Struct Qwen35ForwardConfig 

pub struct Qwen35ForwardConfig {

    pub hidden: u32,
    pub intermediate: u32,
    pub num_layers: u32,
    pub vocab: u32,
    pub group_size: u32,
    pub rms_eps: f32,
    pub num_heads: u32,
    pub num_kv_heads: u32,
    pub head_dim: u32,
    pub rope_theta: f32,
    pub rotary_dim: u32,
    pub num_k_heads_lin: u32,
    pub num_v_heads_lin: u32,
    pub k_head_dim_lin: u32,
    pub v_head_dim_lin: u32,
    pub conv_kernel_size: u32,
}

Compile-time-fixed dimensions for a Qwen3.5 checkpoint. Populate from the model’s config.json (text_config) at startup.

Fields

hidden: u32

intermediate: u32

num_layers: u32

vocab: u32

group_size: u32

rms_eps: f32

num_heads: u32

num_kv_heads: u32

head_dim: u32

rope_theta: f32

rotary_dim: u32

head_dim * partial_rotary_factor — Qwen3.5-VL rotates only the first rotary_dim of each head (=64 of 256 for the 4B checkpoint with partial_rotary_factor = 0.25).

num_k_heads_lin: u32

num_v_heads_lin: u32

k_head_dim_lin: u32

v_head_dim_lin: u32

conv_kernel_size: u32

Implementations

impl Qwen35ForwardConfig

pub const fn qwen3_5_4b_mlx_int8() -> Self

Hardcoded constants for mlx-community/Qwen3.5-4B-MLX-8bit. Matches the Metal example’s dims.rs exactly so the extracted forward path is byte-equivalent to the inlined version.

pub const fn q_total(&self) -> u32

Q_TOTAL = num_heads * head_dim * 2 — Qwen3.5 packs the attention output gate into the same projection as Q, so the q_proj produces a [num_heads, head_dim * 2] interleaved tensor that needs a deinterleave step before normalisation.

pub const fn q_only(&self) -> u32

Q_ONLY = num_heads * head_dim — half of Q_TOTAL, the post-deinterleave Q size.

pub const fn kv_dim(&self) -> u32

KV_DIM = num_kv_heads * head_dim.

pub const fn z_dim_lin(&self) -> u32

Z_DIM_LIN = num_v_heads_lin * v_head_dim_lin.

pub const fn qkv_total_lin(&self) -> u32

QKV_TOTAL_LIN = (num_k_heads_lin + num_k_heads_lin) * k_head_dim_lin + num_v_heads_lin * v_head_dim_lin.

pub const fn num_state_heads(&self) -> u32

NUM_STATE_HEADS = num_v_heads_lin — the number of GDN heads the gate / beta / dt_bias / A_log vectors all run over.

Trait Implementations

impl Clone for Qwen35ForwardConfig

fn clone(&self) -> Qwen35ForwardConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Qwen35ForwardConfig

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

Formats the value using the given formatter.

impl Copy for Qwen35ForwardConfig