Struct Glm5NextMhcKernels 

pub struct Glm5NextMhcKernels {
    pub hc_expand: KernelHandle,
    pub hc_pre: KernelHandle,
    pub hc_mix: KernelHandle,
    pub hc_mix_bf16: KernelHandle,
    pub hc_finish: KernelHandle,
    pub hc_post: KernelHandle,
    pub hc_head: KernelHandle,
}

2026-09-25: Every kernel the GLM-5.3 hyper-connection launches. Copy: the loader resolves them once and every layer holds a copy.

Fields

hc_expand: KernelHandle

2026-09-25: Broadcast the embedding into the hc_mult highway streams; first text layer only. A GLM target cannot use hyper_connection::hc_expand: that kernel is in the DeepSeek-V4 model directory, and a target compiles only common/ and its own model directory.

hc_pre: KernelHandle

2026-09-25: The fused glm5next_hc_pre, one block per token. glm_hc_pre does not launch it; examples/glm5next_hc_split_gate.rs uses it as the oracle for hc_mix + hc_finish.

hc_mix: KernelHandle

2026-09-25: glm5next_hc_mix: one block per (token, mixing row), grid (T, mix_hc), reading an FP32 hc_fn.

hc_mix_bf16: KernelHandle

2026-09-25: glm5next_hc_mix reading a BF16 hc_fn. Widening BF16 to FP32 is exact, so it computes the FP32 kernel’s result on the widened weights. Optional (try_kernel, 0 when absent); glm_hc_pre uses it when the site’s hc_fn_bf16 is set and the handle is not 0.

hc_finish: KernelHandle

2026-09-25: glm5next_hc_finish: from the mixes, post, comb (Sinkhorn) and the collapsed row y.

hc_post: KernelHandle

hc_head: KernelHandle

Implementations

impl Glm5NextMhcKernels

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

2026-09-25: Resolve the kernels. A missing one is an error, except hc_mix_bf16, which is 0 when absent.

Trait Implementations

impl Clone for Glm5NextMhcKernels

fn clone(&self) -> Glm5NextMhcKernels

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for Glm5NextMhcKernels