Aster v0.1.0

Aster v0.1.0

Understand the model. Own the training loop. Follow it into serving.

Aster is a native PyTorch framework for learning and building connected ML workflows—from model definitions and training objectives to fine-tuning, inference, agents, and evaluation.

Aster 首次发布：用可阅读的原生实现，把模型、训练、微调、蒸馏、推理、智能体与评测连接起来。

Documentation · 简体中文 · Learning path · Algorithms & papers

Highlights

• Models and objectives: language, multimodal, action, generative, and world-model components; supervised learning, reinforcement learning, and distillation.
• Shared training: loss normalization, checkpoints, EMA, Muon, and supported parallel/ZeRO layouts.
• Fine-tuning and inference: linear LoRA, shared-base adapters, paged KV, prefix reuse, continuous single-worker batching, and low-bit KV storage.
• Agents and evaluation: bounded tools, selected MCP capabilities, artifact-backed workflows, and reproducible evaluation protocols.
• Learning resources: bilingual guides, an algorithm-to-code index, source tests, and three connected CPU examples.

Get started

Python 3.11+ and a compatible PyTorch installation are required.

git clone --branch v0.1.0 https://github.com/sutiankang/aster.git
cd aster
python -m pip install -e ".[test]"
python -m aster doctor
python examples/quickstart.py

The example trains and merges a tiny LoRA adapter without model downloads, API keys, or a GPU.

If Aster helps you learn or build, give it a star ⭐ to help others discover it.

Verification

• CPU CI for this source revision passed on Python 3.11 and 3.13. Each job: 1,393 passed, 44 skipped, 18 warnings. Skipped tests are not counted as passes.
• Installed-wheel validation: 34 targeted tests and all three CPU workflow examples passed.
• Package contents, metadata, and internal documentation links checked.

Downloads

• Wheel: install the Aster runtime.
• Source archive: includes source, tests, examples, and documentation.
• SHA256SUMS: verify the two package downloads.

Source: b2bcf13. No model weights or benchmark datasets are bundled. This release does not publish a package to PyPI.

Scope and license

APIs are experimental. Supported configurations and remaining work are listed in Status and Roadmap. CPU checks do not establish GPU performance or pretrained-model quality.

A repository-wide license has not yet been granted. See NOTICE for third-party terms and the ongoing source review before reusing or redistributing code.