Struct Lfm2Model 

pub struct Lfm2Model {
    pub embed_tokens: Tensor,
    pub embedding_norm: Option<RMSNorm>,
    pub layers: Vec<Lfm2Block>,
    pub norm_f: RMSNorm,
    pub lm_head: Linear,
    /* private fields */
}

A decoder-only language model: LFM2 (mixed conv/attention blocks) or Llama-style (every block is attention), as built by Lfm2ModelBuilder.

Fields

embed_tokens: Tensor

Embedding table [vocab, hidden].

embedding_norm: Option<RMSNorm>

Optional norm applied right after the embedding (neither Llama nor Hugging Face LFM2 use one; LFM2’s embedding_norm is its final norm).

layers: Vec<Lfm2Block>

Blocks.

norm_f: RMSNorm

Final norm.

lm_head: Linear

Projection to vocabulary logits.

Implementations

impl Lfm2Model

pub fn new( embed_tokens: Tensor, embedding_norm: Option<RMSNorm>, layers: Vec<Lfm2Block>, norm_f: RMSNorm, lm_head: Linear, ) -> Self

Assemble a model from its parts.

pub fn supports_kv_cache(&self) -> bool

Whether every block is attention, so incremental decoding with a KVCache is exact (conv blocks keep no state).

pub fn new_kv_cache(&self, max_seq: usize) -> Option<KVCache>

An empty KV cache for sequences of up to max_seq tokens, or None when supports_kv_cache is false.

pub fn forward( &self, engine: &ComputeEngine, input_ids: &[u32], pos: usize, cache: Option<&mut KVCache>, mask: Option<&UnifiedTensor>, ) -> Tensor

Logits [1, seq, vocab] for input_ids, the first of which is at pos.

pub fn lora_params(&self) -> Vec<Tensor>

Only the LoRA adapter tensors (what LoRA fine-tuning updates).

pub fn params(&self) -> Vec<Tensor>

All trainable tensors.

pub fn named_params(&self) -> HashMap<String, Tensor>

All parameters keyed by Hugging Face-style names.

pub fn named_lora_params(&self) -> HashMap<String, Tensor>

LoRA tensors keyed by name.