Near-roofline NVFP4 quantization kernel #4516

Closed

santoshmo wants to merge 11 commits into

pytorch:mainfrom

santoshmo:nvfp4-rht-cutedsl

Commits on Jun 19, 2026

mxfp4-rht: E2M1 PTX pack helper + bit-exact packing test
Santosh Mohan
committed
mxfp4-rht: FP4 E8M0 scale helpers (floor+rceil), bit-exact vs eager to_mx
Santosh Mohan
committed
mxfp4-rht: register FWHT(32)+sign transform, validated vs hadamard_matrix(32)
Santosh Mohan
committed
mxfp4-rht: fused FWHT+E2M1 CuTeDSL kernel + custom op + gated wrapper
Santosh Mohan
committed
mxfp4-rht: e2e correctness (bit-exact vs emulated to_mx on identical RHT + SQNR vs dense)
Santosh Mohan
committed
mxfp4-rht: wire fused RHT cast into MXTensor.to_mx (opt-in, trailing params)
Santosh Mohan
committed
mxfp4-rht: add cast benchmark mode
Santosh Mohan
committed
mxfp4-rht: align license headers with torchao convention
Santosh Mohan
committed
nvfp4-rht: E4M3 two-level scale helpers, bit-exact vs eager NVFP4
Santosh Mohan
committed
nvfp4-rht: fused FWHT(16)+E2M1 CuTeDSL kernel (E4M3 two-level scale, optional RHT) + op
Santosh Mohan
committed
nvfp4: max-bandwidth CuTeDSL quantize kernel + selectable NVFP4Tensor backend
Santosh Mohan
committed