Struct LayerWeights 

pub struct LayerWeights {
    pub attn_q: WeightStorage,
    pub attn_q_bias: Option<Arc<[f32]>>,
    pub attn_k: WeightStorage,
    pub attn_k_bias: Option<Arc<[f32]>>,
    pub attn_v: WeightStorage,
    pub attn_v_bias: Option<Arc<[f32]>>,
    pub attn_o: WeightStorage,
    pub ffn_gate: WeightStorage,
    pub ffn_up: WeightStorage,
    pub ffn_down: WeightStorage,
    pub attn_norm_gain: Arc<[f32]>,
    pub ffn_norm_gain: Arc<[f32]>,
}

Per-layer weights of a LLaMA transformer block. All tensors are stored as Arc<[f32]> so they can be loaded once and shared across every forward pass with zero copy — each call to LlamaModel::forward clones the Arc (a refcount bump) when it builds fresh const nodes for this layer.

LLaMA proper has no biases anywhere in the attention block, so the *_bias fields are None for LLaMA family models. Qwen2 and a few related architectures do add biases on Q/K/V (but not on the output projection), so the loader stores them here when the safetensors file contains them.

Fields

attn_q: WeightStorage

[dim, dim] query projection. Supports bf16 or f32.

attn_q_bias: Option<Arc<[f32]>>

[dim] query projection bias (Qwen2-style; LLaMA has none).

attn_k: WeightStorage

[dim, dim] key projection.

attn_k_bias: Option<Arc<[f32]>>

[kv_dim] key projection bias.

attn_v: WeightStorage

[dim, dim] value projection.

attn_v_bias: Option<Arc<[f32]>>

[kv_dim] value projection bias.

attn_o: WeightStorage

[dim, dim] output projection.

ffn_gate: WeightStorage

[dim, ffn_dim] gate projection for SwiGLU.

ffn_up: WeightStorage

[dim, ffn_dim] up projection for SwiGLU.

ffn_down: WeightStorage

[ffn_dim, dim] down projection for SwiGLU.

attn_norm_gain: Arc<[f32]>

[dim] RmsNorm gain for the pre-attention norm. Stays f32 — norm gains are small and precision-sensitive.

ffn_norm_gain: Arc<[f32]>

[dim] RmsNorm gain for the pre-FFN norm.

Trait Implementations

impl Clone for LayerWeights

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LayerWeights

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

Formats the value using the given formatter.