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
ReduceRMSNormFusedResidualAdd
struct ReduceRMSNormFusedResidualAdd
Registers the mo.composite.rms_norm_fused_residual_add graph op with the graph compiler.
Implemented traits
Methods
execute
static def execute[epsilon1: Float32, epsilon2: Float32, weight_offset1: Float32, weight_offset2: Float32, multiply_before_cast: Bool = True](output: T, residual_output: T, input: T, residual_input: T, gamma1: T, gamma2: T, ctx: DeviceContext)
Executes the mo.composite.rms_norm_fused_residual_add graph op.
Parameters:
- epsilon1 (
Float32): Stability constant for the first normalization. - epsilon2 (
Float32): Stability constant for the second normalization. - weight_offset1 (
Float32): Scalar offset for the first weight. - weight_offset2 (
Float32): Scalar offset for the second weight. - multiply_before_cast (
Bool): See the graph op signature.
Args:
- output (
T): Output tensor receiving the result. - residual_output (
T): See the graph op signature. - input (
T): Input tensor to reduce. - residual_input (
T): Residual tensor added to the normalized input. - gamma1 (
T): Scale weights for the first normalization. - gamma2 (
T): Scale weights for the second normalization. - ctx (
DeviceContext): Device context used to enqueue the kernel.
Raises:
Error: If the operation parameters are invalid.