Struct Glm5NextDsaKernels 

pub struct Glm5NextDsaKernels {
    pub kpool_compress: KernelHandle,
    pub compact_pools: KernelHandle,
    pub index_scores: KernelHandle,
    pub topk_pools: KernelHandle,
    pub expand_selection: KernelHandle,
    pub k_norm: KernelHandle,
    pub write_geom: KernelHandle,
    pub indexer_store: KernelHandle,
    pub topk_to_mask: KernelHandle,
    pub mla_masked_attn: KernelHandle,
}

Every kernel the DSA path launches.

Resolved with kernel() (not try_kernel): a missing DSA entry point is a hard error, never a silent fallback onto a dense-attention path. A sparse layer that quietly runs dense is a correctness bug that looks like a performance bug.

Fields

kpool_compress: KernelHandle

compact_pools: KernelHandle

index_scores: KernelHandle

topk_pools: KernelHandle

expand_selection: KernelHandle

k_norm: KernelHandle

indexer.k_norm, which is an nn.LayerNorm with a bias — not an RMSNorm.

🪤 Do NOT reach for an RMSNorm kernel here. A .weight-only norm silently drops both the mean subtraction and the bias, and nothing about the shapes says so: k_norm.weight and k_norm.bias are both [index_head_dim]. The binder already lists the bias as REQUIRED for exactly this reason.

✅ No new kernel needed — common/nllb_encoder.cu already carries an in-place BF16 LayerNorm taking (x, weight, bias, rows, dim, eps), and common/ is merged into every target. Found by grepping kernels/ before scoping a build, per the campaign’s standing rule; this is the fifth thing that turned out to already exist.

write_geom: KernelHandle

Derives this step’s selector geometry ON DEVICE from seq_len, so a captured graph replays over the live context instead of the capture-time one. try_kernel — without it the layer keeps the host-scalar path and graphs stay off.

indexer_store: KernelHandle

Places the staged indexer row at a DEVICE-side position and marks it valid. The host k_normed.offset(pos * D * 2) it replaces was the other frozen scalar.

topk_to_mask: KernelHandle

🔬 ORACLE ONLY — see MASKED_ATTN_MAX_KEYS.

mla_masked_attn: KernelHandle

🔬 ORACLE ONLY — see MASKED_ATTN_MAX_KEYS.

Implementations

impl Glm5NextDsaKernels

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

Trait Implementations

impl Clone for Glm5NextDsaKernels

fn clone(&self) -> Glm5NextDsaKernels

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for Glm5NextDsaKernels