Complete, efficient GPU residency for Distributed Machine Learning.
Kleos is an ongoing research project exploring the design of GPU-resident infrastructure for distributed machine learning workloads. The goal is to eliminate CPU bottlenecks by fusing scheduling, communication, and compute directly on the GPU using lightweight, asynchronous primitives.
We aim to achieve complete, efficient GPU residency. We investigate optimizations that approach hardware peak performance for both distributed and single GPU workloads, where bulk-synchronous, CPU-driven orchestration is a limiting factor. To attain this vision, we employ kernel fusion enabled by (1) algorithmic innovations with strong theoretical footing and (2) principled systems design and implementation.
This repository represents a very early-stage release of Kleos infrastructure.
- June 5, 2025 — ⚡️Introducing FlashDMoE, a fused GPU kernel for distributed MoE execution.
➤ Seethis READMEfor details, benchmarks, and usage.
This project is licensed under the BSD 3-Clause License. See LICENSE for full terms.