pub struct Glm5NextDenseMlpWeights {
pub gate_proj: DevicePtr,
pub up_proj: DevicePtr,
pub down_proj: DevicePtr,
}Expand description
A dense SwiGLU MLP in BF16 — layers 0..first_k_dense_replace, and the shared expert of
every routed layer. Same shape, same kernels, different widths.
TP: gate_proj/up_proj are column-parallel (row-sliced, since each is stored [inter, hidden]), down_proj is row-parallel (column-sliced on [hidden, inter]). The output
is therefore a partial sum whenever tp_world_size > 1.
Fields§
§gate_proj: DevicePtr[local_inter, hidden] BF16.
up_proj: DevicePtr[local_inter, hidden] BF16.
down_proj: DevicePtr[hidden, local_inter] BF16 — row-parallel.
Trait Implementations§
Source§impl Clone for Glm5NextDenseMlpWeights
impl Clone for Glm5NextDenseMlpWeights
Source§fn clone(&self) -> Glm5NextDenseMlpWeights
fn clone(&self) -> Glm5NextDenseMlpWeights
Returns a duplicate of the value. Read more
1.0.0 · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
Performs copy-assignment from
source. Read moreSource§impl Debug for Glm5NextDenseMlpWeights
impl Debug for Glm5NextDenseMlpWeights
impl Copy for Glm5NextDenseMlpWeights
Auto Trait Implementations§
impl Freeze for Glm5NextDenseMlpWeights
impl RefUnwindSafe for Glm5NextDenseMlpWeights
impl Send for Glm5NextDenseMlpWeights
impl Sync for Glm5NextDenseMlpWeights
impl Unpin for Glm5NextDenseMlpWeights
impl UnwindSafe for Glm5NextDenseMlpWeights
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