v1.1.0 - VSA-Accelerated Training

VSA-Accelerated Training

This release adds deterministic gradient prediction using Vector Symbolic Architectures for significant training speedups.

New Features

• VSAAccelerator: Wrapper for deterministic gradient prediction
• DeterministicPhaseTrainer Integration: WARMUP → FULL → PREDICT → CORRECT cycle
• VSAConfig: Fine-grained control over training phases
• Ternary Gradient Accumulation: Balanced {-1, 0, +1} representation

Changes

• Update Cargo.toml with vsa-optim optional feature
• Add homepage and documentation URLs to metadata
• Fix rust-version from 1.92 to 1.70

Usage

use axolotl_rs::vsa_accel::{VSAAccelerator, VSAConfig};

let config = VSAConfig::default()
    .with_warmup_steps(10)
    .with_predict_steps(20);

let accelerator = VSAAccelerator::new(config, &var_map, &device)?;

Enable the feature:

cargo add axolotl-rs --features vsa-optim

Testing

All 115 unit tests and integration tests passing.

Documentation

• crates.io
• docs.rs