residual_add_rms_norm_gatef32

Function residual_add_rms_norm_gatef32 

Source
pub fn residual_add_rms_norm_gatef32(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    hidden: DevicePtr,
    src: DevicePtr,
    weight: &DenseWeight,
    output: DevicePtr,
    output_f32: DevicePtr,
    residual: DevicePtr,
    num_tokens: u32,
    hidden_size: u32,
    eps: f32,
    stream: u64,
) -> Result<()>
Expand description

Dual-output fused residual add + RMS norm (ATLAS_FP32_ROUTING).

Same as residual_add_rms_norm (bf16 hidden/residual/output unchanged) but ALSO writes the normed output in FP32 to output_f32 for the MoE router GEMM, removing the norm’s bf16-store rounding from the routing-critical path.

Kernel: residual_add_rms_norm_gatef32(hidden, src, weight, output, output_f32, residual, hidden_size, eps) Grid: (num_tokens, 1, 1) Block: (min(hidden_size, 1024), 1, 1)