Struct ModelWeights 

pub struct ModelWeights {
    pub embed_tokens: DenseWeight,
    pub final_norm: DenseWeight,
    pub lm_head: DenseWeight,
    pub layers: Vec<LayerWeights>,
}

All model weights organized by layer.

Fields

embed_tokens: DenseWeight

Embedding table: [vocab_size, hidden_size] BF16.

final_norm: DenseWeight

Final RMS norm: [hidden_size] BF16.

lm_head: DenseWeight

LM head: [hidden_size, vocab_size] BF16.

layers: Vec<LayerWeights>

Per-layer weights.

Implementations

impl ModelWeights

pub fn from_store( store: &WeightStore, layer_types: &[LayerType], num_experts: usize, gpu: &dyn GpuBackend, config: &ModelConfig, ) -> Result<Self>

Build typed weight references from a flat WeightStore.

layer_types maps layer index → FullAttention or LinearAttention. num_experts is 512 for Qwen3-Next.