It downloads GB's of Nvidia stuff, when the computer has an Intel GPU.

[HubKing]

It chose the CPU mode, so I don't think all those downloads were necessary.
Please. SSD prices have been tripled here.

[pacocartones]

Which version and install method did you use? The answer is different for the latest release and for main. This is from reading the code, not from running it on an Intel machine.

v0.5.6 (latest release, 23 Sep). On Linux and Windows x64, pyproject.toml at the v0.5.6 tag (lines 275-290) takes torch, torchaudio and torchvision from the pytorch-cuda index unconditionally. There is no CPU variant and no driver check, so every install on those platforms downloads the CUDA wheels, even if the app then runs on CPU. If that's your version, the download is expected, and there's no switch to avoid it there.

main since 4 Oct (not in a release yet). electron/scripts/torch-variant.mjs (added in b82bf0d60b) chooses the variant before installing. selectTorchVariant (lines 64-84) picks cuda if nvidiaDriverPresent (lines 5-35) finds any NVIDIA driver file. On Linux those are /proc/driver/nvidia/version, /dev/nvidiactl, a libcuda.so.1 in several lib dirs, or /usr/bin/nvidia-smi; on Windows, nvcuda.dll, nvml.dll or nvidia-smi.exe. If none is found it still asks nvidia-smi for a GPU (lines 37-53), and only then picks cpu. You can force it with OMNIVOICE_TORCH_VARIANT=cpu (or OMNIVOICE_DEVICE=cpu), which makes the sync run with --no-group cuda --group cpu (lines 86-89). There, CUDA would only be downloaded on an Intel-only machine if a leftover NVIDIA driver file is present.

Either way, the wheels already downloaded stay in uv's cache. uv cache clean frees that space, but it clears uv's cache for every project on the machine. All line numbers refer to main at 481cf4c431 and to the v0.5.6 tag.