Struct Qwen2Model 

pub struct Qwen2Model { /* private fields */ }

Complete Qwen2 model for inference.

Assembles embedding, decoder layers, and LM head into a complete model.

Implementations

impl Qwen2Model

pub fn new(config: &Qwen2Config) -> Self

Create a new Qwen2 model from configuration.

Weights are initialized randomly. Use load() to load pre-trained weights.

pub fn new_uninitialized(config: &Qwen2Config) -> Self

Create an uninitialized Qwen2 model with minimal memory allocation.

The model is not ready for inference until weights are loaded.

pub fn forward(&mut self, input_ids: &[u32], position_ids: &[usize]) -> Tensor

Forward pass through the model.

Arguments

• input_ids - Token IDs [seq_len]
• position_ids - Position indices [seq_len]

Returns

Logits tensor [1, seq_len, vocab_size]

pub fn forward_profiled( &mut self, input_ids: &[u32], position_ids: &[usize], ) -> Tensor

Forward with detailed profiling output. Prints timing breakdown for each component.

pub fn generate( &mut self, prompt_ids: &[u32], max_new_tokens: usize, temperature: f32, _top_p: f32, ) -> Vec<u32>

Generate tokens autoregressively.

Arguments

• prompt_ids - Initial prompt token IDs
• max_new_tokens - Maximum number of tokens to generate
• temperature - Sampling temperature (0 = greedy)
• top_p - Nucleus sampling threshold

Returns

Complete sequence including prompt and generated tokens.

pub fn generate_profiled( &mut self, prompt_ids: &[u32], max_new_tokens: usize, temperature: f32, ) -> Vec<u32>

Generate with profiling output (prints timing breakdown).

pub fn config(&self) -> &Qwen2Config

Get model configuration.

pub fn eval(&mut self)

Set model to evaluation mode (no dropout).

pub fn train(&mut self)

Set model to training mode.

pub fn enable_cache(&mut self)

Enable KV cache for efficient generation.

pub fn disable_cache(&mut self)

Disable KV cache.

pub fn clear_cache(&mut self)

Clear KV cache.

pub fn num_layers(&self) -> usize

Get number of layers.

pub fn weight_names(&self) -> Vec<String>

Get list of weight names following HuggingFace convention.

Returns names like:

• model.embed_tokens.weight
• model.layers.0.self_attn.q_proj.weight
• model.norm.weight
• lm_head.weight

pub fn weight_info(&self) -> HashMap<String, Vec<usize>>

Get weight shapes as a map from name to shape.

pub fn weights(&self) -> HashMap<String, Vec<f32>>

Extract accessible weights as a map from name to f32 data.

Returns a map suitable for serialization to SafeTensors format. Note: Currently returns weights from components with public accessors. Full weight export will be enabled when nn modules expose weight accessors.

pub fn num_parameters(&self) -> usize

Get total number of parameters in the model.

pub fn embed_tokens_mut(&mut self) -> &mut Embedding

Get mutable reference to embedding layer.

pub fn layer_mut(&mut self, idx: usize) -> Option<&mut Qwen2DecoderLayer>

Get mutable reference to decoder layer at index.

pub fn norm_mut(&mut self) -> &mut RMSNorm

Get mutable reference to final norm layer.

pub fn lm_head_mut(&mut self) -> &mut Linear

Get mutable reference to language model head.

pub fn lm_head(&self) -> &Linear

Get reference to language model head (for testing/inspection).

pub fn load_from_safetensors(&mut self, path: &Path) -> Result<usize, String>

Load weights from SafeTensors format.

Arguments

• path - Path to .safetensors file

Returns

Number of weights loaded

Errors

Returns error if file cannot be read or weights don’t match.

pub fn from_safetensors( config: &Qwen2Config, path: &Path, ) -> Result<Self, String>

Load model from SafeTensors file.

Creates a new model with the given config and loads weights from file.

pub fn load_from_apr(&mut self, path: &Path) -> Result<usize, String>

Load weights from APR v2 format file.

Per Native Library Mandate (Spec §2.4): Uses mmap via bundle::MappedFile for zero-copy tensor access. This is the REQUIRED approach for APR files.

Note: APR canonical names don’t have the “model.” prefix (it’s stripped during import per format/converter.rs). We look for names without prefix.

Returns

Number of weights loaded

Errors

Returns error if file cannot be read or weights don’t match.

pub fn from_apr(config: &Qwen2Config, path: &Path) -> Result<Self, String>

Load model from APR v2 format file.

Creates a new model with the given config and loads weights from file.

Trait Implementations

impl Debug for Qwen2Model

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.