Windows: not picking up GPU/CUDA install

[houstonhaynes]

Version Information
0.6.2

Describe the bug
CUDA "sniffing" is failing

See output log

2025-09-11 17:53:22 - witticism.core.dependency_validator - INFO - [DEPENDENCY_VALIDATOR] VALIDATION_START: checking all dependencies
2025-09-11 17:53:27 - witticism.core.dependency_validator - WARNING - [DEPENDENCY_VALIDATOR] PLATFORM_ERROR: Windows power management check failed - Command '['powershell', '-Command', 'Get-WmiObject', '-Class', 'Win32_PowerPlan']' timed out after 5 seconds
2025-09-11 17:53:27 - witticism.core.dependency_validator - INFO - [DEPENDENCY_VALIDATOR] CUDA_UNAVAILABLE: will use CPU mode
2025-09-11 17:53:27 - witticism.core.dependency_validator - INFO - [DEPENDENCY_VALIDATOR] VALIDATION_COMPLETE: checked 10 dependencies

[houstonhaynes]

You know that VBS is a virus, right? 😏

[houstonhaynes]

The issue is a version mismatch - WhisperX (through ctranslate2) expects cuDNN 8, but PyTorch 2.5.1 comes with cuDNN 9. It doesn't crash any more but I'm not getting audio to translate - having tried through two mics that I use regularly for video conferencing.

[houstonhaynes]

The VBS file had some missing """

Set objShell = CreateObject("WScript.Shell")
objShell.Run "powershell.exe -WindowStyle Hidden -ExecutionPolicy Bypass -File ""C:\Users\h3tec\AppData\Roaming\Microsoft\Windows\Start Menu\Programs\Startup\WitticismAutoStart.ps1""", 0, False

[houstonhaynes]

I seems to be working pretty hard for presumably hitting the GPU - when it was falling back to CPU mode it was "working" but slowly. Now it doesn't output any text at all.

[houstonhaynes]

 witticism --log-level DEBUG
2025-09-12 10:55:55 - witticism.main - INFO - [WITTICISM] STARTUP: version=0.6.2, args={'model': None, 'log_level': 'DEBUG', 'reset_config': False}
2025-09-12 10:55:55 - witticism.main - INFO - [WITTICISM] DEPENDENCY_CHECK: validating system and Python dependencies
2025-09-12 10:55:55 - witticism.core.dependency_validator - INFO - [DEPENDENCY_VALIDATOR] VALIDATION_START: checking all dependencies
2025-09-12 10:55:55 - witticism.core.dependency_validator - DEBUG - [DEPENDENCY_VALIDATOR] PYTHON_OK: torch v2.5.1+cu121
2025-09-12 10:55:55 - witticism.core.dependency_validator - DEBUG - [DEPENDENCY_VALIDATOR] PYTHON_OK: whisperx vunknown
2025-09-12 10:55:55 - witticism.core.dependency_validator - DEBUG - [DEPENDENCY_VALIDATOR] PYTHON_OK: PyQt5 vunknown
2025-09-12 10:55:55 - witticism.core.dependency_validator - DEBUG - [DEPENDENCY_VALIDATOR] PYTHON_OK: pyaudio v0.2.14
2025-09-12 10:55:55 - witticism.core.dependency_validator - DEBUG - [DEPENDENCY_VALIDATOR] PYTHON_OK: numpy v2.1.2
2025-09-12 10:55:55 - witticism.core.dependency_validator - DEBUG - [DEPENDENCY_VALIDATOR] PYTHON_OK: scipy v1.16.2
2025-09-12 10:55:56 - witticism.core.dependency_validator - DEBUG - [DEPENDENCY_VALIDATOR] PYTHON_OK: transformers v4.56.1
2025-09-12 10:55:56 - witticism.core.dependency_validator - DEBUG - [DEPENDENCY_VALIDATOR] SYSTEM_OK: Python 3.12
2025-09-12 10:56:01 - witticism.core.dependency_validator - WARNING - [DEPENDENCY_VALIDATOR] PLATFORM_ERROR: Windows power management check failed - Command '['powershell', '-Command', 'Get-WmiObject', '-Class', 'Win32_PowerPlan']' timed out after 5 seconds
2025-09-12 10:56:01 - witticism.core.dependency_validator - INFO - [DEPENDENCY_VALIDATOR] CUDA_OK: NVIDIA GeForce RTX 3050 with CUDA 12.1
2025-09-12 10:56:01 - witticism.core.dependency_validator - INFO - [DEPENDENCY_VALIDATOR] VALIDATION_COMPLETE: checked 10 dependencies
2025-09-12 10:56:01 - witticism.main - WARNING - [WITTICISM] DEPENDENCY_OPTIONAL_MISSING: 1 optional features will be disabled
2025-09-12 10:56:01 - witticism.main - INFO - [WITTICISM] FEATURE_DISABLED: windows_power_management - Command '['powershell', '-Command', 'Get-WmiObject', '-Class', 'Win32_PowerPlan']' timed out after 5 seconds
2025-09-12 10:56:01 - witticism.main - INFO - [WITTICISM] TRAY_EARLY_INIT: creating system tray for startup notifications
2025-09-12 10:56:01 - witticism.main - INFO - [WITTICISM] TRAY_EARLY_READY: system tray available for startup error notifications
2025-09-12 10:56:01 - witticism.main - INFO - [WITTICISM] ENGINE_INIT: model_size=base.en, device=auto, compute_type=auto, language=en
2025-09-12 10:56:01 - witticism.core.whisperx_engine - INFO - [WHISPERX_ENGINE] INIT: device=cuda, compute_type=auto, model_size=base.en, language=en
2025-09-12 10:56:01 - witticism.main - INFO - [WITTICISM] SLEEP_MONITORING: enabling proactive suspend/resume detection
2025-09-12 10:56:01 - witticism.platform.sleep_monitor - INFO - [SLEEP_MONITOR] FACTORY: creating Windows power event monitor
2025-09-12 10:56:01 - witticism.platform.sleep_monitor - DEBUG - [SLEEP_MONITOR] WINDOWS_MONITOR: starting PowerShell power event monitoring
2025-09-12 10:56:01 - witticism.platform.sleep_monitor - INFO - [SLEEP_MONITOR] STARTED: Windows power event monitoring active
2025-09-12 10:56:01 - witticism.core.whisperx_engine - INFO - Sleep monitoring enabled for proactive CUDA recovery
2025-09-12 10:56:01 - witticism.main - INFO - [WITTICISM] SLEEP_MONITORING_ACTIVE: CUDA protection enabled for initialization
2025-09-12 10:56:01 - witticism.main - INFO - [WITTICISM] CUDA_VALIDATION: performing startup health check with sleep monitor protection
2025-09-12 10:56:01 - witticism.core.whisperx_engine - INFO - [WHISPERX_ENGINE] CUDA_HEALTH_CHECK: Performing startup CUDA health validation
2025-09-12 10:56:01 - witticism.core.whisperx_engine - INFO - [WHISPERX_ENGINE] CUDA_HEALTHY: startup validation passed - gpu=NVIDIA GeForce RTX 3050, memory=0.00GB
2025-09-12 10:56:01 - witticism.main - INFO - [WITTICISM] CUDA_VALIDATION_PASSED: startup health check successful
2025-09-12 10:56:01 - witticism.main - INFO - [WITTICISM] MODEL_LOADING: starting WhisperX model loading
2025-09-12 10:56:01 - witticism.core.whisperx_engine - INFO - [WHISPERX_ENGINE] MODEL_LOADING: transcription model 'base.en' on cuda with auto
2025-09-12 10:56:04 - matplotlib - DEBUG - matplotlib data path: C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\matplotlib\mpl-data
2025-09-12 10:56:04 - matplotlib - DEBUG - CONFIGDIR=C:\Users\h3tec\.matplotlib
2025-09-12 10:56:04 - matplotlib - DEBUG - interactive is False
2025-09-12 10:56:04 - matplotlib - DEBUG - platform is win32
2025-09-12 10:56:04 - matplotlib - DEBUG - CACHEDIR=C:\Users\h3tec\.matplotlib
2025-09-12 10:56:04 - matplotlib.font_manager - DEBUG - Using fontManager instance from C:\Users\h3tec\.matplotlib\fontlist-v390.json
2025-09-12 10:56:05 - torio._extension.utils - DEBUG - Loading FFmpeg6
2025-09-12 10:56:05 - torio._extension.utils - DEBUG - Failed to load FFmpeg6 extension.
Traceback (most recent call last):
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 116, in _find_ffmpeg_extension
    ext = _find_versionsed_ffmpeg_extension(ffmpeg_ver)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 108, in _find_versionsed_ffmpeg_extension
    _load_lib(lib)
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 94, in _load_lib
    torch.ops.load_library(path)
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torch\_ops.py", line 1350, in load_library
    ctypes.CDLL(path)
  File "C:\Users\h3tec\AppData\Local\Programs\Python\Python312\Lib\ctypes\__init__.py", line 379, in __init__
    self._handle = _dlopen(self._name, mode)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: Could not find module 'C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\lib\libtorio_ffmpeg6.pyd' (or one of its dependencies). Try using the full path with constructor syntax.
2025-09-12 10:56:05 - torio._extension.utils - DEBUG - Loading FFmpeg5
2025-09-12 10:56:05 - torio._extension.utils - DEBUG - Failed to load FFmpeg5 extension.
Traceback (most recent call last):
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 116, in _find_ffmpeg_extension
    ext = _find_versionsed_ffmpeg_extension(ffmpeg_ver)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 108, in _find_versionsed_ffmpeg_extension
    _load_lib(lib)
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 94, in _load_lib
    torch.ops.load_library(path)
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torch\_ops.py", line 1350, in load_library
    ctypes.CDLL(path)
  File "C:\Users\h3tec\AppData\Local\Programs\Python\Python312\Lib\ctypes\__init__.py", line 379, in __init__
    self._handle = _dlopen(self._name, mode)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: Could not find module 'C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\lib\libtorio_ffmpeg5.pyd' (or one of its dependencies). Try using the full path with constructor syntax.
2025-09-12 10:56:05 - torio._extension.utils - DEBUG - Loading FFmpeg4
2025-09-12 10:56:05 - torio._extension.utils - DEBUG - Failed to load FFmpeg4 extension.
Traceback (most recent call last):
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 116, in _find_ffmpeg_extension
    ext = _find_versionsed_ffmpeg_extension(ffmpeg_ver)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 108, in _find_versionsed_ffmpeg_extension
    _load_lib(lib)
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 94, in _load_lib
    torch.ops.load_library(path)
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torch\_ops.py", line 1350, in load_library
    ctypes.CDLL(path)
  File "C:\Users\h3tec\AppData\Local\Programs\Python\Python312\Lib\ctypes\__init__.py", line 379, in __init__
    self._handle = _dlopen(self._name, mode)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: Could not find module 'C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\lib\libtorio_ffmpeg4.pyd' (or one of its dependencies). Try using the full path with constructor syntax.
2025-09-12 10:56:05 - torio._extension.utils - DEBUG - Loading FFmpeg
2025-09-12 10:56:05 - torio._extension.utils - DEBUG - Failed to load FFmpeg extension.
Traceback (most recent call last):
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 116, in _find_ffmpeg_extension
    ext = _find_versionsed_ffmpeg_extension(ffmpeg_ver)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\torio\_extension\utils.py", line 106, in _find_versionsed_ffmpeg_extension
    raise RuntimeError(f"FFmpeg{version} extension is not available.")
RuntimeError: FFmpeg extension is not available.
2025-09-12 10:56:05 - speechbrain.utils.checkpoints - DEBUG - Registered checkpoint save hook for _speechbrain_save
2025-09-12 10:56:05 - speechbrain.utils.checkpoints - DEBUG - Registered checkpoint load hook for _speechbrain_load
2025-09-12 10:56:05 - speechbrain.utils.checkpoints - DEBUG - Registered checkpoint save hook for save
2025-09-12 10:56:05 - speechbrain.utils.checkpoints - DEBUG - Registered checkpoint load hook for load
2025-09-12 10:56:06 - speechbrain.utils.quirks - INFO - Applied quirks (see `speechbrain.utils.quirks`): [disable_jit_profiling, allow_tf32]
2025-09-12 10:56:06 - speechbrain.utils.quirks - INFO - Excluded quirks specified by the `SB_DISABLE_QUIRKS` environment (comma-separated list): []
2025-09-12 10:56:06 - speechbrain.utils.checkpoints - DEBUG - Registered checkpoint save hook for _save
2025-09-12 10:56:06 - speechbrain.utils.checkpoints - DEBUG - Registered checkpoint load hook for _recover
>>Performing voice activity detection using Pyannote...
2025-09-12 10:56:11 - fsspec.local - DEBUG - open file: C:/Users/h3tec/pipx/venvs/witticism/Lib/site-packages/whisperx/assets/pytorch_model.bin
2025-09-12 10:56:11 - fsspec.local - DEBUG - open file: C:/Users/h3tec/pipx/venvs/witticism/Lib/site-packages/whisperx/assets/pytorch_model.bin
2025-09-12 10:56:11 - pytorch_lightning.utilities.migration.utils - INFO - Lightning automatically upgraded your loaded checkpoint from v1.5.4 to v2.5.5. To apply the upgrade to your files permanently, run `python -m pytorch_lightning.utilities.upgrade_checkpoint C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\whisperx\assets\pytorch_model.bin`
Model was trained with pyannote.audio 0.0.1, yours is 3.4.0. Bad things might happen unless you revert pyannote.audio to 0.x.
Model was trained with torch 1.10.0+cu102, yours is 2.5.1+cu121. Bad things might happen unless you revert torch to 1.x.
2025-09-12 10:56:11 - witticism.core.whisperx_engine - INFO - [WHISPERX_ENGINE] MODEL_LOADED: transcription model loaded in 10.27s
2025-09-12 10:56:11 - witticism.core.whisperx_engine - INFO - [WHISPERX_ENGINE] ALIGN_MODEL_LOADING: alignment model for language 'en' on cuda
2025-09-12 10:56:13 - witticism.core.whisperx_engine - INFO - [WHISPERX_ENGINE] ALIGN_MODEL_LOADED: alignment model loaded in 1.12s
2025-09-12 10:56:13 - witticism.core.whisperx_engine - INFO - [WHISPERX_ENGINE] ALL_MODELS_LOADED: all models loaded successfully
2025-09-12 10:56:13 - witticism.main - INFO - [WITTICISM] MODEL_LOADING_COMPLETE: all models loaded successfully
2025-09-12 10:56:13 - witticism.main - INFO - [WITTICISM] AUDIO_INIT: sample_rate=48000, channels=1, vad_aggressiveness=2
2025-09-12 10:56:13 - witticism.core.audio_capture - INFO - AudioCapture initialized: 48000Hz, 1 channel(s)
2025-09-12 10:56:13 - witticism.main - INFO - [WITTICISM] PIPELINE_INIT: min_audio_length=0.5s, max_audio_length=30.0s
2025-09-12 10:56:13 - witticism.core.transcription_pipeline - INFO - TranscriptionPipeline initialized
2025-09-12 10:56:13 - witticism.core.transcription_pipeline - INFO - TranscriptionPipeline started
2025-09-12 10:56:13 - witticism.main - INFO - [WITTICISM] OUTPUT_INIT: mode=type
2025-09-12 10:56:13 - witticism.utils.output_manager - INFO - OutputManager initialized with mode: type
2025-09-12 10:56:13 - witticism.core.hotkey_manager - INFO - [HOTKEY_MANAGER] PTT_KEY_CHANGED: from f9 to f9
2025-09-12 10:56:13 - witticism.core.hotkey_manager - INFO - [HOTKEY_MANAGER] CONFIG_PTT_KEY: using configured PTT key 'F9'
2025-09-12 10:56:13 - witticism.core.hotkey_manager - INFO - [HOTKEY_MANAGER] INIT: mode=push_to_talk, ptt_key=f9
2025-09-12 10:56:13 - witticism.main - INFO - [WITTICISM] INIT_COMPLETE: all components initialized successfully
2025-09-12 10:56:13 - witticism.core.hotkey_manager - INFO - [HOTKEY_MANAGER] STARTED: keyboard listener active
2025-09-12 10:56:13 - witticism.main - INFO - [WITTICISM] STARTUP_COMPLETE: application ready and running
2025-09-12 10:56:19 - witticism.core.hotkey_manager - DEBUG - [HOTKEY_MANAGER] PTT_START: f9 pressed - recording started
2025-09-12 10:56:19 - witticism.core.audio_capture - INFO - Started recording on device None
2025-09-12 10:56:19 - witticism.core.audio_capture - INFO - Push-to-talk started
2025-09-12 10:56:21 - witticism.core.hotkey_manager - DEBUG - [HOTKEY_MANAGER] PTT_STOP: f9 released - recording stopped
2025-09-12 10:56:21 - witticism.core.audio_capture - INFO - Stopped recording. Captured 1.83 seconds
2025-09-12 10:56:21 - witticism.core.audio_capture - INFO - Push-to-talk stopped. Duration: 1.83s
C:\Users\h3tec\pipx\venvs\witticism\Lib\site-packages\pyannote\audio\utils\reproducibility.py:74: ReproducibilityWarning: TensorFloat-32 (TF32) has been disabled as it might lead to reproducibility issues and lower accuracy.
It can be re-enabled by calling
   >>> import torch
   >>> torch.backends.cuda.matmul.allow_tf32 = True
   >>> torch.backends.cudnn.allow_tf32 = True
See https://github.com/pyannote/pyannote-audio/issues/1370 for more details.

  warnings.warn(
No active speech found in audio

[Aaronontheweb]

sorry I missed this @houstonhaynes - was OOF for the past few days. Yeah, we definitely need to fix the CUDA detection for Windows. I don't have good hardware to test it on (all four of my CUDA machines run Linux)

[Aaronontheweb]

As for the .vbs issue - that must be an artifact in the installer script. Need to get rid of it probably.

I started having some degradation on one of my CUDA Linux machines too so there is likely a bug in here that needs fixing.

[houstonhaynes]

I actually went and built my own F# library just to play with Vosk and kept it at Whisper.cpp streaming. The only thing that irks me more than .NET interop is the Python heterogyny. 😺😆 https://github.com/speakeztech/WhisperFS