deepfakes/faceswap · error · FaceswapError

There was an error reading from the Nvidia Machine Learning

Error message

There was an error reading from the Nvidia Machine Learning Library. The most likely cause is incorrectly installed drivers. If this is the case, Please remove and reinstall your Nvidia drivers before reporting. Original Error: {str(err)}

What it means

Fallback except in the same PyNVML init path: any exception other than the three known NVML errors (e.g. NVML_UnknownError, version-glyph errors like NVML_ERROR_FUNCTION_NOT_FOUND after nvmlInit succeeded partially, or OS-level errors) is re-raised as FaceswapError with the original message preserved.

Source

Thrown at lib/gpu_stats/nvidia.py:55

        Raises
        ------
        FaceswapError
            If the NVML library could not be successfully loaded
        """
        if self._is_initialized:
            return
        try:
            self._log("debug", "Initializing PyNVML for Nvidia GPU.")
            pynvml.nvmlInit()
        except (pynvml.NVMLError_LibraryNotFound,  # pylint:disable=no-member
                pynvml.NVMLError_DriverNotLoaded,  # pylint:disable=no-member
                pynvml.NVMLError_NoPermission) as err:  # pylint:disable=no-member
            msg = ("There was an error reading from the Nvidia Machine Learning Library. The most "
                   "likely cause is incorrectly installed drivers. If this is the case, Please "
                   "remove and reinstall your Nvidia drivers before reporting. Original "
                   f"Error: {str(err)}")
            raise FaceswapError(msg) from err
        except Exception as err:  # pylint:disable=broad-except
            msg = ("An unhandled exception occurred reading from the Nvidia Machine Learning "
                   f"Library. Original error: {str(err)}")
            raise FaceswapError(msg) from err

        os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
        super()._initialize()

    def _shutdown(self) -> None:
        """Cleanly close access to NVML and set :attr:`_is_initialized` back to ``False``. """
        self._log("debug", "Shutting down NVML")
        pynvml.nvmlShutdown()
        super()._shutdown()

    def _get_device_count(self) -> int:
        """Detect the number of GPUs attached to the system.

        Returns

View on GitHub (pinned to f530cb7508)

Solutions

1. Read the appended 'Original error' to identify the NVML code (e.g. NVML_ERROR_FUNCTION_NOT_FOUND) and match driver vs pynvml versions
2. Clean-reinstall the NVIDIA driver AND update the nvidia-ml-py/pynvml package so both come from the same driver generation
3. Verify `nvidia-smi` works; if not, reload the kernel modules (sudo rmmod nvidia_uvm && sudo modprobe nvidia_uvm) or reboot
4. Bypass with --cpu while the GPU stack is repaired

Example fix

# before
python faceswap.py --nvidia ...  # An unhandled exception occurred reading from the Nvidia ML Library

# after
pip install -U nvidia-ml-py          # match pynvml to driver
sudo apt install --reinstall nvidia-driver-XXX
nvidia-smi                           # verify
python faceswap.py --nvidia ...

import shutil, subprocess

def nvidia_driver_ok():
    return shutil.which("nvidia-smi") is not None \
        and subprocess.run(["nvidia-smi"], capture_output=True).returncode == 0

assert nvidia_driver_ok(), "fix NVIDIA driver before --nvidia"

def nvml_reachable() -> bool:
    """True when NVML initializes cleanly (driver loaded & permitted)."""
    try:
        import pynvml
        pynvml.nvmlInit()
        pynvml.nvmlShutdown()
        return True
    except Exception:
        return False

from lib.utils import FaceswapError
try:
    gpu_stats = NvidiaStats(log="debug")
except FaceswapError as err:
    if "Nvidia Machine Learning Library" in str(err):
        log_driver_fix_instructions(); use_cpu()
    else:
        raise

When it happens

Trigger: `faceswap.py --nvidia ...` where nvmlInit() raises something unexpected: driver/library version skew (newer driver userspace, older libnvidia-ml), NVML functions missing in the installed driver version, driver lost GPU access mid-boot, or exotic virtualization (vGPU/legacy passthrough) errors.

Common situations: Mixed-version installs (libnvidia-ml from one release, kernel module from another); VMs with GPU passthrough; driver crashes that leave NVML in a bad state; pynvml package version newer than the driver's NVML API.

Related errors

• An unhandled exception occurred initializing the device via
• Hex color codes should start with a '#' and be 6 characters
• You do not have enough GPU memory available to train the sel
• Landmark based masks cannot be created for {self._landmark_t
• There is a mismatch between the number of frames found in th

AI-assisted analysis of deepfakes/faceswap@f530cb7508 (2026-08-15). Data as JSON: /api/errors/133145adf1f96299. Report an issue: GitHub.