pub struct W4a16BatchmTiers { /* private fields */ }Expand description
2026-09-25: Resolved handles for the narrow w4a16_gemv_batch{M} family.
A zero handle means “this target did not load that tier”; consumers check
.0 != 0, and Self::kernel returns a zero handle rather than panicking
when nothing in the family covers m.
Implementations§
Source§impl W4a16BatchmTiers
impl W4a16BatchmTiers
Sourcepub fn resolve(gpu: &dyn GpuBackend) -> Self
pub fn resolve(gpu: &dyn GpuBackend) -> Self
2026-09-25: Resolve every tier in the family. A missing kernel is a zero
handle, and select_tier skips it.
Sourcepub fn edge(&self, w: &QuantizedWeight) -> u32
pub fn edge(&self, w: &QuantizedWeight) -> u32
2026-09-28: The row edge for w, from its activation stamp (edge_of).
Sourcepub fn declares_a4(&self, w: &QuantizedWeight) -> bool
pub fn declares_a4(&self, w: &QuantizedWeight) -> bool
2026-09-28: w declares FP4 activations and this backend has the W4A4 kernels. The
published tier decides what that means for a dense FFN’s narrow steps
(WeightQuantTier::ffn_small_batch_w4a4).
Sourcepub fn ffn_edge(&self, w: &QuantizedWeight) -> u32
pub fn ffn_edge(&self, w: &QuantizedWeight) -> u32
2026-09-28: Self::edge capped at w4a4_proj::W4A4_MAX_M: the dense FFN’s
narrow arms stop at 32 rows, and 33..=64 go to forward_prefill.
Sourcepub fn kernel_for(&self, m: u32, w: &QuantizedWeight) -> KernelHandle
pub fn kernel_for(&self, m: u32, w: &QuantizedWeight) -> KernelHandle
2026-09-28: Self::kernel for a projection over w, within Self::edge.
Sourcepub fn kernel(&self, m: u32) -> KernelHandle
pub fn kernel(&self, m: u32) -> KernelHandle
2026-09-25: Narrowest resolved tier covering m rows, or KernelHandle(0) when
this family cannot serve m, for a projection without a policy stamp (the lm_head,
the MTP head): within the edge of an unstamped weight (2026-09-28: the W4A16 edge,
or under nvfp4 with --w4a4-downcast the lever’s edge, as before the tiers).
Trait Implementations§
Source§impl Clone for W4a16BatchmTiers
impl Clone for W4a16BatchmTiers
Source§fn clone(&self) -> W4a16BatchmTiers
fn clone(&self) -> W4a16BatchmTiers
1.0.0 · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
source. Read moreSource§impl Debug for W4a16BatchmTiers
impl Debug for W4a16BatchmTiers
Source§impl Default for W4a16BatchmTiers
2026-09-25: The “no NVFP4 kernels” state: an all-zero table, for which kernel
returns a zero handle at every m.
impl Default for W4a16BatchmTiers
2026-09-25: The “no NVFP4 kernels” state: an all-zero table, for which kernel
returns a zero handle at every m.