A fix release: Maya-XS updates start cleanly and give back their disk space, and Qwen CYBER-FROST runs on 8 GB cards again.
- DeepSeek Maya-XS v1 -> v2 update: a start could stop with "changed after the pack was built" even after
--setup --repack, because the old copy of the n-gram tables was checked instead of the shared one the model reads (#777, thanks @waxinz). Running the setup once also turns that old copy into a link to the shared one, giving back about 161 GB. - Qwen CYBER-FROST on 8 GB cards: v1.1.48 could run out of GPU memory on the first request or during the speed calibration. Maya now keeps room for the first request's buffers on small cards; bigger cards are unchanged. On an RTX 3050 8 GB it writes about as fast as before or faster, and the speed calibration completes.
Maya is free; if it helps you, you can sponsor it: https://github.com/sponsors/mw00