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

Mojo function

rms_norm_fused_residual_add

def rms_norm_fused_residual_add[dtype: DType, rank: Int, Input0Fn: def[width: Int, rank: Int](IndexList[rank]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable, Input1Fn: def[width: Int, rank: Int](IndexList[rank]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable, Output0Fn: def[width: SIMDLength, rank: Int, alignment: Int](IndexList[rank], SIMD[dtype, width]) -> None & RegisterPassable & ImplicitlyCopyable, OutputResidualFn: def[width: SIMDLength, rank: Int, alignment: Int](IndexList[rank], SIMD[dtype, width]) -> None & RegisterPassable & ImplicitlyCopyable, AxisSizeT: CoordLike, /, target: StringSpan[ImmStaticOrigin] = StringSpan("cpu"), multiply_before_cast: Bool = True, reduce_dim: Int = (rank - Int(1))](input_0_fn: Input0Fn, input_1_fn: Input1Fn, output_0_fn: Output0Fn, output_residual_fn: OutputResidualFn, shape: Coord, axis_size: AxisSizeT, gamma1: TileTensor[dtype, Storage=gamma1.Storage, address_space=gamma1.address_space, linear_idx_type=gamma1.linear_idx_type], epsilon1: Scalar[dtype], weight_offset1: Scalar[dtype], gamma2: TileTensor[dtype, Storage=gamma2.Storage, address_space=gamma2.address_space, linear_idx_type=gamma2.linear_idx_type], epsilon2: Scalar[dtype], weight_offset2: Scalar[dtype], context: Optional[DeviceContext] = None) where identical(Input0Fn.dtype, dtype) where identical(Input1Fn.dtype, dtype) where identical(Output0Fn.dtype, dtype) where identical(OutputResidualFn.dtype, dtype)