CoolFace
Modelpublic

sakamakismile/Huihui-gemma-4-26B-A4B-it-qat-abliterated-pad768-NVFP4

sourceHugging Facegemmaupdated 4mo agoView on Hugging Face
0likes695downloads
5 commits on main
bc857db4mo ago

Spec-decode section: add concurrent MTP results (341.5/557.6 JA agg; pays to ~4 streams, break-even at 8)

sakamakismile
1d3b0f24mo ago

README: point to the MTP bundle repo

sakamakismile
90ed2054mo ago

Add Speculative Decoding section: native gemma4_mtp (QAT assistant) 134/163 tok/s; EAGLE-3 fails on Japanese

sakamakismile
51f01764mo ago

QAT cures MoE W4A4 Japanese long-form collapse; +TP=2 vs TP=4 table (TP=2 sweet spot)

sakamakismile
e7fb71d4mo ago

initial commit

sakamakismile