pub struct Dflash2Head {
pub attn_conv: Vec<Dflash2Conv>,
pub mlp_conv: Vec<Dflash2Conv>,
pub hidden_proj: GpuTensor,
pub pred_codebook: Vec<u8>,
pub succ_codebook: Vec<u8>,
pub rank: usize,
pub top_k: usize,
pub conv_k: usize,
pub group_size: usize,
pub vocab: usize,
}Fields§
§attn_conv: Vec<Dflash2Conv>§mlp_conv: Vec<Dflash2Conv>candidate_selector.hidden_projection.weight [rank, hidden].
pred_codebook: Vec<u8>Codebooks [V, rank] raw bf16, HOST-resident: the walk gathers ~1+16 rows per
draft slot (~70KB/round) — host math beside the round’s existing chain dtoh,
no device residency for 2x127MB tables. Checkpoint quirk: stored WITHOUT the
.weight suffix (reference from_pretrained installs a key_mapping).
succ_codebook: Vec<u8>§rank: usize§top_k: usize§conv_k: usize§group_size: usize§vocab: usizeImplementations§
Auto Trait Implementations§
impl Freeze for Dflash2Head
impl RefUnwindSafe for Dflash2Head
impl Send for Dflash2Head
impl Sync for Dflash2Head
impl Unpin for Dflash2Head
impl UnsafeUnpin for Dflash2Head
impl UnwindSafe for Dflash2Head
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