PyTorch Cache on Mac: Is ~/.cache/torch Safe to Delete?

#What PyTorch stores here

PyTorch Hub decides its download directory in a fixed order, according to the PyTorch docs:

1. A path set in code with torch.hub.set_dir().
2. $TORCH_HOME/hub, if TORCH_HOME is set.
3. $XDG_CACHE_HOME/torch/hub, if XDG_CACHE_HOME is set.
4. ~/.cache/torch/hub otherwise.

Without any of those set, the default on a Mac is ~/.cache/torch. Inside you'll usually find:

#PyTorch never cleans it up

The "Caching logic" section of the PyTorch docs states: "By default, we don't clean up files after loading it." Each new architecture or weights version adds a file, and vision or speech checkpoints often weigh hundreds of megabytes to a few gigabytes. Hub repos for every project you tried stay as well.

#What happens after you delete it

The next torch.hub.load or weights=... call downloads the files again and verifies them. That can be slow on large checkpoints, and it fails in environments without internet access. Training code, datasets you keep elsewhere, and your own saved models aren't affected. One nuance: torch.hub.load runs code from the downloaded repo, so a fresh download also fetches that repo's current code. That only matters if you relied on an older snapshot of an unpinned branch.

#Find out what is cached

du -sh ~/.cache/torch
du -sh ~/.cache/torch/hub/* ~/.cache/torch/hub/checkpoints/* 2>/dev/null | sort -h
python3 -c "import torch; print(torch.hub.get_dir())"

The last line prints the hub directory PyTorch will really use, which is useful if TORCH_HOME is set in your shell or a notebook kernel.

#Clearing it by hand

PyTorch has no cache-clean command, so manual removal is the method:

1. Stop running notebooks and training jobs.
2. In Finder, choose Go → Go to Folder… and enter ~/.cache/torch.
3. Move individual files in hub/checkpoints or repo folders in hub that you no longer need to the Trash, or remove the whole torch folder.

Deleting a single checkpoint is often enough. File names include a hash fragment (for example resnet50-<hash>.pth), so different weight versions of one model sit side by side, and only the one your code currently requests is needed.

On machines where you want downloads in one known place, setting TORCH_HOME to a dedicated folder makes the cache easier to find and clear later.

#Which checkpoints to keep

Checkpoints your current projects load on every run are worth keeping, more so if you sometimes work offline. Weights from finished experiments can go. Model files from the Transformers ecosystem are in the Hugging Face cache, and the developer caches overview lists the other places Python tooling keeps downloads.

#Sources

• PyTorch docs: torch.hub (where downloaded models are saved, caching logic)
• torchvision docs: Models and pre-trained weights