Struct LoadedModel 

pub struct LoadedModel {
    pub model: Box<dyn Module>,
    pub tokenizer: Tokenizer,
    pub device: Device,
    pub dtype: DType,
    pub adapter_layers: Option<AdapterLayers>,
    pub trainable_params: VarMap,
}

Loaded model with configuration.

Fields

model: Box<dyn Module>

Model weights and forward pass

tokenizer: Tokenizer

Tokenizer

device: Device

Device where model is loaded

dtype: DType

Model dtype

adapter_layers: Option<AdapterLayers>

Adapter layers (if using LoRA/QLoRA)

trainable_params: VarMap

Trainable parameters (LoRA weights)

Implementations

impl LoadedModel

pub fn forward(&self, input_ids: &Tensor) -> Result<Tensor, AxolotlError>

Run forward pass on input tokens.

Errors

Returns an error if the forward pass fails.

pub fn forward_with_adapters( &self, input_ids: &Tensor, ) -> Result<Tensor, AxolotlError>

Run forward pass with adapter layers.

IMPORTANT: Current implementation does NOT properly integrate adapters. LoRA adapters need to be injected at each attention/MLP layer, not applied post-hoc to logits. This requires custom model architecture (LoraLlama).

For now, this returns base model output. Gradient flow is maintained through the trainable LoRA parameters in trainable_params VarMap.

Errors

Returns an error if the forward pass fails.

pub fn trainable_tensors(&self) -> Vec<Var>

Get trainable parameters for optimizer.

Returns only the LoRA A/B matrices, not the frozen base model weights.

pub fn trainable_param_count(&self) -> usize

Count trainable parameters.