Struct Glm5NextDsaLayerKernels 

pub struct Glm5NextDsaLayerKernels {
    pub gemm: KernelHandle,
    pub gemm_f32: KernelHandle,
    pub gemv: KernelHandle,
    pub gemv_f32: KernelHandle,
    pub gemv_batchm: KernelHandle,
    pub rms_norm: KernelHandle,
    pub latent_write: KernelHandle,
}

Every kernel a DSA block launches, beyond the selection set.

Fields

gemm: KernelHandle

C = A @ B^T, BF16 out.

gemm_f32: KernelHandle

Same, FP32 out — the selector wants q_idx and the head weights in FP32.

gemv: KernelHandle

M=1 twins of the two above. gemv_f32 may be a 0 handle on a target that predates dense_gemv_bf16_fp32out; gemm refuses nothing and falls back to the tile arm.

gemv_f32: KernelHandle

gemv_batchm: KernelHandle

🔴 dense_gemv_bf16_batchm — 2 ..= 8 rows in ONE weight sweep, the arm that makes a K-token verify pay for q_a/q_b/kv_a/kv_b/o once instead of K times. 0 = unavailable.

rms_norm: KernelHandle

🪤 vanilla = x * rms * w. The other rms_norm adds 1 to the weight.

latent_write: KernelHandle

RMSNorm + FP8 + paged slot write, GLM-target.

Implementations

impl Glm5NextDsaLayerKernels

pub fn resolve(gpu: &dyn GpuBackend) -> Result<Self>

Trait Implementations

impl Clone for Glm5NextDsaLayerKernels

fn clone(&self) -> Glm5NextDsaLayerKernels

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for Glm5NextDsaLayerKernels