search

dbolya / tomesd

Speed up Stable Diffusion with this one simple trick!
MIT License
1.29k stars 78 forks source link

Question: how to save optimized model? #26

Closed recoilme closed 1 year ago

recoilme commented 1 year ago

i try to load to pipe, apply patch and save result to diffusers format and convert to ckpt, but get same file size.. 2.13Gb Is it possible to store/save optimized model?

output during conversion:

Reshaping encoder.mid.attn_1.q.weight for SD format
Reshaping encoder.mid.attn_1.k.weight for SD format
Reshaping encoder.mid.attn_1.v.weight for SD format
Reshaping encoder.mid.attn_1.proj_out.weight for SD format
Reshaping decoder.mid.attn_1.q.weight for SD format
Reshaping decoder.mid.attn_1.k.weight for SD format
Reshaping decoder.mid.attn_1.v.weight for SD format
Reshaping decoder.mid.attn_1.proj_out.weight for SD format
dbolya commented 1 year ago

ToMe has no parameters doesn't affect the weights of the model at all, so no. The tomesd code needs to be run for tokens to be merged.