pub struct GraphMoeDetachedShared<'a> {
pub gate: GraphW<'a>,
pub up: GraphW<'a>,
pub down: GraphW<'a>,
pub width: usize,
}Expand description
A SwiGLU shared MoE expert whose intermediate width or packed layout does not match the routed expert bank. Most MoE checkpoints put the shared expert in the same packed bank as the routed experts, but Flint keeps its 1024-wide shared expert beside 512-wide routed experts. Keeping this descriptor separate lets the graph run the existing routed-bank kernels and add the shared expert through the ordinary dense q4tp path, rather than silently falling all the way back to per-op execution.
Fields§
§gate: GraphW<'a>§up: GraphW<'a>§down: GraphW<'a>§width: usizeAuto Trait Implementations§
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