Struct ModelConfig 

#[non_exhaustive]
pub struct ModelConfig {

    pub model_type: Option<String>,
    pub num_hidden_layers: Option<u32>,
    pub num_attention_heads: Option<u32>,
    pub num_key_value_heads: Option<u32>,
    pub head_dim: Option<u32>,
    pub hidden_size: Option<u32>,
    pub torch_dtype: Option<String>,
    pub sliding_window: Option<u32>,
    pub sliding_window_pattern: Option<u32>,
    pub use_sliding_window: Option<bool>,
    pub kv_lora_rank: Option<u32>,
    pub qk_rope_head_dim: Option<u32>,
    pub layer_types: Option<Vec<String>>,
    pub hybrid_override_pattern: Option<String>,
    pub attn_layer_indices: Option<Vec<u32>>,
    pub full_attention_interval: Option<u32>,
    pub mamba_n_heads: Option<u32>,
    pub mamba_d_head: Option<u32>,
    pub mamba_d_state: Option<u32>,
    pub mamba_d_conv: Option<u32>,
    pub mamba_n_groups: Option<u32>,
}

Attention- and cache-relevant fields parsed from a model’s config.json.

Every field is Option because configs vary across architecture families; the KV-cache estimator decides which combinations are computable and which fall back to an “unavailable” verdict. Legacy n_layer / n_head / n_head_kv spellings are absorbed by serde aliases on the private deserialization struct. Drives KV-cache budgeting for inspect --check-gpu --context.

Fields (Non-exhaustive)

Non-exhaustive structs could have additional fields added in future. Therefore, non-exhaustive structs cannot be constructed in external crates using the traditional Struct { .. } syntax; cannot be matched against without a wildcard ..; and struct update syntax will not work.

model_type: Option<String>

Architecture tag (e.g. "llama", "qwen3", "gemma2", "deepseek_v2").

num_hidden_layers: Option<u32>

Number of transformer layers (num_hidden_layers / n_layer).

num_attention_heads: Option<u32>

Number of query attention heads (num_attention_heads / n_head).

num_key_value_heads: Option<u32>

Number of key/value heads for GQA (num_key_value_heads / num_kv_heads / n_head_kv). Absent ⇒ MHA (equals num_attention_heads).

head_dim: Option<u32>

Explicit per-head dimension when stated (Gemma = 256, Qwen3 = 128). Absent ⇒ derived as hidden_size / num_attention_heads.

hidden_size: Option<u32>

Model hidden size, used to derive head_dim when it is not explicit.

torch_dtype: Option<String>

Activation dtype spelling ("bfloat16", "float16", …); sizes the KV-cache element via torch_dtype_bytes.

sliding_window: Option<u32>

Sliding-window span in tokens when the model uses windowed attention. null / absent ⇒ full attention.

sliding_window_pattern: Option<u32>

Global-attention period for mixed local/global layouts (Gemma-3: every N-th layer is a full-attention layer).

use_sliding_window: Option<bool>

Explicit on/off switch for sliding-window attention — Qwen2/3 ship a sliding_window value but disable it with false.

kv_lora_rank: Option<u32>

MLA latent-KV rank (DeepSeek). Presence marks multi-head latent attention, where the naive KV formula does not apply.

qk_rope_head_dim: Option<u32>

MLA decoupled-RoPE key dimension (DeepSeek); part of the latent-KV size used by the documented MLA estimate.

layer_types: Option<Vec<String>>

Per-layer kind tags for hybrid models ("attention" / "mamba" / "linear_attention" / …). Primary hybrid-layout signal (Granite-4).

hybrid_override_pattern: Option<String>

Nemotron-H layer-layout string ("M-M-M-M*-…": * = attention, M = Mamba, - = FFN-only). Alternative hybrid-layout signal.

attn_layer_indices: Option<Vec<u32>>

Explicit indices of the attention layers (Bamba). Alternative hybrid-layout signal; the remaining layers are recurrent.

full_attention_interval: Option<u32>

Period of full-attention layers — every N-th layer is attention, the rest recurrent (Qwen3-Next). Alternative hybrid-layout signal.

mamba_n_heads: Option<u32>

Mamba2 SSM head count (mamba_n_heads / mamba_num_heads).

mamba_d_head: Option<u32>

Mamba2 SSM per-head dimension (mamba_d_head / mamba_head_dim).

mamba_d_state: Option<u32>

Mamba2 SSM state size (mamba_d_state / ssm_state_size).

mamba_d_conv: Option<u32>

Mamba2 causal-convolution width (mamba_d_conv / conv_kernel).

mamba_n_groups: Option<u32>

Mamba2 group count for the convolution (mamba_n_groups / n_groups).

Trait Implementations

impl Clone for ModelConfig

fn clone(&self) -> ModelConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ModelConfig

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

Formats the value using the given formatter.

impl Default for ModelConfig

fn default() -> ModelConfig

Returns the “default value” for a type.

impl Serialize for ModelConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.