Enum OpClass 

pub enum OpClass {
    Ffn = 0,
    Matvec = 1,
    Matmat = 2,
    Batch = 3,
    MatmatWide = 4,
    MatvecHead = 5,
    GemmNt = 6,
}

GPU-eligible op classes, each with an independent probe.

Variants

Ffn = 0

Whole FFN chain in one submission (dense / MoE block).

Matvec = 1

Large hybrid CPU∥GPU matvec (lm_head class).

Matmat = 2

Prefill GEMM (matmat).

Batch = 3

Batched matvecs of one input (QKV).

MatmatWide = 4

Prefill GEMM at image-diffusion widths (b ≥ 128). Probed apart from Matmat: one imagegen process runs BOTH populations (prompt encode b≈40 where the GPU wins big, DiT b≥256 where the CPU AMX arm is competitive) — a single shared verdict locks the wrong arm for whichever population samples second.

MatvecHead = 5

The lm_head itself, apart from the merely-large matvecs. Same reasoning as MatmatWide, and DeepSeek-V4 is where it bit: its attention projections are 37M weights and its head is 529M, so the projections’ verdict — CPU, honestly measured at 0.19 ms — decided for a matvec fourteen times their size that took 11 ms a token on the host.

GemmNt = 6

The blocked f32 GEMM (fcd_ops::gemm_nt): attention’s QKᵀ and AV, and the VAE decoders’ projections. It used to take every job over 4 M MACs on sight, with no CPU arm to lose to — which on the MiniMax-H3 video decoder was three times SLOWER than the host it displaced. Its population is per-head slices, nothing like the weight GEMMs above, so it probes on its own.

Trait Implementations

impl Clone for OpClass

fn clone(&self) -> OpClass

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for OpClass