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

composite_rms_norm_rope_shape

def composite_rms_norm_rope_shape[dtype: DType, cos_sin_dtype: DType, rank: Int](input: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input.static_spec], weight: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=weight.static_spec], epsilon: Float32, weight_offset: Scalar[dtype], cos_vals: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=cos_vals.static_spec], sin_vals: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=sin_vals.static_spec]) -> IndexList[rank]

Computes the output shape for the mo.composite.rms_norm_rope graph op.

Parameters:

dtype (DType): Element type of the input and weight tensors.
cos_sin_dtype (DType): Element type of the cos_vals and sin_vals RoPE tables.
rank (Int): Number of dimensions in the input and output tensors.

Args:

input (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input.static_spec]): Activation tensor normalized by RMS then rotated by RoPE; the last dimension must be even.
weight (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=weight.static_spec]): Per-column scale weights applied after RMS normalization.
epsilon (Float32): Small constant added inside the RMS normalization square root for numerical stability.
weight_offset (Scalar[dtype]): Scalar offset added to weight before scaling.
cos_vals (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=cos_vals.static_spec]): Cosine table used by the RoPE rotation, matching input rank.
sin_vals (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=sin_vals.static_spec]): Sine table used by the RoPE rotation, matching input rank.

Returns:

IndexList[rank]: The output shape, which matches the input shape.