Struct DflashLayer

Source

pub struct DflashLayer {Show 22 fields
    pub input_layernorm: DenseWeight,
    pub post_attention_layernorm: DenseWeight,
    pub q_proj: DenseWeight,
    pub k_proj: DenseWeight,
    pub v_proj: DenseWeight,
    pub o_proj: DenseWeight,
    pub q_norm: DenseWeight,
    pub k_norm: DenseWeight,
    pub gate_proj: DenseWeight,
    pub up_proj: DenseWeight,
    pub down_proj: DenseWeight,
    pub q_proj_fp8: Option<Fp8DenseWeight>,
    pub k_proj_fp8: Option<Fp8DenseWeight>,
    pub v_proj_fp8: Option<Fp8DenseWeight>,
    pub o_proj_fp8: Option<Fp8DenseWeight>,
    pub gate_proj_fp8: Option<Fp8DenseWeight>,
    pub up_proj_fp8: Option<Fp8DenseWeight>,
    pub down_proj_fp8: Option<Fp8DenseWeight>,
    pub attention_conv_base: Option<DenseWeight>,
    pub attention_conv_proj: Option<DenseWeight>,
    pub mlp_conv_base: Option<DenseWeight>,
    pub mlp_conv_proj: Option<DenseWeight>,
}

Expand description

Per-drafter-layer Qwen3-style weights. Phase 1 is BF16-only; Phase G (2026-05-28) adds optional FP8 weight fields populated at model load when ATLAS_DFLASH_DRAFTER_FP8=1. The BF16 fields are always present (Fp8 path falls back to them for any GEMM whose Fp8 weight is None).

Fields§

§input_layernorm: DenseWeight§post_attention_layernorm: DenseWeight§q_proj: DenseWeight§k_proj: DenseWeight§v_proj: DenseWeight§o_proj: DenseWeight§q_norm: DenseWeight§k_norm: DenseWeight§gate_proj: DenseWeight§up_proj: DenseWeight§down_proj: DenseWeight§q_proj_fp8: Option<Fp8DenseWeight>§k_proj_fp8: Option<Fp8DenseWeight>§v_proj_fp8: Option<Fp8DenseWeight>§o_proj_fp8: Option<Fp8DenseWeight>§gate_proj_fp8: Option<Fp8DenseWeight>§up_proj_fp8: Option<Fp8DenseWeight>§down_proj_fp8: Option<Fp8DenseWeight>§attention_conv_base: Option<DenseWeight>

attention_conv.base_kernel [2 applications, kernel_size, hidden].

§attention_conv_proj: Option<DenseWeight>

attention_conv.kernel_projection.weight [2*kernel*groups, hidden].

§mlp_conv_base: Option<DenseWeight>

mlp_conv.base_kernel, same shape as attention_conv_base.

§mlp_conv_proj: Option<DenseWeight>

mlp_conv.kernel_projection.weight, same shape as attention_conv_proj.

Auto Trait Implementations§

§

DflashLayer

Struct DflashLayer Copy item path

Fields§

Auto Trait Implementations§

impl Freeze for DflashLayer

impl RefUnwindSafe for DflashLayer

impl Send for DflashLayer

impl Sync for DflashLayer

impl Unpin for DflashLayer

impl UnwindSafe for DflashLayer

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 DflashLayer

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>,