ViDubb Crash with cuDNN (libcudnn_cnn) on Modern CUDA 12 / PyTorch 2.x Setups

[badbandsduna-dot]

System Info:

-OS: Ubuntu 24.04 (WSL, Miniconda)

-GPU: NVIDIA RTX 3060 (CUDA 12)

-Python: 3.10 (Conda env)

-PyTorch: 2.8.0+cu128

-cuDNN: v9.1 (all expected libcudnn_cnn.so.* symlinks present)

✅ What Works:

-TTS (with patched espeak_wrapper for modern espeak)

-Background sound extraction

-Video/file download from YouTube

-Gradio UI launches and starts processing

❌ What Fails

Pipeline crashes with:

Unable to load any of {libcudnn_cnn.so.9.1.0, libcudnn_cnn.so.9.1, libcudnn_cnn.so.9, libcudnn_cnn.so}
Invalid handle. Cannot load symbol cudnnCreateConvolutionDescriptor
Aborted (core dumped)

Occurs even if lip sync (Wav2Lip) is disabled via UI

ldd shows all libcudnn_cnn dependencies are present

🔍 Troubleshooting Attempted

-Installed latest CUDA/cuDNN via Conda and system packages

-Manually symlinked all relevant libcudnn_cnn.so.* versions

-Patched TTS/espeak wrapper for Ubuntu 24.04 compatibility

-Set and verified LD_LIBRARY_PATH, no missing libs

-Disabled Wav2Lip via UI and manually — crash still occurs

-Confirmed Wav2Lip was trained with PyTorch 1.10 (running on 2.8)

🧠 Suspected Root Cause

-Wav2Lip code is imported or initialized even when disabled in the Gradio UI.

-Wav2Lip depends on legacy CUDA/cuDNN APIs that are incompatible with modern PyTorch 2.x and CUDA 12+.

-The function cudnnCreateConvolutionDescriptor no longer exists in modern libcudnn_cnn, leading to a hard crash.

🙏 Feature/Dev Requests:

1.Add a strict toggle to fully disable all Wav2Lip imports and GPU code when lip sync is OFF

2.Refactor or upgrade Wav2Lip module to support PyTorch ≥ 2.x + CUDA 11+

3.Document compatible CUDA/cuDNN/PyTorch versions or provide Docker/conda setup for legacy environments

4.(Optional) Add a pre-run check to warn about known incompatibilities with CUDA/cuDNN on modern systems

📌 Additional Notes

-Manually commenting out Wav2Lip-related imports in app.py allows the rest of the app to run without issue

-Crash always points to:

-Cannot load symbol cudnnCreateConvolutionDescriptor

-despite all expected cuDNN libs being present and visible in LD_LIBRARY_PATH.

Files attached:

ViDubb Repo Logs.txt