Struct GgufHyperparams 

pub struct GgufHyperparams {

    pub n_layer: u32,
    pub n_embd: u32,
    pub n_head: u32,
    pub n_kv_head: u32,
    pub rope_freq_base: f32,
    pub rope_scale: f32,
    pub head_dim: u32,
    pub head_dim_swa: u32,
    pub sliding_window: u32,
    pub shared_kv_layers: u32,
    pub logit_softcap: f32,
    pub architecture: u32,
    pub arch_flags: u32,
}

Architecture hyper-parameters parsed from the GGUF KV section.

Fields

n_layer: u32

n_embd: u32

n_head: u32

n_kv_head: u32

Grouped-query KV heads; 0 means MHA (n_kv_head == n_head).

rope_freq_base: f32

llama.rope.freq_base (FLOAT32 in GGUF); 0 → DEFAULT_ROPE_FREQ_BASE.

rope_scale: f32

Linear RoPE scale from llama.rope.scale_linear / llama.rope.scaling.factor; 0 → 1.0.

head_dim: u32

Explicit head dim (*.attention.key_length); 0 → derive n_embd / n_head.

head_dim_swa: u32

SWA / local-attention head dim (*.attention.key_length_swa); 0 → same as head_dim.

sliding_window: u32

Sliding-window size in tokens; 0 → full context attention only.

shared_kv_layers: u32

Last N layers share KV from the last non-shared layer of the same type (Gemma 4).

logit_softcap: f32

Final logit softcapping (Gemma 2+); 0 → disabled.

architecture: u32

[ARCH_*] id from general.architecture (and tensor-feature refinement).

arch_flags: u32

[ARCH_FLAG_*] bitmask.

Implementations

impl GgufHyperparams

pub fn effective_rope_freq_base(&self) -> f32

pub fn effective_rope_scale(&self) -> f32

Effective position divisor for RoPE (scaled_pos = pos / scale).

pub fn effective_head_dim(&self) -> u32

Nominal head_dim (n_embd / n_head if head_dim == 0).

pub fn head_dim(&self) -> u32

pub fn effective_n_kv_head(&self) -> u32

pub fn q_heads_per_kv(&self) -> u32

pub fn gqa_ratio(&self) -> u32

pub fn head_dim_swa_or(&self) -> u32

pub fn architecture_name(&self) -> &'static str

Human-readable architecture name for logs / errors.

pub fn decode_supported(&self) -> Result<(), String>

Whether the native decode path can run this architecture coherently.

Gemma 4 (E2B/E4B) requires PLE, dual-RoPE SWA/global head dims, QK-norm, post-norms, variable FFN width, and shared KV — none of which the Llama-shaped decode path implements. Running it produces multilingual garbage (measured 2026-07-09 on gemma-4-E2B-it-Q4_K_M). Override with QUALIA_LLM_FORCE_UNSUPPORTED_ARCH=1 only for bring-up.

Trait Implementations

impl Clone for GgufHyperparams

fn clone(&self) -> GgufHyperparams

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GgufHyperparams

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

Formats the value using the given formatter.

impl Default for GgufHyperparams

fn default() -> GgufHyperparams

Returns the “default value” for a type.

impl PartialEq for GgufHyperparams

fn eq(&self, other: &GgufHyperparams) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Copy for GgufHyperparams

impl StructuralPartialEq for GgufHyperparams