10 Commits

Author	SHA1	Message	Date
kijai	845eb14425	Handle thinking tokens different only for Gemma4	2026-04-13 23:40:52 +03:00
kijai	e0cccbd4c9	Fix image encoder attention mask type ... So it works with basic attention	2026-04-13 23:29:27 +03:00
kijai	387e8d8a4c	small fixes	2026-04-12 18:27:26 +03:00
kijai	0fc398a821	Various fixes	2026-04-12 18:11:07 +03:00
kijai	17ccff25be	Cleanup	2026-04-12 16:53:12 +03:00
kijai	6b803abe5a	update comment	2026-04-12 16:30:49 +03:00
kijai	6718be09ba	cleanup, enable fused rms norm by default	2026-04-10 15:28:26 +03:00
kijai	05eaceafa1	Cleanup, video fixes	2026-04-07 12:37:29 +03:00
kijai	93e8635110	parity with reference implementation ... outputs can 100% match transformers with same sdpa flags, checkpoint this and then optimize	2026-04-07 01:15:04 +03:00
kijai	832753f497	initial gemma4 support	2026-04-03 03:46:45 +03:00