[feat] Add flydsl based grouped gemm#384

Merged

RuibinCheung merged 13 commits into

mainfrom

dev/kyle/flydsl_grp_gemm

Jun 22, 2026

Commits on Jun 17, 2026

flydsl grouped fp8 GEMM: M-grouped fwd(NT)/dgrad(NN) + variable-K wgrad(TN)

kyle-256
authored and
kyle
committed
flydsl grouped fp8: wire backend into grouped GEMM dispatch

kyle-256
authored and
kyle
committed
flydsl grouped fp8: merge per-layout persistent/non-persistent NT & NN kernels

kyle-256
authored and
kyle
committed
flydsl grouped fp8: drop persistent wgrad, keep only the masked chunked kernel

kyle-256
authored and
kyle
committed
flydsl grouped fp8: int64 indexing for inputs + output (handles A/B/C > 2^31 elems / > 4GB)

kyle-256
authored and
kyle
committed
flydsl dense fp8: int64 inputs (foldable + <=4GB traversal) + mode-split launch cache

kyle-256
authored and
kyle
committed
flydsl grouped fp8 bwd: small-M dgrad/wgrad autotune + skew-robust wgrad

kyle-256
authored and
kyle
committed
flydsl grouped wgrad: share masked/persistent tile decode + i64 rebase

kyle-256
authored and
kyle
committed
rename fp8_gemm_helper -> gemm_helper + rebase main
kyle
committed
test(grouped_gemm): add FlyDSL backend coverage to tensorwise tests
kyle
committed
fix(flydsl grouped): nt_vmcnt=-1 causes data hazard in non-persistent NT/NN kernels
kyle-256
committed
fix(flydsl utils): restore missing helpers + fix output store bugs
kyle-256
committed

Commits on Jun 22, 2026

feat(flydsl gemm): i64 traversal SRD re-base to lift the k*n / k*m < 2^32 cap
kyle-256
committed