Training workbench to make your model glow
Twinkle is a lightweight, client-server training framework engineered with modular, high-cohesion interfaces for LLM training.
- Loosely Coupled Architecture — Standardized interfaces with backward compatibility. Use only what you need.
- Multiple Runtime Modes — Run locally with torchrun, scale across Ray clusters, or deploy as HTTP services.
- Multi-Framework Support — Works with both Transformers and Megatron backends.
- Multi-Tenancy Training — Train multiple LoRAs on a shared base model with isolated configurations.
- Training as a Service — Built-in capabilities for automated cluster management and dynamic scaling.
- Full Training Control — Retain control over forward, backward, and step operations for easy debugging.
- Source Code: github.com/modelscope/twinkle
- Documentation: modelscope.github.io/twinkle-web/docs/
- Releases: github.com/modelscope/twinkle/releases
Apache License 2.0