pub struct DflashLayerWeights {Show 15 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 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 raw weights (BF16). Same shape across all 8 layers.
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§attention_conv_base: Option<DenseWeight>attention_conv.base_kernel [2, kernel_size, hidden] BF16 —
static tap weights (index 0 = prepare/pre-sublayer, 1 = finish/post).
attention_conv_proj: Option<DenseWeight>attention_conv.kernel_projection.weight
[2 * kernel_size * groups, hidden] BF16 — dynamic tap generator.
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§
impl Freeze for DflashLayerWeights
impl RefUnwindSafe for DflashLayerWeights
impl Send for DflashLayerWeights
impl Sync for DflashLayerWeights
impl Unpin for DflashLayerWeights
impl UnwindSafe for DflashLayerWeights
Blanket Implementations§
Source§impl<T> BorrowMut<T> for Twhere
T: ?Sized,
impl<T> BorrowMut<T> for Twhere
T: ?Sized,
Source§fn borrow_mut(&mut self) -> &mut T
fn borrow_mut(&mut self) -> &mut T
Mutably borrows from an owned value. Read more