Release v0.2.3

Maintenance release.

• Support manual sharding (#816)
• Optimize cross-mesh communication (#773, #798)
• Add publications (#885)
• Support new models: CodeGen, BLOOM-Z (#774, #844)
• Add priority scheduling in the model server (#852)
• Add guidance on strategy inspection (#879)
• Inference stage construction (#793, #799)
• Misc bug fixes (#876, #878, #873)

Release v0.2.2

• Rebase to jax v0.3.22 (#782)
• Support TPU backend for ShardParallel (#764)
• Misc improvements for LLM serving

Release v0.2.1

• Fine-tune OPT with 3d parallelism #716
• Support bigscience/bloom #707
• Add model serving controller #720
• Jaxlib update

Release v0.2.0

• Improve OPT generation speed (#661)
• Add logprobs endpoint (#679)
• Support global norm in pipeline parallelism (#672)
• Rebase jax to v0.3.15 (#682)
• Improve the stability of the server

Release v0.1.6

• Improve OPT-175B demo https://opt.alpa.ai/. Our website has served 20k+ requests during the last week.
• Add OPT fine-tuning examples https://github.com/alpa-projects/alpa/tree/main/examples/opt_finetune
• Various bug fixes

Release v0.1.5

• Improve OPT-175B demo https://opt.alpa.ai/
• Improve version management
• Various bug fixes

Release v0.1.4

• [Feature] Serialize Parallel Plan (#587)
• [Test] Move test install to alpa package (#584)
• [CI/CD] New PyPI Infra (#577)

Release v0.1.3

• [Tutorial] Serve OPT-175B (https://alpa-projects.github.io/tutorials/opt_serving.html)
  • New website
  • Better weight conversion scripts
• [API Change] Move layer construction decorators to ParallelMethod (#565)

Release v0.1.1

• Interface change (#540)
• Minor bug fixes

Release v0.1.0

Major feature updates:

• Support serving Meta's open source OPT model 175B
• Distributed weight init
• Add a set of strategy-specific interfaces such as DataParallel, Zero2Parallel, Zero3Parallel

Minors:

• enhancement on the PyTorch frontend
• many bugfixes
• Doc update: Performance tuning guide, and a slide deck for Alpa