pub enum HeadMode {
All,
LastRow,
Skip,
}Expand description
What the forward’s exit computes (chunked long-context prefill skips the head: the [t, vocab] logits block of a big chunk is gigabytes and reads/writes no state).
Variants§
All
Exit mixer + lm_head on every row ([t, vocab] logits) — the historical shape.
LastRow
Exit mixer on the chunk, lm_head on the LAST row only ([vocab] logits).
Skip
No exit mixer, no lm_head, empty return (mid-prefill chunks).
Trait Implementations§
impl Copy for HeadMode
impl Eq for HeadMode
impl StructuralPartialEq for HeadMode
Auto Trait Implementations§
impl Freeze for HeadMode
impl RefUnwindSafe for HeadMode
impl Send for HeadMode
impl Sync for HeadMode
impl Unpin for HeadMode
impl UnsafeUnpin for HeadMode
impl UnwindSafe for HeadMode
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