12 Commits

Author	SHA1	Message	Date
kijai	dafa2695d4	Code cleanup, don't force the fp32 layers as it has minimal effect	2025-12-03 16:08:47 +02:00
kijai	b24bac604a	Update chunking	2025-12-03 16:08:47 +02:00
kijai	a3ce1e02d7	Further reduce peak VRAM consumption by chunking ffn	2025-12-03 16:08:47 +02:00
kijai	1baf5ec4af	Reduce peak VRAM usage a bit	2025-12-03 16:08:47 +02:00
kijai	f1a5f6f5b3	Rever RoPE scaling to simpler one	2025-12-03 16:08:47 +02:00
kijai	68c712ef7e	Don't scale RoPE for lite model as that just doesn't work...	2025-12-03 16:08:46 +02:00
kijai	dd318ada2f	Support block replace patches (SLG mostly)	2025-12-03 16:08:46 +02:00
kijai	292c3576c2	Fix I2V, add necessary latent post process nodes	2025-12-03 16:08:46 +02:00
kijai	8f02217f85	Code cleanup, optimizations, use fp32 for all layers originally at fp32	2025-12-03 16:08:46 +02:00
kijai	0920cdcf63	Add transformer_options for attention	2025-12-03 16:08:46 +02:00
kijai	5f1346ccd1	Fix fp8	2025-12-03 16:08:46 +02:00
kijai	530df3d710	Add Kandinsky5 model support ... lite and pro T2V tested to work	2025-12-03 16:08:46 +02:00