pub enum StepPlan<'a> {
Decode {
tokens: Vec<u32>,
ctx: DecodeCtxCommit,
readback: Readback<'a>,
},
DecodeFed {
sources: Vec<FeedSource>,
masks: Vec<RowMask>,
ctx: DecodeCtxCommit,
},
}Expand description
2026-09-25: One device step. The sequence rows it advances travel beside it as
&mut S.
Variants§
Decode
2026-09-25: Batched decode of one token per row, the context commit, then the readback.
Fields
§
ctx: DecodeCtxCommitDecodeFed
2026-09-25: A decode step launched while the previous one is still in flight:
each row’s input is resolved on the device from sources, the
readback is the masked feed argmax, and the host bookkeeping the
plain step performs at launch (tokens.push, seq_len += 1) is
deferred to the core, which applies it once the previous step is
committed. The context commit runs between the forward and the
readback, as it does for the plain step.
Trait Implementations§
Auto Trait Implementations§
impl<'a> Freeze for StepPlan<'a>
impl<'a> RefUnwindSafe for StepPlan<'a>
impl<'a> Send for StepPlan<'a>
impl<'a> Sync for StepPlan<'a>
impl<'a> Unpin for StepPlan<'a>
impl<'a> !UnwindSafe for StepPlan<'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