ComfyUI: ModuleNotFoundError: No module named 'torchaudio' (fix)

ModuleNotFoundError: No module named 'torchaudio' stops ComfyUI v0.37.0 and older at startup. We reproduced it on v0.37.0, fixed it with a matching torchaudio, and checked that v0.38.0 no longer imports it.

1. The error

ComfyUI v0.37.0 on torch 2.9.0+cpu, all requirements installed except torchaudio, python main.py --cpu:

Traceback (most recent call last):
  File "ComfyUI/main.py", line 251, in <module>
    import execution
  File "ComfyUI/execution.py", line 24, in <module>
    from latent_preview import set_preview_method
  File "ComfyUI/latent_preview.py", line 5, in <module>
    from comfy.sd import VAE
  File "ComfyUI/comfy/sd.py", line 15, in <module>
    import comfy.ldm.lightricks.vae.audio_vae
  File "ComfyUI/comfy/ldm/lightricks/vae/audio_vae.py", line 4, in <module>
    import torchaudio
ModuleNotFoundError: No module named 'torchaudio'

It is not only the audio nodes: through v0.37.0, comfy/sd.py imports the LTX audio VAE, which imports torchaudio at the top, so ComfyUI exits before the server starts. torchaudio is listed in requirements.txt up to v0.37.0, so this happens when the install skipped it, e.g. a manual pip install torch torchvision or a requirements file edited to avoid reinstalling torch.

2. Fix A: install the torchaudio that matches your torch

torchaudio must be the same release and build as torch (2.9.0 with 2.9.0, same CPU or CUDA tag). Check torch first, then install from the same index:

python -c "import torch; print(torch.__version__)"
# 2.9.0+cpu  ->
pip install torchaudio==2.9.0 --index-url https://download.pytorch.org/whl/cpu

For a CUDA build use the matching index (e.g. /whl/cu128 for +cu128). After installing torchaudio 2.9.0+cpu, the same command initialised normally and --quick-test-for-ci exited with status 0. A torchaudio from a different release fails differently: libtorchaudio.so undefined symbol.

3. Fix B: update ComfyUI

In v0.38.0 torchaudio is gone from requirements.txt, from comfy_extras/nodes_audio.py and from comfy/ldm/lightricks/vae/audio_vae.py (we checked each file at both tags), so current ComfyUI does not need it. If you still see the error after updating, a custom node imports torchaudio: the traceback path under custom_nodes/ names it, and Fix A applies.

Frequently asked questions

Does ComfyUI still need torchaudio?

Not from v0.38.0: it was removed from requirements.txt and the core audio files. Up to v0.37.0 it is required at startup.

Which torchaudio version should I install?

The one with the same version and build tag as your torch, e.g. 2.9.0+cpu with torch 2.9.0+cpu, from the same PyTorch index.

Written with AI assistance. Measured on 2026-10-24 on Linux, CPython 3.11.17: ComfyUI v0.37.0 with torch 2.9.0+cpu, without and with torchaudio 2.9.0+cpu; v0.38.0 checked in source. Output pasted unedited except shortened paths. PlainTables is our extension; ComfyUI and torchaudio are free.