Disk space check before download/installation

[mirh]

This probably applies both to the temp download folder and also to the target installation disk.

Deleting existing virtual environment...
Creating virtual environment...
> C:\Users\user\AppData\Local\Temp\2tHK8V6PQNi8BAls37RH9oI97Ms\resources\bin\uv.exe venv --relocatable --prompt invoke --python 3.11 --python-preference only-managed K:\AI\InvokeAI\.venv
Using CPython 3.11.11
Creating virtual environment at: K:\AI\InvokeAI\.venv
Activate with: K:\AI\InvokeAI\.venv\Scripts\activate
Installing invokeai package...
> C:\Users\user\AppData\Local\Temp\2tHK8V6PQNi8BAls37RH9oI97Ms\resources\bin\uv.exe pip install --python 3.11 --python-preference only-managed invokeai[xformers]==v5.9.1 --force-reinstall --compile-bytecode --index=https://download.pytorch.org/whl/cu124
Resolved 169 packages in 2.08s
  × Failed to download `torch==2.4.1+cu124`
  ├─▶ Failed to extract archive
  ╰─▶ failed to write to file
      `C:\Users\user\AppData\Local\uv\cache\.tmpKcFtn6\torch/lib/dnnl.lib`:
      There is not enough space on the disk. (os error 112)
  help: `torch` (v2.4.1+cu124) was included because `invokeai` (v5.9.1)
        depends on `torch`
Failed to install invokeai python package: Process exited with code 1

FWIW this also broke my old installation, which wasn't recognized anymore afterwards.

[psychedelicious]

Unfortunately, it is not easy to check for disk space before installing:

• Python package sizes may change as the packages are updated.
• Different platforms have differently-sized dependencies. For example, CUDA torch is drastically different from MPS torch.

I don't think we can really do much about this problem, sorry. You can try following the troubleshooting steps to clear the uv cache which may free up some disk space. You could also try purging the pip cache with pip cache purge.

It's expected that a failed update breaks an existing installation. You'll need to use the launcher's repair mode to fix it after freeing disk space.

[mirh]

I was going to say that pypi lets you know the size in advance.. Though to be the fairest that's just the wheel size, not the installed one(which can be easily 4x bigger). And it's not pytorch (which is the biggest thing).
So.. uhm, maybe you could just have some rough overall hardcoded check? Like for instance, I can see my .venv is 6.50GB folder here.
And yes, you could already call it a day with half of that if you were to select the CPU-only version, but otherwise I think deviations shouldn't be too big (rocm has a similar weight to cuda).

[psychedelicious]

Indeed, the wheel is tiny relative to the dependencies.

On the main branch HEAD, .venv is:

• Linux (CUDA): ~6.9GB
• macOS (MPS): ~1.7GB

Last week we updated a lot of deps and removed some unused deps. If I check out a ref prior to that cleanup, the numbers are different:

• Linux (CUDA): ~7.1GB
• macOS (MPS): ~1.4GB

There are other complications:

• uv caches packages. This cache location is determined by the system - it may or may not be located on the same disk as the Invoke install.
• uv attempts to hardlink packages into the venv but may, for a variety of reasons, fall back to copying packages.
• uv may clear its cache at any time. The system or user may also clear the cache.
• Users can install multiple versions of the app through the launcher.

So the actual disk space required to install can vary wildly, from ~1.4GB (macOS w/ hardlinked packages) to ~14GB (Linux w/ CUDA and copied packages).

I don't plan on working on this in the near term, but I'd be happy to review a contribution that provides a rough estimate to users. I suspect the uv hardlink vs copy behavior may be the trickiest part to predict.