ersistent OSError: [WinError 127] Error loading c10_cuda.dll during torch import on Windows (Pinokio path)

[arshanarchi]

Hello,

I am trying to set up and run this project from a Pinokio installation path (C:\pinokio\api\echomimic2.git\app), but I am encountering a persistent OSError when the application tries to import the torch library. The specific error is:

OSError: [WinError 127] The specified procedure could not be found. Error loading "C:\pinokio\api\echomimic2.git\app\env\lib\site-packages\torch\lib\c10_cuda.dll" or one of its dependencies.
This error occurs during the import torch step when running python app_en.py (or similar entry point scripts).

Here is information about my system and environment:

Operating System: Windows 10 (Version 10.0.22631.5262)
GPU: NVIDIA GeForce RTX 4060
NVIDIA Driver Version: 576.02
CUDA Version (supported by driver, from nvidia-smi): 12.9
Python Version (in Conda env): 3.10
Conda Environment Name: echomimic (created with python=3.10)
Project Path: C:\pinokio\api\echomimic2.git\app
I have already performed extensive troubleshooting, but the error persists. Here are the steps I have taken:

Verified the env directory exists: Confirmed C:\pinokio\api\echomimic2.git\app\env exists, but it was not recognized as a Conda environment initially.
Created a new Conda environment: Created a new environment (echomimic) as the previous env directory was not recognized.
Attempted installation with Python 3.8, then 3.9, then 3.10: Initial attempts failed on dependency conflicts (openvcv-python, gradiod_client). Settled on Python 3.10 as recommended by the main antgroup/echomimic_v2 README.
Corrected gradiod_client typo: Found and corrected a typo in requirements.txt from gradiod_client==1.4.3 to gradio_client==1.4.3, as the package was not found on PyPI.
Installed dependencies from requirements.txt: Successfully ran pip install -r requirements.txt after fixing the typo and using Python 3.10. All packages reported as successfully installed.
Explicitly reinstalled PyTorch for CUDA 12.8: Ran pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128 within the activated echomimic environment, based on my CUDA 12.9 driver support and PyTorch website recommendations. pip reported requirements already satisfied, indicating the packages were present.
Installed/Repaired Microsoft Visual C++ Redistributables: Downloaded and installed/repaired the latest x64 and x86 Visual C++ Redistributable packages and restarted the computer.
Checked and cleaned System PATH: Found entries for CUDA 11.8 and 12.6 in the system's Environment PATH variable. Removed these older versions and restarted the computer.
Clean Reinstallation of NVIDIA Driver: Used DDU to perform a clean uninstallation of the NVIDIA driver in Safe Mode, then installed the latest driver (version 576.02) downloaded from the NVIDIA website. Restarted the computer.
Attempted to run the application after each step: Activated the echomimic environment and tried running python app_en.py after each troubleshooting step, but the OSError: [WinError 127] persists unchanged.
Additionally, I have noticed that my NVIDIA Control Panel/application is not opening, which might be related to the driver issue, although the clean reinstallation did not resolve the application loading error or the c10_cuda.dll error.

I am seeking assistance to understand why this c10_cuda.dll loading error is occurring and how to resolve it, as I have exhausted the standard troubleshooting steps. Any guidance or specific knowledge related to this error within this project or the Pinokio environment would be greatly appreciated.

Thank you for your time and help.