Reproduce “Petals: Collaborative Inference and Fine-tuning of Large Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alexander Borzunov, Dmitry Baranchuk, Tim Dettmers, Max Ryabinin, Younes Belkada, Artem Chumachenko, Pavel Samygin, Colin Raffel
PETALS enables collaborative inference and parameter-efficient fine-tuning of large language models (LLMs) such as BLOOM-176B, allowing their use on distributed consumer-grade GPUs at interactive speeds of approximately one inference step per second.
Key insights
- Combining 8-bit mixed matrix decomposition for weights and dynamic blockwise quantization for activations significantly reduces memory and bandwidth requirements with minimal impact on model quality.
- A decentralized pipeline parallel execution strategy, leveraging community-contributed GPU resources, can achieve interactive inference speeds (approx. 1 step/second) for LLMs like BLOOM-176B, surpassing local offloading methods for low-latency tasks.
- Robust fault tolerance protocols, dynamic load balancing, and efficient client-side routing are crucial for maintaining performance and reliability in a peer-to-peer network of diverse and potentially transient devices.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.