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
Methods
execute
static def execute[epsilon: Float32, weight_offset: Float32, transpose_b: Bool = True](normed_output: T, unnormed_output: T, input: T, weight: T, gamma: T, 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:
- epsilon (
Float32): Small constant added to the squared mean before the reciprocal square root for numerical stability. - weight_offset (
Float32): Reserved for API consistency with other RMS norm ops; not consumed by this kernel. - transpose_b (
Bool): Whether to transposeweightbefore the matmul (defaults toTrue).
Args:
- normed_output (
T): Output tensor holding the RMS-normalized leading columns of the matmul result. - unnormed_output (
T): Output tensor holding the trailing columns of the matmul result, passed through unchanged. - input (
T): Input activation tensorxof the GEMV. - weight (
T): Weight matrixWof shape(N, K)(rank 2). - gamma (
T): RMS normalization scale vector (rank 1). - ctx (
DeviceContext): The device context used to enqueue the kernel.