tvm_tirx.tirx.tile.views.GemmAsync

Contents

tvm_tirx.tirx.tile.views.GemmAsync#

class tvm_tirx.tirx.tile.views.GemmAsync#

Bases: TensorCall

General matrix multiplication asynchronously.

Supports two arg layouts: - Regular (6 args): C, A, B, transA, transB, accum - Block-scaled (8 args): C, A, B, SFA, SFB, transA, transB, accum

kind = 'gemm_async'#

output#

lhs#

rhs#

property is_block_scaled: bool#: Whether this is a block-scaled MMA operation.

property sfa#: Get the scale factor tensor for A (None for regular MMA).

property sfb#: Get the scale factor tensor for B (None for regular MMA).

property transA#

property transB#

property accum#

property srcs: list[Expr]#: Get the source matrices (including scale factors if block-scaled).

property dsts: list[Expr]#: Get the destination matrix.