Add TurboQuant3/4 modes to quantized_scaled_dot_product_attention#3453

Closed

dedalien wants to merge 23 commits into

ml-explore:mainfrom

dedalien:turboq/integrate-generic-quant-sdpa

Commits on Apr 23, 2026

first attempt

CC-Yeh
authored and
dedalien
committed
fix

CC-Yeh
authored and
dedalien
committed
Unify mxfp4/8 paths and optimize mxfp8 fused calculation

CC-Yeh
authored and
dedalien
committed
supports nvpf4

CC-Yeh
authored and
dedalien
committed
supports affine 4/8 bits

CC-Yeh
authored and
dedalien
committed
supports affine 2/3/5/6 bits

CC-Yeh
authored and
dedalien
committed
clean up

CC-Yeh
authored and
dedalien
committed
adapt #3023

CC-Yeh
authored and
dedalien
committed
Limit affine SDPA to group_size=32 and bits={4,6,8}

CC-Yeh
authored and
dedalien
committed
fix group_size for nvfp4 and simplify code

CC-Yeh
authored and
dedalien
committed
refactor: use function_constant to reduce template specializations(binary size)

CC-Yeh
authored and
dedalien
committed
tune blocks

CC-Yeh
authored and
dedalien
committed
fix

CC-Yeh
authored and
dedalien
committed
cleanup

CC-Yeh
authored and
dedalien
committed
cleanup + refactor

CC-Yeh
authored and
dedalien
committed
enable causal

CC-Yeh
authored and
dedalien
committed
support sinks

CC-Yeh
authored and
dedalien
committed
cleanup

CC-Yeh
authored and
dedalien
committed
cleanup

CC-Yeh
authored and
dedalien
committed
improvements

CC-Yeh
authored and
dedalien
committed
Support group_size=64 for affine quantized SDPA

dogukanveziroglu
authored and
dedalien
committed

Commits on Apr 26, 2026