Module: aprender::models

Public module of the aprender-core crate.

Source

crates/aprender-core/src/models.rs or directory.

Example

use aprender::models::Qwen2Model;
// See `cargo doc -p aprender-core --open` for full API reference.

Module summary

aprender::models is where pre-implemented transformer architectures live — BERT (encoder, cross-encoder) for embeddings and reranking, Qwen2 for decoder-only generation. These are reference implementations of the architectures used by the trained checkpoints aprender ships, designed for training and offline analysis. For production inference at scale, use the realizar runtime (re-exported as aprender-serve); these in-crate models are for training, fine-tuning, and architecture experiments.

Key types

TypeDescription
Qwen2ModelQwen2 decoder-only transformer with grouped-query attention and SwiGLU MLP.
BertConfigHyperparameters for BERT (hidden size, attention heads, layers, intermediate size, max positions).
BertEncoderPre-norm BERT encoder for sentence embeddings.
CrossEncoderCross-attention BERT used for pairwise scoring / reranking.

Inside models::qwen2, the building blocks Embedding, Qwen2MLP, Qwen2DecoderLayer, GroupedQueryAttention, and Qwen2Config are publicly accessible so you can compose custom variants.

Usage patterns

Pattern 1: Construct a BERT encoder from a config

use aprender::models::{BertConfig, BertEncoder};

let config = BertConfig {
    hidden_size: 384,
    num_attention_heads: 12,
    num_hidden_layers: 6,
    intermediate_size: 1536,
    max_position_embeddings: 512,
    vocab_size: 30_522,
    ..BertConfig::default()
};
let encoder = BertEncoder::new(&config);
println!("encoder ready with {} layers", config.num_hidden_layers);

Pattern 2: Inspect Qwen2 components

use aprender::models::qwen2::{Qwen2MLP, Embedding};

// MLP block — SwiGLU: gate_proj + up_proj into down_proj
let mlp = Qwen2MLP::placeholder(896, 4864);

// Token embedding table for a 151,936-token vocab
let emb = Embedding::placeholder(151_936, 896);
println!("hidden size = {}", emb.weight().shape()[1]);

See also

  • nn — the layer primitives (Linear, RMSNorm, attention) these models build on
  • format — load weights from GGUF / SafeTensors / APR
  • text — tokenization (BPE, chat templates) for prompt preparation
  • autograd — Tensor and backward used by these architectures during training

Full API

Run cargo doc -p aprender-core --open for the rendered rustdoc, or browse docs.rs/aprender for the published version.