = None
    _preloaded = False