Struct ModelCapabilities 

pub struct ModelCapabilities {
    pub has_ssm_layers: bool,
    pub has_attention_layers: bool,
    pub has_moe_layers: bool,
    pub supports_thinking: bool,
    pub supports_vision: bool,
    pub has_mtp: bool,
    pub ssm_architecture: SsmArchitecture,
    pub attention_type: AttentionType,
    pub has_nested_config: bool,
}

Feature flags derived from model config at parse time.

These replace scattered is_nemotron_h(), is_qwen35() checks with model-agnostic predicates. New models get capabilities automatically from their config — no code changes needed.

Fields

has_ssm_layers: bool

Model has SSM/recurrent layers (GDN or Mamba-2).

has_attention_layers: bool

Model has full attention layers.

has_moe_layers: bool

Model has MoE (Mixture of Experts) layers.

supports_thinking: bool

Model supports <think> reasoning tokens. Derived from tokenizer vocabulary, not model type name.

supports_vision: bool

Model has vision encoder (multimodal).

has_mtp: bool

Model has MTP (Multi-Token Prediction) draft head.

ssm_architecture: SsmArchitecture

SSM architecture family (determines state layout).

attention_type: AttentionType

Attention architecture (standard GQA vs MLA compressed latent).

has_nested_config: bool

Model wraps language model config in a nested field (e.g., text_config).

Implementations

impl ModelCapabilities

pub fn from_config(config: &ModelConfig) -> Self

Derive capabilities from a ModelConfig.

This is the SSOT for model feature detection. When adding a new model, ensure its config fields populate the right capabilities here — then all downstream code works automatically.

Trait Implementations

impl Clone for ModelCapabilities

fn clone(&self) -> ModelCapabilities

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ModelCapabilities

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

Formats the value using the given formatter.