skyrl-tx 0.2.1

Unified API for training and inference

SkyRL tx: Unified API for training and inference

⚠️ The project is currently very early with lots of missing features (e.g. currently LoRA is only supported for the MLP layer, model sharding is in a very early state). Many of these are easy to implement and we welcome contributions! ⚠️

SkyRL tx is an open-source cross-platform library that allows users to set up their own service exposing a Tinker like REST API for neural network forward and backward passes. It unifies inference and training into a single, common API, abstracting away the infrastructure challenges of managing GPUs.

The t in tx stands for transformers, training, or tinker, and the x stands for "cross-platform".

Getting Started

See the following SkyRL tx blog posts for more info and examples:

• Initial blog post with the piglatin training example
• v0.0.2 release with a Qwen/Qwen3-4B training example
• v0.0.3 release with a Qwen/Qwen3-Coder-30B-A3B training and a sampling example
• v0.1.0 release with an end-to-end RL example

See also our talk SkyRL tx: A unified training and inference engine at Ray Summit 2025.

Features

✅ Implemented

• Training: MultiLoRA fine-tuning with gradient accumulation
• Inference: Text generation with
  • Temperature sampling
  • Stop token support
• API: REST API compatible with Tinker specification

🚧 In Progress

• Model sharding improvements
• Additional LoRA layer support

Project Status

This is a very early release of SkyRL tx. While the project is functional end-to-end, there is still a lot of work to be done. We are sharing it with the community to invite feedback, testing, and contributions.