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

group_norm_cpu

def group_norm_cpu[dtype: DType, rank: Int, //, InputFn: def[width: Int](Coord[*?]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable, GammaFn: def[width: Int](Coord[*?]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable, BetaFn: def[width: Int](Coord[*?]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable](input_fn: InputFn, gamma_fn: GammaFn, beta_fn: BetaFn, shape: Coord, epsilon: Float32, output: TileTensor[dtype, Engine=output.Engine, address_space=output.address_space, linear_idx_type=output.linear_idx_type], num_groups: Int, ctx: Optional[DeviceContext] = None)

Computes group normalization on CPU.

Reduces a single-pass Welford mean/variance over each (batch, group) block of channels_per_group * spatial elements, then applies the per-channel gamma/beta affine transform. Parallelizes across N * num_groups blocks.

Parameters:

• ​dtype (DType): Element type of the input and output tensors.
• ​rank (Int): Tensor rank of the input and output tensors (3 or 4).
• ​InputFn (def[width: Int](Coord[*?]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable): Type of the input load closure.
• ​GammaFn (def[width: Int](Coord[*?]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable): Type of the gamma load closure.
• ​BetaFn (def[width: Int](Coord[*?]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable): Type of the beta load closure.

Args:

• ​input_fn (InputFn): Function called to generate an input value.
• ​gamma_fn (GammaFn): Function called to generate a gamma value.
• ​beta_fn (BetaFn): Function called to generate a beta value.
• ​shape (Coord): The shape of the input/output tensor.
• ​epsilon (Float32): Small constant for numerical stability.
• ​output (TileTensor[dtype, Engine=output.Engine, address_space=output.address_space, linear_idx_type=output.linear_idx_type]): Output tensor receiving the normalized result.
• ​num_groups (Int): Number of groups the channel dimension is split into.
• ​ctx (Optional[DeviceContext]): Optional device context used to size the CPU thread pool.