NewBytesweep Storage Cleaner is now on Google Play for Android
Bytesweep
Download

Getting a hundred gigabytes of model weights off your boot drive

When a boot drive is full of model weights, moving them is usually better than deleting them. Weights are read once when a model loads and then live in memory, so an external SSD costs you a slower first load and nothing after that. What matters is using each tool’s own setting for the move, not dragging directories across and leaving symlinks behind.

Pick the file system before anything else

You make this choice once, and changing it later means copying everything off and reformatting.

Decision chart: if only Macs will use the drive, format it APFS; if only Linux machines, ext4; if it must also work with Windows, exFAT, which is fine for Ollama, LM Studio and ComfyUI but not for the Hugging Face cache, because exFAT has no symlinks.
exFAT is the only format a Mac and a Windows PC can both write, and it’s the one format the Hugging Face cache can’t use properly.

exFAT is the obvious choice for a drive shared between machines, and it’s wrong for the Hugging Face cache. That cache is built on symlinks: every snapshot directory is a set of links into a shared blob store. exFAT has no symlinks, so huggingface_hub falls back to storing a full copy of each file in every snapshot, and the sharing between revisions is gone. NTFS isn’t a way out on a Mac either, since macOS can read NTFS but not write it without extra software.

So: APFS for a drive only Macs will use, ext4 for one only Linux will use. If it has to work with Windows too, use exFAT and keep the Hugging Face cache somewhere else.

Ollama

Ollama reads OLLAMA_MODELS. The macOS desktop app doesn’t read your shell profile, so the variable has to go to launchd:

launchctl setenv OLLAMA_MODELS /Volumes/AI/ollama
mv ~/.ollama/models /Volumes/AI/ollama
# quit Ollama from the menu bar, then start it again

launchctl setenv doesn’t survive a restart of the Mac, so run it again after a reboot, or use the model location field in Settings if your version of the Ollama app has one.

On Linux with the systemd service, the variable goes in the unit:

sudo systemctl edit ollama
# [Service]
# Environment="OLLAMA_MODELS=/mnt/ai/ollama"
sudo chown -R ollama:ollama /mnt/ai/ollama
sudo systemctl restart ollama

Check with ollama list. If it comes back empty, Ollama is looking at the new location and the data hasn’t moved across. The layout of Ollama’s store covers what’s inside it.

Hugging Face

HF_HOME moves the whole Hugging Face directory, including the hub cache and the datasets cache, which defaults to $HF_HOME/datasets. Set HF_DATASETS_CACHE only if you want datasets somewhere different from models.

export HF_HOME=/Volumes/AI/huggingface

Copy the existing cache with rsync -a, which keeps the symlinks as symlinks. A copy that follows links (cp -L, or rsync without -a or -l) writes a full copy of each blob for every snapshot that links to it.

rsync -a ~/.cache/huggingface/ /Volumes/AI/huggingface/

Put the export in ~/.zshrc for your shells. Jupyter kernels and processes started by an IDE may not inherit it, and reading the Hugging Face cache shows how to check which location a process is actually using.

LM Studio, ComfyUI and the Python caches

LM Studio has a setting for it: open My Models and change the models folder there. That’s the whole procedure. Read the layout of its store first if you plan to share files with llama.cpp.

ComfyUI has no environment variable for its models. It reads extra_model_paths.yaml in its install directory, which can point each kind of model at a folder on the external drive. ComfyUI’s model sprawl goes through the file.

For the rest of the Python ecosystem, two variables do most of the work:

export TORCH_HOME=/Volumes/AI/torch          # torchvision and torch.hub weights
export XDG_CACHE_HOME=/Volumes/AI/cache      # a great many tools default under this

PyTorch and TensorFlow’s own caches lists which tools follow XDG_CACHE_HOME and which don’t.

Move one store at a time, largest first, and check the tool still works before you start the next. A symlink from the old location to the new one usually works too, but it can fail badly when an app replaces the directory instead of writing into it, so use the supported setting wherever there is one.

What you give up

Load time is the only cost, and it’s arithmetic: size divided by read speed. A 40 GB model read at 2.8 GB/s, which a good Thunderbolt SSD manages, takes about fifteen seconds. Over USB 3.2 at 1 GB/s it’s about forty. Once loaded, inference runs at the same speed as from the internal drive, because nothing more is being read. A spinning disk or a network share turns the load into minutes; gigabit Ethernet moves roughly 0.11 GB/s, so the same model takes around six minutes.

Plan for the drive not being attached. Each tool fails differently when its store is missing: Ollama shows an empty model list, Hugging Face tries to download everything again, and ComfyUI starts with no checkpoints. Check the drive is mounted before you conclude something deleted your models.

An SSD doesn’t spin down, but macOS can put an idle external disk to sleep depending on your energy settings. If loads fail intermittently after the machine has been idle, turn off the setting that puts disks to sleep.

Keep reading