pub struct Step35Aux {
pub rope_freqs: Option<Vec<(usize, CudaSlice<f32>)>>,
}Expand description
step35 model-level auxiliaries. Deliberately NOT folded into GemmaAux: every gemma4 path
does gemma4_aux.as_ref().unwrap() and would then also fire on a step35 model.
Fields§
§rope_freqs: Option<Vec<(usize, CudaSlice<f32>)>>rope_freqs.weight [n_rot_full/2] llama3-style freq factors. Upstream applies them to
FULL-attention layers ONLY (rope_factors = is_swa ? nullptr : get_rope_factors(...),
step35.cpp:246) — the SWA layers pass a null factor pointer. Step-3.7-Flash ships [64] F32.
Keep one copy on every PP device: this model-level tensor is read by full-attention
layers on both sides of the cut, and a primary-only copy would be a mapped peer read.
Implementations§
Auto Trait Implementations§
impl Freeze for Step35Aux
impl RefUnwindSafe for Step35Aux
impl Send for Step35Aux
impl Sync for Step35Aux
impl Unpin for Step35Aux
impl UnsafeUnpin for Step35Aux
impl UnwindSafe for Step35Aux
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