pub struct Qwen35ForwardConfig {Show 16 fields
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,
}Expand description
2026-09-25: Dimensions of a Qwen3.5 checkpoint, read by both layer forwards.
Fields§
§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: u322026-09-25: RoPE rotates only the first rotary_dim elements of each head;
the rest pass through unchanged (rope_apply.metal).
num_k_heads_lin: u32§num_v_heads_lin: u32§k_head_dim_lin: u32§v_head_dim_lin: u32§conv_kernel_size: u32Implementations§
Source§impl Qwen35ForwardConfig
impl Qwen35ForwardConfig
Sourcepub const fn qwen3_5_4b_mlx_int8() -> Self
pub const fn qwen3_5_4b_mlx_int8() -> Self
2026-09-25: The dimensions the metal_qwen35_inference example runs with; its
default model directory is Qwen3.5-4B-MLX-8bit.
Sourcepub const fn q_total(&self) -> u32
pub const fn q_total(&self) -> u32
2026-09-25: q_proj output width. Qwen3.5 packs the attention output gate
into the Q projection: each head’s row is [Q_h | gate_h], which
qwen35_qkv_split separates before the norm.
Sourcepub const fn q_only(&self) -> u32
pub const fn q_only(&self) -> u32
2026-09-25: Q width after the gate is split off: half of Self::q_total.
pub const fn kv_dim(&self) -> u32
pub const fn z_dim_lin(&self) -> u32
pub const fn qkv_total_lin(&self) -> u32
Sourcepub const fn num_state_heads(&self) -> u32
pub const fn num_state_heads(&self) -> u32
2026-09-25: The number of GDN heads that gate, beta, dt_bias and A_log are indexed by.
Trait Implementations§
Source§impl Clone for Qwen35ForwardConfig
impl Clone for Qwen35ForwardConfig
Source§fn clone(&self) -> Qwen35ForwardConfig
fn clone(&self) -> Qwen35ForwardConfig
Returns a duplicate of the value. Read more
1.0.0 · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
Performs copy-assignment from
source. Read moreSource§impl Debug for Qwen35ForwardConfig
impl Debug for Qwen35ForwardConfig
impl Copy for Qwen35ForwardConfig
Auto Trait Implementations§
impl Freeze for Qwen35ForwardConfig
impl RefUnwindSafe for Qwen35ForwardConfig
impl Send for Qwen35ForwardConfig
impl Sync for Qwen35ForwardConfig
impl Unpin for Qwen35ForwardConfig
impl UnwindSafe for Qwen35ForwardConfig
Blanket Implementations§
Source§impl<T> BorrowMut<T> for Twhere
T: ?Sized,
impl<T> BorrowMut<T> for Twhere
T: ?Sized,
Source§fn borrow_mut(&mut self) -> &mut T
fn borrow_mut(&mut self) -> &mut T
Mutably borrows from an owned value. Read more