Struct Glm5NextKdaKernels

Source

pub struct Glm5NextKdaKernels {Show 16 fields
    pub gemm: KernelHandle,
    pub gemv: KernelHandle,
    pub gemv_batchm: KernelHandle,
    pub conv_decode: KernelHandle,
    pub conv_prefill: KernelHandle,
    pub l2: KernelHandle,
    pub gate: KernelHandle,
    pub chunk_prepare: KernelHandle,
    pub chunk_scan: KernelHandle,
    pub recurrent: KernelHandle,
    pub recurrent_smem: KernelHandle,
    pub o_norm: KernelHandle,
    pub split_widen: KernelHandle,
    pub sigmoid: KernelHandle,
    pub fill: KernelHandle,
    pub pack: KernelHandle,
}

Fields§

§gemm: KernelHandle§gemv: KernelHandle

🔴 The DECODE weight kernel. dense_gemm_bf16 tiles 16x16 over (N, M); at M=1 the grid collapses and it measured 58 GB/s against a 254 GB/s part — 32 % of the whole GLM decode step (2026-08-28 profile). Every KDA projection is M=1 at decode.

§gemv_batchm: KernelHandle

🔴 The BATCHED weight kernel, 2 ..= 8 rows in ONE weight sweep. This is what makes a K-token speculative verify cost one pass over q/k/v/f_a/f_b/g_a/g_b/o instead of K. 0 on a backend without it — ops::dense_mm_bf16 then falls back to the tile GEMM.

§conv_decode: KernelHandle§conv_prefill: KernelHandle§l2: KernelHandle§gate: KernelHandle§chunk_prepare: KernelHandle§chunk_scan: KernelHandle§recurrent: KernelHandle§recurrent_smem: KernelHandle

1R+1W sibling of recurrent, bit-identical: the decayed state column lives in shared memory between the two passes instead of being re-read from global. try_kernel — a target without it falls back to the 2R+2W kernel.

§o_norm: KernelHandle§split_widen: KernelHandle§sigmoid: KernelHandle§fill: KernelHandle§pack: KernelHandle

Struct Glm5NextKdaKernels Copy item path

Fields§

Implementations§

impl Glm5NextKdaKernels

pub const ENTRY_POINTS: usize = 14

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

Trait Implementations§

impl Clone for Glm5NextKdaKernels

fn clone(&self) -> Glm5NextKdaKernels

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

impl Copy for Glm5NextKdaKernels

Auto Trait Implementations§

impl Freeze for Glm5NextKdaKernels

impl RefUnwindSafe for Glm5NextKdaKernels

impl Send for Glm5NextKdaKernels

impl Sync for Glm5NextKdaKernels

impl Unpin for Glm5NextKdaKernels

impl UnwindSafe for Glm5NextKdaKernels

Blanket Implementations§

impl<T> Any for Twhere T: 'static + ?Sized,

fn type_id(&self) -> TypeId

impl<T> Borrow<T> for Twhere T: ?Sized,

fn borrow(&self) -> &T

impl<T> BorrowMut<T> for Twhere T: ?Sized,

fn borrow_mut(&mut self) -> &mut T

impl<T> CloneToUninit for Twhere T: Clone,

unsafe fn clone_to_uninit(&self, dest: *mut u8)

impl<T> From<T> for T

fn from(t: T) -> T

impl<T> Instrument for T

fn instrument(self, span: Span) -> Instrumented<Self>

fn in_current_span(self) -> Instrumented<Self>

impl<T, U> Into<U> for Twhere U: From<T>,

fn into(self) -> U

impl<T> ToOwned for Twhere T: Clone,

type Owned = T

fn to_owned(&self) -> T

fn clone_into(&self, target: &mut T)

impl<T, U> TryFrom<U> for Twhere U: Into<T>,

type Error = Infallible

fn try_from(value: U) -> Result<T, <T as TryFrom<U>>::Error>

impl<T, U> TryInto<U> for Twhere U: TryFrom<T>,

type Error = <U as TryFrom<T>>::Error

fn try_into(self) -> Result<U, <U as TryFrom<T>>::Error>

impl<V, T> VZip<V> for Twhere V: MultiLane<T>,

fn vzip(self) -> V

impl<T> WithSubscriber for T

fn with_subscriber<S>(self, subscriber: S) -> WithDispatch<Self>where S: Into<Dispatch>,

fn with_current_subscriber(self) -> WithDispatch<Self>

Struct Glm5NextKdaKernels

impl<T> Any for T
where T: 'static + ?Sized,

impl<T> Borrow<T> for T
where T: ?Sized,

impl<T> BorrowMut<T> for T
where T: ?Sized,

impl<T> CloneToUninit for T
where T: Clone,

impl<T, U> Into<U> for T
where U: From<T>,

impl<T> ToOwned for T
where T: Clone,

impl<T, U> TryFrom<U> for T
where U: Into<T>,

impl<T, U> TryInto<U> for T
where U: TryFrom<T>,

impl<V, T> VZip<V> for T
where V: MultiLane<T>,

fn with_subscriber<S>(self, subscriber: S) -> WithDispatch<Self>
where S: Into<Dispatch>,