For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

naive_block_scaled_matmul_kernel

def naive_block_scaled_matmul_kernel[c_type: DType, a_type: DType, b_type: DType, a_scales_type: DType, b_scales_type: DType, accum_type: DType, a_layout: Layout, b_layout: Layout, c_layout: Layout, a_scale_layout: Layout, b_scale_layout: Layout, scaling_kind: UMMAKind, SF_VECTOR_SIZE: Int, transpose_b: Bool = True, elementwise_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None] = None](c: LayoutTensor[c_type, c_layout, MutAnyOrigin], a: LayoutTensor[a_type, a_layout, MutAnyOrigin], b: LayoutTensor[b_type, b_layout, MutAnyOrigin], a_scales: LayoutTensor[a_scales_type, a_scale_layout, MutAnyOrigin], b_scales: LayoutTensor[b_scales_type, b_scale_layout, MutAnyOrigin], alpha: Float32)

Naive GPU kernel that emulates a block-scaled matmul using TCGEN scale factors.

Both A and B must be in K-major format with 5D TCGEN scale-factor layouts. Each thread accumulates one output element by iterating over K, applying per-block scale factors, and optionally invoking an elementwise epilogue lambda.

Parameters:

c_type (DType): Element type of the output matrix c.
a_type (DType): Element type of the LHS input matrix a.
b_type (DType): Element type of the RHS input matrix b.
a_scales_type (DType): Element type of the a_scales block scale-factor tensor.
b_scales_type (DType): Element type of the b_scales block scale-factor tensor.
accum_type (DType): Element type used for the dot-product accumulator.
a_layout (Layout): Memory layout of the LHS input matrix a.
b_layout (Layout): Memory layout of the RHS input matrix b.
c_layout (Layout): Memory layout of the output matrix c.
a_scale_layout (Layout): Memory layout of the a_scales block scale-factor tensor.
b_scale_layout (Layout): Memory layout of the b_scales block scale-factor tensor.
scaling_kind (UMMAKind): UMMAKind variant selecting the block-scaled MMA instruction.
SF_VECTOR_SIZE (Int): Number of elements covered by each block scale factor.
transpose_b (Bool): Whether b is stored transposed (defaults to True).
elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Optional epilogue lambda applied to the matmul result (defaults to None).