pub struct GlmVitBlock {Show 14 fields
pub norm1_w: DevicePtr,
pub qkv_w: DevicePtr,
pub qkv_b: DevicePtr,
pub q_norm_w: DevicePtr,
pub k_norm_w: DevicePtr,
pub proj_w: DevicePtr,
pub proj_b: DevicePtr,
pub norm2_w: DevicePtr,
pub gate_w: DevicePtr,
pub gate_b: DevicePtr,
pub up_w: DevicePtr,
pub up_b: DevicePtr,
pub down_w: DevicePtr,
pub down_b: DevicePtr,
}Expand description
One ViT block’s weights. Every tensor is BF16 on disk and stays BF16 — the
tower is NOT quantised in nvidia/GLM-5.3-Flash-NVFP4 (347/347 tensors are
BF16; only the text/MoE stack is NVFP4).
Fields§
§norm1_w: DevicePtr§qkv_w: DevicePtr§qkv_b: DevicePtr§q_norm_w: DevicePtrPer-head RMSNorm on Q, [head_dim]. Applied after the QKV split and
BEFORE the rotary transform — the order is the reference’s.
k_norm_w: DevicePtr§proj_w: DevicePtr§proj_b: DevicePtr§norm2_w: DevicePtr§gate_w: DevicePtr§gate_b: DevicePtr§up_w: DevicePtr§up_b: DevicePtr§down_w: DevicePtr§down_b: DevicePtrAuto Trait Implementations§
impl Freeze for GlmVitBlock
impl RefUnwindSafe for GlmVitBlock
impl Send for GlmVitBlock
impl Sync for GlmVitBlock
impl Unpin for GlmVitBlock
impl UnwindSafe for GlmVitBlock
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