perf(BACKEND-ROCM): cooperate across Q8_K activation blocks - #2472

Merged

mudler merged 12 commits into

mudler:mainfrom

VikashLoomba:row/BACKEND-ROCM-Q8K-COOP

Sep 2, 2026

Commits on Aug 29, 2026

spec(BACKEND-ROCM): decompose the Q8_K activation quantizer
VikashLoomba
committed
87b4688
spec(BACKEND-ROCM): bind the reviewed Q8_K gate prompt
VikashLoomba
committed
212cce2
spec(BACKEND-ROCM): bound the Q8_K default to gfx1100
VikashLoomba
committed
6ae9f62
perf(BACKEND-ROCM): cooperate across each Q8_K activation block
VikashLoomba
committed
70d8d31
test(BACKEND-ROCM): witness the selected Q8_K kernel on device
VikashLoomba
committed
be70d25
perf(BACKEND-ROCM): select cooperative Q8_K quantization on gfx1100
VikashLoomba
committed
17f4568

Commits on Sep 1, 2026

Commits on Sep 2, 2026

merge: origin/main into row/BACKEND-ROCM-Q8K-COOP -- take the f16-codec header extraction
mudler
committed
48e814b