Struct KernelCaps 

pub struct KernelCaps {
    pub w8a8_decode: bool,
    pub w8a8_moe_decode: bool,
    pub w8a8_block_scaled_decode: bool,
    pub fp8_lm_head_batched: bool,
}

2026-09-28: Which kernels the policy may rely on. The W8A8 decode family reports its presence here (metrale_model_layers::layers::kernel_caps).

Fields

w8a8_decode: bool

2026-09-28: A per-token E4M3 activation quantizer plus E4M3xE4M3 small-M GEMMs for the dense and non-expert projections. Until they land, an FP8-declared layer served at its FP8 weights runs the W8A16 kernels: its activations are above the declared precision.

w8a8_moe_decode: bool

2026-09-28: The same for MoE experts (routed and shared), whose grouped kernels land separately (is_expert_module).

w8a8_block_scaled_decode: bool

2026-09-28: The W8A8 decode of non-expert projections whose FP8 weights are 128x128 block-scaled (HF fp8, e.g. the attention and GDN of Qwen3.6-35B-A3B-FP8), validated separately from the per-channel dense family (WeightQuantPolicy::fp8_block_scaled_decode_act).

fp8_lm_head_batched: bool

2026-09-28: An FP8 lm_head that serves every decode row count in one pass. The FP8 head today launches once per row, which collapses throughput at width (C16 88.9 tok/s against 200.0 on the NVFP4 head, C128 not finishing), so declared takes a declared FP8 head only once this is set (LmHeadChoice::PendingFp8Kernel).

Trait Implementations

impl Clone for KernelCaps

fn clone(&self) -> KernelCaps

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for KernelCaps

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for KernelCaps

fn default() -> KernelCaps

Returns the “default value” for a type.

impl PartialEq for KernelCaps

fn eq(&self, other: &KernelCaps) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Copy for KernelCaps

impl Eq for KernelCaps

impl StructuralPartialEq for KernelCaps