Struct Qwen30bCfg 

pub struct Qwen30bCfg {
    pub d: usize,
    pub dff: usize,
    pub ne: usize,
    pub k: usize,
    pub layers: usize,
    pub h: usize,
    pub kv: usize,
    pub hd: usize,
    pub vocab: usize,
    pub rope_theta: f64,
    pub rms_eps: f64,
    pub attn_split: usize,
}

The Qwen3-30B-A3B architecture (n_layer is overridable to keep local emit/compile fast; the per-layer shape is identical to the real model).

Fields

d: usize

dff: usize

ne: usize

k: usize

layers: usize

h: usize

kv: usize

hd: usize

vocab: usize

rope_theta: f64

RoPE base frequency (<arch>.rope.freq_base in the GGUF). Read it from the checkpoint — it was a hardcoded 1e6 in the emitted CUDA until Qwen3.8-27B arrived with 1e7, which would have silently mis-rotated every query and key while still producing plausible-looking tokens.

rms_eps: f64

RMSNorm epsilon (<arch>.attention.layer_norm_rms_epsilon). Also formerly hardcoded.

attn_split: usize

Split-K factor for the flash-attention lane. Per-chip tuning knob, not a model property: measured best at 4 on the A100 and 8 as the portable default.

Implementations

impl Qwen30bCfg

pub fn real() -> Self