Skip to main content

ArchitectureConfig

Trait ArchitectureConfig 

Source
pub trait ArchitectureConfig: Send + Sync {
Show 14 methods // Required methods fn num_attention_heads(&self) -> usize; fn num_key_value_heads(&self) -> usize; fn hidden_size(&self) -> usize; fn intermediate_size(&self) -> usize; fn vocab_size(&self) -> usize; fn max_position_embeddings(&self) -> usize; fn rope_theta(&self) -> Option<f32>; fn rope_dim(&self) -> Option<usize>; fn layer_norm_eps(&self) -> f32; fn use_rms_norm(&self) -> bool; // Provided methods fn head_dim(&self) -> usize { ... } fn use_parallel_residual(&self) -> bool { ... } fn attention_type(&self) -> AttentionType { ... } fn default_dtype(&self) -> Kind { ... }
}
Expand description

Core configuration for model architectures

Required Methods§

Source

fn num_attention_heads(&self) -> usize

Number of attention heads for queries

Source

fn num_key_value_heads(&self) -> usize

Number of key-value heads (for MQA/GQA)

Source

fn hidden_size(&self) -> usize

Hidden dimension size

Source

fn intermediate_size(&self) -> usize

Intermediate/FFN dimension size

Source

fn vocab_size(&self) -> usize

Vocabulary size

Source

fn max_position_embeddings(&self) -> usize

Maximum position embeddings

Source

fn rope_theta(&self) -> Option<f32>

RoPE base frequency (if applicable)

Source

fn rope_dim(&self) -> Option<usize>

RoPE dimension count (if different from head_dim)

Source

fn layer_norm_eps(&self) -> f32

Layer normalization epsilon

Source

fn use_rms_norm(&self) -> bool

Whether to use RMSNorm instead of LayerNorm

Provided Methods§

Source

fn head_dim(&self) -> usize

Head dimension (hidden_size / num_attention_heads)

Source

fn use_parallel_residual(&self) -> bool

Whether this architecture uses parallel attention and FFN

Source

fn attention_type(&self) -> AttentionType

Get attention type (MHA, MQA, GQA)

Source

fn default_dtype(&self) -> Kind

Default dtype for this architecture

Implementors§