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) -> DType { ... }
}Expand description
Core configuration for model architectures
Required Methods§
Sourcefn num_attention_heads(&self) -> usize
fn num_attention_heads(&self) -> usize
Number of attention heads for queries
Sourcefn num_key_value_heads(&self) -> usize
fn num_key_value_heads(&self) -> usize
Number of key-value heads (for MQA/GQA)
Hidden dimension size
Sourcefn intermediate_size(&self) -> usize
fn intermediate_size(&self) -> usize
Intermediate/FFN dimension size
Sourcefn vocab_size(&self) -> usize
fn vocab_size(&self) -> usize
Vocabulary size
Sourcefn max_position_embeddings(&self) -> usize
fn max_position_embeddings(&self) -> usize
Maximum position embeddings
Sourcefn rope_theta(&self) -> Option<f32>
fn rope_theta(&self) -> Option<f32>
RoPE base frequency (if applicable)
Sourcefn layer_norm_eps(&self) -> f32
fn layer_norm_eps(&self) -> f32
Layer normalization epsilon
Sourcefn use_rms_norm(&self) -> bool
fn use_rms_norm(&self) -> bool
Whether to use RMSNorm instead of LayerNorm
Provided Methods§
Sourcefn use_parallel_residual(&self) -> bool
fn use_parallel_residual(&self) -> bool
Whether this architecture uses parallel attention and FFN
Sourcefn attention_type(&self) -> AttentionType
fn attention_type(&self) -> AttentionType
Get attention type (MHA, MQA, GQA)
Sourcefn default_dtype(&self) -> DType
fn default_dtype(&self) -> DType
Default dtype for this architecture