Open-source distributed-training tooling for PyTorch, published by TsugiCinema, Inc.
pip install tsugi gives you one import namespace over two independent,
Apache-2.0 SDKs that each target a different bottleneck in large-model training:
| Package | Install | Imports as | What it does |
|---|---|---|---|
| tsugi | pip install tsugi | tsugi | Meta-package that re-exports both SDKs under one namespace. |
| tsugi-mend | pip install tsugi-mend | tsugi.mend | Cross-rack distributed-training reducer. Integrates Decoupled DiLoCo, DES-LOC, async tensor parallelism, and FALCON fail-slow mitigation to recover the cross-rack synchronization tax on commodity-network clusters. |
| tsugi-kpool | pip install tsugi-kpool | tsugi.kpool | K-Pool LoRA SDK. A drop-in extension to PyTorch + PEFT for distributed LoRA fine-tuning with K-of-N adapter routing and a buffer-convergence aggregation trigger. |
The two SDKs share zero code. The meta-package wraps them at the import surface for convenience; either can be installed and used on its own.
Everything here is licensed under the Apache License, Version 2.0, with
its full automatic patent grant. See each repository's LICENSE and NOTICE
for the per-package patent posture.
- Homepage: https://tsugilabs.ai
tsugi(meta-package): https://github.com/tsugiai/tsugitsugi-mend(cross-rack reducer): https://github.com/tsugiai/tsugi-mendtsugi-kpool(K-Pool LoRA): https://github.com/tsugiai/tsugi-kpool