Pull requests / #897

#897 setup: the disk check counts only what step 6 still writes

closed · @bsvinay · 0 comentarios · En GitHub

Setup & installMulti-GPUModels & quantsLinux

Descripción

A re-run of setup on an already installed model can fail on the disk check even though nothing new needs to be written.

Seen on: Swift 1.5 IQ3_XXS in the low-RAM mode (52 GB RAM, 2x RX 7900 XTX, Linux), models already downloaded and the pack already prepared in Strata-data/packs/swift-iq3_xxs (experts.bin + index.txt present). Setup stopped at step 2 with:

    [X] not enough free disk space in /root/.ollama/models/swift-IQ3_XXS: need ~52 GB

The check always added the low-RAM copy (arena_gb + 1) and 8 GB for the MTP draft layer, even when both are already on the disk.

Change: count only what step 6 will still write.
- the pack's experts.bin (the low-RAM copy, or the AVX-512 Q2_0 conversion) counts only when it is not already there (experts.bin + index.txt);
- the MTP draft layer counts 2 GB instead of 8 when mtp/rt/experts.bin exists.

A fresh install asks for the same space as before. With the change, the same re-run passes the check and completes (steps 5-7 skip what is there).

En el sitio

Enlaces a install, modelos, releases.