feat: add FP8 blockwise dequantize/GEMM operators and FP8(E4M3) KV cache support in paged attention - #1568

Open

shsaihdsaiudh wants to merge 11 commits into

InfiniTensor:mainfrom

shsaihdsaiudh:feat/fp8-blockwise-dequantize

Commits on Sep 20, 2026

feat: add fp8_blockwise_dequantize operator
shsaihdsaiudh
committed
9c05256
build: allow cudnn=n NVIDIA builds and add sm_120

silasyyyang
authored and
shsaihdsaiudh
committed
1bd0bc9
feat: add fp8_blockwise_gemm operator (fused FP8 blockwise GEMM)

authored and
shsaihdsaiudh
committed
50dfacd
feat: add FP8(E4M3) KV cache support to paged attention operators
shsaihdsaiudh
committed
b4e42cc
feat: pass through KV scale params in non-nvidia paged attention backends

silasyyyang
authored and
shsaihdsaiudh
committed
d284f99
feat: add split-kv flash-decoding to FP8 paged attention decode
shsaihdsaiudh
committed
ca23006
feat: add tensor-core mma path to fp8_blockwise_gemm for 9 <= M <= 32
shsaihdsaiudh
committed
4363be5
fix: restore M_TILE template parameter on launch_mtile

silasyyyang
authored and
shsaihdsaiudh
committed
57124d8
fix: pass CUDA gencode via cugencodes so device-link inherits the arch
shsaihdsaiudh
committed
2c66ab5
fix: define paged-attention decode output for zero-length sequences
shsaihdsaiudh
committed
cfac104
fix: route F8 caches to the infiniop fallback in paged_caching dispatch
shsaihdsaiudh
committed
264ebc1