Idea: MFlux should _default_ to downloading _suitable_ pre-quantized MFlux models... instead of unquantized HF Models #670
ianscrivener
started this conversation in
Ideas & Suggestions: Features
Say a new user runs mflux-generate-z-image-turbo -p 'a puffin on a cliff top' on a 16GB Mac that has no modelos downloaded..
Currently MFlux will try to download the full 32.9Gb Tongyi-MAI/Z-Image-Turbo models and _try_ to quantize it to a bf16 format models. This will likely crash.
Instead MFlux should (1) check the available memory and (2) download a suitably sized models... perhaps q3 text_encoder and a q5 transformer models.
All reactions
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment