Struct DecisionModel 

pub struct DecisionModel { /* private fields */ }

The full decision model: encoder + head.

Implementations

impl DecisionModel

pub fn load( weights_path: &Path, encoder_config: &Value, head_layers: usize, n_act: usize, device: Device, ) -> Result<Self>

Load a checkpoint from model.safetensors + the encoder config directory.

• weights_path — path to model.safetensors.
• encoder_config — parsed encoder/config.json.
• head_layers — cfg.head_layers.
• n_act — len(cfg.act_costs) + 1.

pub fn hidden_size(&self) -> usize

Encoder hidden size.

pub fn embed( &self, input_ids: &[Vec<u32>], attention: &[Vec<u32>], ) -> Result<Vec<Vec<f32>>>

Mean-pool the encoder’s last_hidden_state over non-padding tokens. input_ids/attention are padded rectangular batches. Returns [n, hidden]. Mirrors embed_fn_from_agent in laya/shortlist.py.

pub fn forward(&self, b: &CollatedBatch) -> Result<ForwardOutput>

Run one collated batch. Returns (logits, act_logits) as row-major host vectors: logits[row][0..kmax] (masked options filled with -1e4) and act_logits[row][0..n_act].