Torch is now required dependency?

[weiji14]

Describe the bug
A clear and concise description of what the bug is.

Got this ModuleNotFoundError at conda-forge/python-cudnn-frontend-feedstock#32 (comment)

 │ │ Traceback (most recent call last):
 │ │   File "$SRC_DIR/conda_build_script.py", line 1, in <module>
 │ │     import cudnn
 │ │   File "$PREFIX/lib/python3.10/site-packages/cudnn/__init__.py", line 223, in <module>
 │ │     from .wrapper import Graph
 │ │   File "$PREFIX/lib/python3.10/site-packages/cudnn/wrapper.py", line 40, in <module>
 │ │     import torch
 │ │ ModuleNotFoundError: No module named 'torch'

Error introduced in PR #174 (v1.15.0)

Expected behavior
A clear and concise description of what you expected to happen.

PyTorch is not required to import cudnn, since it is not listed in requirements.txt

System Environment (please complete the following information):

• cudnn_frontend version: v1.5.0
• cudnn_backend version: v9.13.1.26
• GPU arch: N/A
• cuda runtime version: 12.9
• cuda driver version: N/A
• host compiler: ?
• OS: Ubuntu 24.04.3 LTS

API logs
Please attach API logs for both cudnn_frontend and cudnn_backend.

// For cudnn_frontend
export CUDNN_FRONTEND_LOG_FLIE=fe.log
export CUDNN_FRONTEND_LOG_INFO=1

// For cudnn_backend
export CUDNN_LOGLEVEL_DBG=3
export CUDNN_LOGDEST_DBG=be.log

To Reproduce
Steps to reproduce the behavior:

1. '...'
2. '....'
3. '....'

Additional context
Add any other context about the problem here.

Unsure if torch is meant to be a required or optional dependency now. If it has become required, lemme know and I can update the packaging info on conda-forge.

[Anerudhan]

Hi @weiji14
Thanks for pointing this out.

It should not be, but it is. We will try to rectify this as part of a patch release.

Just curious, what framework do you use to allocate/manage device tensor memory?

Thanks

[weiji14]

Thanks, looking forward to the patch release.

I mostly use pytorch to allocate tensor memory, but also interested in supporting other frameworks (e.g. cupy) 🙂