KOSMOS: Self-Scheduling Persistent Kernels for Communication-Compute Overlap

Researchers present KOSMOS, a library for self-scheduling persistent kernels designed to improve communication-compute overlap in transformer models. The system uses queues ordered by consumer demand to dynamically switch workgroups between communication and computation roles without relying on a scheduler or tuned split. Evaluated on AMD hardware, KOSMOS accelerates end-to-end training for tensor and sequence parallelism as well as DeepSeek-V3 and Qwen3-235B proxies. Open article →