Struct Arch 

pub struct Arch {

    pub arch: String,
    pub block_count: u64,
    pub context_length: u64,
    pub embedding_length: u64,
    pub feed_forward_length: u64,
    pub head_count: u64,
    pub head_count_kv: u64,
    pub head_dim: u64,
    pub rms_eps: f32,
    pub rope_freq_base: f32,
    pub value_length: u64,
    pub rope_dimension_count: u64,
    pub full_attention_interval: u64,
    pub nextn_predict_layers: u64,
    pub ssm_conv_kernel: u64,
    pub ssm_state_size: u64,
    pub ssm_group_count: u64,
    pub ssm_time_step_rank: u64,
    pub ssm_inner_size: u64,
}

The transformer architecture fields the hand-written model lanes read from GGUF metadata. Every field is a key under the architecture prefix and reads 0 (or the documented default) when absent.

Fields

arch: String

block_count: u64

context_length: u64

context_length — the trained context window (Qwen3.8: 262144).

embedding_length: u64

feed_forward_length: u64

head_count: u64

head_count_kv: u64

head_dim: u64

rms_eps: f32

rope_freq_base: f32

value_length: u64

attention.value_length — V head dim, which need not equal the K head dim.

rope_dimension_count: u64

rope.dimension_count — how many of each head’s dims actually rotate. Qwen3.8 rotates 64 of 256; treating this as the full head_dim silently corrupts three quarters of every query and key.

full_attention_interval: u64

full_attention_interval — every Nth block is full attention, the rest are linear (Gated DeltaNet). Qwen3.8: 4, so blocks 3, 7, 11 … are attention and the other 48 are GDN.

nextn_predict_layers: u64

nextn_predict_layers — MTP draft blocks appended to block_count. Qwen3.8 ships 1, inside the same file, so block_count (65) is one MORE than the model’s real depth (64).

ssm_conv_kernel: u64

ssm.conv_kernel — taps in GDN’s depthwise causal conv (4).

ssm_state_size: u64

ssm.state_size — the D of each head’s D×D associative memory (128).

ssm_group_count: u64

ssm.group_count — GDN key heads (16).

ssm_time_step_rank: u64

ssm.time_step_rank — GDN value heads (48). Three value heads share each key head.

ssm_inner_size: u64

ssm.inner_size — the GDN value width, time_step_rank · state_size (6144).

Implementations

impl Arch

pub fn model_layers(&self) -> u64

Blocks that are actually part of the model — block_count minus the MTP draft blocks appended to it.

pub fn is_full_attention_layer(&self, l: u64) -> bool

Is block l a full-attention block? With full_attention_interval = N, every Nth block is, counting from 1, i.e. indices N-1, 2N-1, …. A checkpoint without the key is uniformly full attention.

Trait Implementations

impl Clone for Arch

fn clone(&self) -> Arch

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Arch

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

Formats the value using the given formatter.