tilelang.tileop.gemm_blockscaled¶

Common block-scaled GEMM tile op.

Submodules¶

Classes¶

GemmBlockScaled

Block-scaled GEMM tile op: C (+)= (A * SFA) @ (B * SFB).

Package Contents¶

class tilelang.tileop.gemm_blockscaled.GemmBlockScaled¶

Bases: tilelang.tileop.gemm.Gemm

Block-scaled GEMM tile op: C (+)= (A * SFA) @ (B * SFB).

A GemmNode subclass on the C++ side, so it shares the dense GEMM’s operand layouts, warp partition and scheduling, and is lowered through the same tl.gemm.infer_layout / tl.gemm.lower entry points, which dispatch on this Python class. The extra FFI fields are sfaRegion, sfbRegion and sfKStart. The C++ selector returns a block-scaled instruction key (cuda.tcgen05.blockscaled, cuda.mma.blockscaled) that the backend registry maps to an implementation class built on GemmBlockScaledMixin, so dense implementation classes never see the scale factors.

property SFARegion¶
property SFBRegion¶
property sf_k_start¶
property is_blockscaled: bool¶
Return type:

bool