pub struct Glm5NextKdaWeights {Show 13 fields
pub q_proj: DenseWeight,
pub k_proj: DenseWeight,
pub v_proj: DenseWeight,
pub conv: DenseWeight,
pub f_a: DenseWeight,
pub f_b: DenseWeight,
pub dt_bias: DevicePtr,
pub a_log: DevicePtr,
pub b_proj: DenseWeight,
pub g_a: DenseWeight,
pub g_b: DenseWeight,
pub o_norm: DenseWeight,
pub o_proj: DenseWeight,
}Expand description
One KDA block’s device weights. Torch Linear layout [out, in], BF16, except the two F32
gate parameters. There is no Z tensor — the output gate is low-rank g_a/g_b.
Fields§
§q_proj: DenseWeight§k_proj: DenseWeight§v_proj: DenseWeight§conv: DenseWeight[conv_dim, conv_kernel] BF16 = concat([q, k, v]).squeeze(1).
f_a: DenseWeight§f_b: DenseWeight§dt_bias: DevicePtr[heads * head_dim] F32 — per channel.
a_log: DevicePtr[heads] F32 — per head. The asymmetry with dt_bias is the highest-risk line.
b_proj: DenseWeight§g_a: DenseWeight§g_b: DenseWeight§o_norm: DenseWeight[head_dim] BF16.
o_proj: DenseWeightAuto Trait Implementations§
impl Freeze for Glm5NextKdaWeights
impl RefUnwindSafe for Glm5NextKdaWeights
impl Send for Glm5NextKdaWeights
impl Sync for Glm5NextKdaWeights
impl Unpin for Glm5NextKdaWeights
impl UnwindSafe for Glm5NextKdaWeights
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