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 struct

MatmulFusedPartialRMSNorm

struct MatmulFusedPartialRMSNorm

Fuses GEMV (M=1 matmul) with partial RMS normalization.

Computes y = x @ W.T, then applies RMS normalization to the first N_normed columns while passing the remaining columns through unchanged.

Implemented traits

AnyType, Deinitable, Movable

Methods

execute

static def execute[dtype: DType, rank: Int, target: StringSpan[ImmStaticOrigin], transpose_b: Bool = True](normed_output: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=normed_output.static_spec], unnormed_output: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=unnormed_output.static_spec], input: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input.static_spec], weight: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=weight.static_spec], gamma: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=gamma.static_spec], epsilon: Float32, weight_offset: Scalar[dtype], ctx: DeviceContext)

Execute fused GEMV + partial RMS norm.

Calls gemv_and_partial_norm from nn.gemv_partial_norm which computes y = x @ W.T, then partitions y into normed and unnormed outputs.

Parameters:

• ​dtype (DType): Element type of all input and output tensors.
• ​rank (Int): Tensor rank of the input and output tensors.
• ​target (StringSpan[ImmStaticOrigin]): The target GPU device.
• ​transpose_b (Bool): Whether to transpose weight before the matmul (defaults to True).

Args:

• ​normed_output (ManagedTensorSlice[IOSpec[_, _].Output, static_spec=normed_output.static_spec]): Output tensor holding the RMS-normalized leading columns of the matmul result.
• ​unnormed_output (ManagedTensorSlice[IOSpec[_, _].Output, static_spec=unnormed_output.static_spec]): Output tensor holding the trailing columns of the matmul result, passed through unchanged.
• ​input (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input.static_spec]): Input activation tensor x of the GEMV.
• ​weight (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=weight.static_spec]): Weight matrix W of shape (N, K) (rank 2).
• ​gamma (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=gamma.static_spec]): RMS normalization scale vector (rank 1).
• ​epsilon (Float32): Small constant added to the squared mean before the reciprocal square root for numerical stability.
• ​weight_offset (Scalar[dtype]): Reserved for API consistency with other RMS norm ops; not consumed by this kernel.
• ​ctx (DeviceContext): The device context used to enqueue the kernel.