Moxzi Alpha 9 adds GPU training and persistent actors to Motoko 2.0

Moxzi has released Alpha 9, an update that brings GPU-powered machine learning, model sharing and improved recovery features to its Motoko development environment for the Internet Computer (ICP).

The release combines work from Alpha 8 and Alpha 9, as Alpha 8 was not publicly released. It also introduces support for Motoko 2.0 across the toolchain, updating language syntax, persistence defaults, diagnostics and runtime behaviour.

Moxzi is a Motoko compiler written in Motoko, with runtimes that allow Motoko actors to run on the Internet Computer, servers and browsers. The latest update focuses on expanding what these actors can do, including training neural networks and retaining learned state across compatible software upgrades.

One of the main additions is GPU-enabled machine learning for hosted Motoko actors. With a compatible host configuration, actors can train and run neural networks using GPU or CPU processing. The system uses the Burn machine learning framework, with Apple Metal and CPU support, while CUDA and Vulkan builds are available through feature options. CUDA support had not yet been qualified on NVIDIA hardware for this release.

The GPU functionality runs on off-chain hosts rather than inside the Internet Computer’s replicated canisters. This distinction means the host’s configuration and security boundaries remain important, and the release does not provide the same consensus or certification guarantees as on-chain execution.

Alpha 9 also allows actors to publish model weight snapshots and grant selected readers access. Other actors can use shared weights or combine compatible models through weighted mixing. A distillation feature also allows a smaller model to learn from the predictions of a larger or differently structured model.

The update improves how local actors retain their state. These actors can exist within another actor’s container, allowing multiple entities to maintain their own identities and state without requiring a separate canister for each one. Under the new persistence defaults, local actors stored in stable fields can retain their state and handles across compatible upgrades.

Reliability improvements address recovery after crashes, including the restoration of model weights, optimiser state, stored values and publication permissions. Alpha 9 introduces checkpoints based on accumulated device processing time, with a default target of 30 seconds of device work between checkpoints. The project notes that this is a replay target rather than a guarantee that a restart will take 30 seconds.

Performance tests reported improvements across several workloads. Text generation using a key-value attention cache increased from 1,044 to 14,681 characters per second, while an eight-expert model’s training step on Metal fell from two seconds to 0.22 seconds. Weight synchronisation between two hosts behind Ray dropped from about 40 seconds to 1.2 seconds.

The developers caution that these results apply to the tested workloads and should not be interpreted as universal performance gains for every application or model.

Other changes include stricter compilation and migration checks, browser execution and memory limits, improved saved-world restoration, and tighter controls around host identity, storage isolation, timers and recovery. Developers upgrading from earlier persistence configurations are advised to rehearse migrations, as the move to enhanced orthogonal persistence is irreversible.

Moxzi Alpha 9 is available as alpha software under the Business Source License 1.1. Installation instructions, GPU training examples and the full release notes are available through the Moxzi website and its documentation.