Struct Qwen35Kernels

Source

pub struct Qwen35Kernels {Show 27 fields
    pub rms: KernelHandle,
    pub rope: KernelHandle,
    pub kvap: KernelHandle,
    pub attn: KernelHandle,
    pub sg: KernelHandle,
    pub add_rms: KernelHandle,
    pub qkv_split: KernelHandle,
    pub conv1d: KernelHandle,
    pub gdn_gate: KernelHandle,
    pub sigmoid: KernelHandle,
    pub gdn_dec: KernelHandle,
    pub kvap_turbo8: KernelHandle,
    pub attn_turbo8: KernelHandle,
    pub kvap_turbo4: KernelHandle,
    pub attn_turbo4: KernelHandle,
    pub kvap_turbo3: KernelHandle,
    pub attn_turbo3: KernelHandle,
    pub kvap_turbo2: KernelHandle,
    pub attn_turbo2: KernelHandle,
    pub kvap_bf16k_turbo4v: KernelHandle,
    pub attn_bf16k_turbo4v: KernelHandle,
    pub kvap_bf16k_turbo3v: KernelHandle,
    pub attn_bf16k_turbo3v: KernelHandle,
    pub kvap_bf16k_turbo2v: KernelHandle,
    pub attn_bf16k_turbo2v: KernelHandle,
    pub wht: KernelHandle,
    pub wht_inv: KernelHandle,
}

Expand description

Pre-resolved kernel handles. Resolve once at startup; pass & to every per-layer call so name-lookup overhead doesn’t appear in the hot path.

Fields§

§rms: KernelHandle§rope: KernelHandle§kvap: KernelHandle§attn: KernelHandle§sg: KernelHandle§add_rms: KernelHandle§qkv_split: KernelHandle§conv1d: KernelHandle§gdn_gate: KernelHandle§sigmoid: KernelHandle§gdn_dec: KernelHandle§kvap_turbo8: KernelHandle

TurboQuant KV cache paths (Turbo8/4/3/2): quantizing appends, dequantizing decode attentions, and the WHT rotation bookends. Resolved unconditionally (the kernels live in the common set) so a turbo cache can never silently fall back to the bf16 kernels.

§attn_turbo8: KernelHandle§kvap_turbo4: KernelHandle§attn_turbo4: KernelHandle§kvap_turbo3: KernelHandle§attn_turbo3: KernelHandle§kvap_turbo2: KernelHandle§attn_turbo2: KernelHandle§kvap_bf16k_turbo4v: KernelHandle§attn_bf16k_turbo4v: KernelHandle§kvap_bf16k_turbo3v: KernelHandle§attn_bf16k_turbo3v: KernelHandle§kvap_bf16k_turbo2v: KernelHandle§attn_bf16k_turbo2v: KernelHandle§wht: KernelHandle§wht_inv: KernelHandle

Qwen35Kernels

Struct Qwen35Kernels Copy item path

Fields§

Implementations§

impl Qwen35Kernels

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

Auto Trait Implementations§

impl Freeze for Qwen35Kernels

impl RefUnwindSafe for Qwen35Kernels

impl Send for Qwen35Kernels

impl Sync for Qwen35Kernels

impl Unpin for Qwen35Kernels

impl UnwindSafe for Qwen35Kernels

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> 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, 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 Qwen35Kernels

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

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>,