Trait DecodeBackbone 

pub trait DecodeBackbone: PersistentDecodeModel {
    // Required methods
    fn decode_family(&self) -> &'static str;
    fn decode_dims(&self) -> DecodeDims;
    fn decode_shape_key(&self) -> u64;
    fn decode_mask_plan(&self) -> MaskPlan;
    fn decode_rope_plan(&self) -> RopePlan;
    fn decode_token_embedding(&self) -> Arc<[f32]> ;
    fn decode_apply_layer(
        &self,
        layer_idx: usize,
        inputs: &DecodeLayerInputs<'_>,
    ) -> Result<Tensor>;
    fn decode_final_norm_and_head(&self, h: &Tensor) -> Result<Tensor>;
}

Everything the shared build path needs from one model family.

Four of increment 3’s seven per-family items are served through this trait rather than copied; the two that are genuinely per-family — the attention block and the token-data hook — are the two methods that take real work.

Required Methods

fn decode_family(&self) -> &'static str

Family name, threaded into every error this shared path raises.

Not cosmetic. The 23 hunks that made collapsing two build bodies look hard turned out to be mostly error-message prefixes, so sharing the body makes diagnostics lie by default: without this, a Qwen3Moe failure reports "forward_with_kv_context: …" and reads as a Llama bug.

fn decode_dims(&self) -> DecodeDims

fn decode_shape_key(&self) -> u64

Identity of the held decode plan — family + the config values that change graph structure + this weight set. Must fold in MaskPlan::mix_into.

fn decode_mask_plan(&self) -> MaskPlan

fn decode_rope_plan(&self) -> RopePlan

Per-layer RoPE-base assignment. Required, not defaulted, on purpose: a default would be silently correct for the single-base families and silently WRONG for a multi-base one that forgot to override it. Every family stating it is the compiler enumerating them.

fn decode_token_embedding(&self) -> Arc<[f32]>

The [vocab, hidden] embedding table the token lookup reads.

fn decode_apply_layer( &self, layer_idx: usize, inputs: &DecodeLayerInputs<'_>, ) -> Result<Tensor>

One transformer layer: attention (writing this step’s K/V into the cache buffers) plus the FFN. This is the per-family architecture — partial rotary, GQA replication, norm placement, bias presence, MoE routing all live here and nowhere else on this path.

layer_idx is passed because per-layer variation is real (mask variant selection is resolved by the caller, but Gemma3’s RoPE base and LFM2’s state kind are not).

fn decode_final_norm_and_head(&self, h: &Tensor) -> Result<Tensor>

Final norm + LM head → [batch, seq, vocab].

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors