Struct LayerDev

Source

pub struct LayerDev {Show 41 fields
    pub il: u32,
    pub ratio: usize,
    pub expert_kind: ExpertKind,
    pub wq_a: DenseBf16,
    pub wq_b: DenseBf16,
    pub wkv: DenseBf16,
    pub wo_a: DenseBf16,
    pub wo_b: DenseBf16,
    pub q_norm: CudaSlice<f32>,
    pub kv_norm: CudaSlice<f32>,
    pub attn_norm: CudaSlice<f32>,
    pub ffn_norm: CudaSlice<f32>,
    pub sink: CudaSlice<f32>,
    pub cmp: Option<CmpDev>,
    pub idx: Option<IdxDev>,
    pub hc_attn_fn: CudaSlice<f32>,
    pub hc_ffn_fn: CudaSlice<f32>,
    pub hc_attn_base: Vec<f32>,
    pub hc_attn_scale: Vec<f32>,
    pub hc_ffn_base: Vec<f32>,
    pub hc_ffn_scale: Vec<f32>,
    pub hc_attn_base_dev: CudaSlice<f32>,
    pub hc_attn_scale_dev: CudaSlice<f32>,
    pub hc_ffn_base_dev: CudaSlice<f32>,
    pub hc_ffn_scale_dev: CudaSlice<f32>,
    pub gate_bias_dev: Option<CudaSlice<f32>>,
    pub tid2eid_dev: Option<CudaSlice<i32>>,
    pub experts_s2_dev: CudaSlice<f32>,
    pub gate_w: CudaSlice<f32>,
    pub gate_bias: Option<Vec<f32>>,
    pub tid2eid: Option<Vec<i64>>,
    pub experts_w: CudaSlice<u8>,
    pub experts_sc: CudaSlice<u8>,
    pub experts_s2: Vec<f32>,
    pub shared_w: [DenseBf16; 3],
    pub wq_a_fp8: Option<Fp8Dense>,
    pub wq_b_fp8: Option<Fp8Dense>,
    pub wkv_fp8: Option<Fp8Dense>,
    pub wo_a_fp8: Option<Fp8Dense>,
    pub wo_b_fp8: Option<Fp8Dense>,
    pub shared_fp8: [Option<Fp8Dense>; 3],
}

Fields§

§il: u32§ratio: usize§expert_kind: ExpertKind§wq_a: DenseBf16§wq_b: DenseBf16§wkv: DenseBf16§wo_a: DenseBf16§wo_b: DenseBf16§q_norm: CudaSlice<f32>§kv_norm: CudaSlice<f32>§attn_norm: CudaSlice<f32>§ffn_norm: CudaSlice<f32>§sink: CudaSlice<f32>§cmp: Option<CmpDev>§idx: Option<IdxDev>§hc_attn_fn: CudaSlice<f32>§hc_ffn_fn: CudaSlice<f32>§hc_attn_base: Vec<f32>§hc_attn_scale: Vec<f32>§hc_ffn_base: Vec<f32>§hc_ffn_scale: Vec<f32>§hc_attn_base_dev: CudaSlice<f32>§hc_attn_scale_dev: CudaSlice<f32>§hc_ffn_base_dev: CudaSlice<f32>§hc_ffn_scale_dev: CudaSlice<f32>§gate_bias_dev: Option<CudaSlice<f32>>§tid2eid_dev: Option<CudaSlice<i32>>

i32 cast of tid2eid, range- and distinctness-validated at LOAD (the legacy path asserts per token at route time; the device route kernel cannot).

§experts_s2_dev: CudaSlice<f32>§gate_w: CudaSlice<f32>§gate_bias: Option<Vec<f32>>§tid2eid: Option<Vec<i64>>§experts_w: CudaSlice<u8>§experts_sc: CudaSlice<u8>§experts_s2: Vec<f32>§shared_w: [DenseBf16; 3]§wq_a_fp8: Option<Fp8Dense>§wq_b_fp8: Option<Fp8Dense>§wkv_fp8: Option<Fp8Dense>§wo_a_fp8: Option<Fp8Dense>§wo_b_fp8: Option<Fp8Dense>§shared_fp8: [Option<Fp8Dense>; 3]

Auto Trait Implementations§

§

LayerDev

Struct LayerDev Copy item path

Fields§

Auto Trait Implementations§

impl Freeze for LayerDev

impl RefUnwindSafe for LayerDev

impl Send for LayerDev

impl Sync for LayerDev

impl Unpin for LayerDev

impl UnsafeUnpin for LayerDev

impl UnwindSafe for LayerDev

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

fn into(self) -> U

impl<T> Same for T

type Output = 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>

Struct LayerDev

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>,