Bug: TorchAudio/PyTorch CUDA version mismatch in v0.24.0
Describe the bug
ComfyUI fails to start with a RuntimeError due to a CUDA version mismatch between PyTorch and TorchAudio.
Error
RuntimeError: Detected that PyTorch and TorchAudio were compiled with different CUDA versions.
PyTorch has CUDA version 13.0 whereas TorchAudio has CUDA version 12.8.
Please install the TorchAudio version that matches your PyTorch version.
Full traceback path
main.py -> execution.py -> latent_preview.py -> comfy/sd.py
-> comfy/ldm/lightricks/vae/audio_vae.py -> torchaudio
Environment
- Image:
ghcr.io/clsferguson/comfyui-docker:v0.24.0
- GPU: NVIDIA A40
- PyTorch version (inside container):
2.12.0+cu130
- TorchAudio CUDA version:
12.8
Root cause
The image appears to have been built with PyTorch from the cu130 (nightly) wheel index, but TorchAudio was pulled from the cu128 stable index. These need to match.
Per the repo README, the intended base is CUDA 12.8 with PyTorch from the cu128 wheel index — so either PyTorch was inadvertently upgraded to a nightly cu130 build, or TorchAudio was not updated to match.
Workaround
Manually reinstalling a matching TorchAudio nightly inside the running container resolves the issue:
docker exec -it ComfyUI-latest bash -c \
"pip install --upgrade --pre torchaudio \
--index-url https://download.pytorch.org/whl/nightly/cu130"
docker restart ComfyUI-latest
Expected behavior
PyTorch and TorchAudio should be pinned to the same CUDA version at build time so the container starts without errors.
Steps to reproduce
- Pull
ghcr.io/clsferguson/comfyui-docker:v0.24.0
- Start the container
- Observe the RuntimeError in container logs
Bug: TorchAudio/PyTorch CUDA version mismatch in v0.24.0
Describe the bug
ComfyUI fails to start with a
RuntimeErrordue to a CUDA version mismatch between PyTorch and TorchAudio.Error
Full traceback path
Environment
ghcr.io/clsferguson/comfyui-docker:v0.24.02.12.0+cu13012.8Root cause
The image appears to have been built with PyTorch from the
cu130(nightly) wheel index, but TorchAudio was pulled from thecu128stable index. These need to match.Per the repo README, the intended base is CUDA 12.8 with PyTorch from the
cu128wheel index — so either PyTorch was inadvertently upgraded to a nightly cu130 build, or TorchAudio was not updated to match.Workaround
Manually reinstalling a matching TorchAudio nightly inside the running container resolves the issue:
Expected behavior
PyTorch and TorchAudio should be pinned to the same CUDA version at build time so the container starts without errors.
Steps to reproduce
ghcr.io/clsferguson/comfyui-docker:v0.24.0