pub struct HcSiteWeights {
pub hc_fn: DevicePtr,
pub hc_base: DevicePtr,
pub hc_scale: DevicePtr,
pub lowrank: Option<HcLowRank>,
}Expand description
Per-block Manifold-Constrained Hyper-Connection (mHC) parameters for one
site (attention or FFN). All buffers are float32 device pointers, matching
the checkpoint dtype. See ops::hc_pre / ops::hc_post.
Fields§
§hc_fn: DevicePtrMix projection fn: [mix_hc, hc_mult*hidden] f32, where
mix_hc = (2 + hc_mult) * hc_mult.
hc_base: DevicePtrMix bias base: [mix_hc] f32.
hc_scale: DevicePtrMix scale: [3] f32 (pre / post / comb scalars).
lowrank: Option<HcLowRank>Qwen low-rank variant. Some => dispatch the low-rank kernels and
IGNORE hc_fn/hc_base/hc_scale (which are NULL in that case).
Auto Trait Implementations§
impl Freeze for HcSiteWeights
impl RefUnwindSafe for HcSiteWeights
impl Send for HcSiteWeights
impl Sync for HcSiteWeights
impl Unpin for HcSiteWeights
impl UnwindSafe for HcSiteWeights
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