pub enum Readback<'a> {
Argmax,
ArgmaxMasked {
masks: Vec<RowMask>,
},
HostLogits {
into: &'a mut Vec<u8>,
},
}Expand description
2026-09-25: How the logits of a decode step come back to the host.
Variants§
Argmax
2026-09-25: One argmax id per row, computed on the device.
ArgmaxMasked
2026-09-25: The device argmax with the per-row two-id mask applied exactly as
the host re-pick would (see argmax_feed.cu); the ids also land
in the device feed cells for a following fed step.
HostLogits
2026-09-25: The whole logits block, copied into into (resized by the router to
rows * vocab * elem_bytes; the core samples over it).
Trait Implementations§
Auto Trait Implementations§
impl<'a> Freeze for Readback<'a>
impl<'a> RefUnwindSafe for Readback<'a>
impl<'a> Send for Readback<'a>
impl<'a> Sync for Readback<'a>
impl<'a> Unpin for Readback<'a>
impl<'a> !UnwindSafe for Readback<'a>
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