KOSMOS：面向通信-计算重叠的自调度持久内核库
KOSMOS: Self-scheduling Persistent Kernels for Communication-Compute Overlap
AI 导读
KOSMOS 是一个自调度持久内核库，其 workgroup 通过按消费者需求排序的队列，在数据就绪时自行切换通信与计算角色，无需调度器、代价模型或预设切分。
正文 · 原文
Abstract:Tensor and sequence parallelism (TP+SP) and expert parallelism (EP) put collectives on the critical path of transformer layers. State-of-the-art fused kernels mitigate this with fixed roles or schedules, but this rigidity leaves large idle windows. Predicting a communication-computation split does not remove them either; a predetermined split cannot follow data readiness. We present KOSMOS, a library for self-scheduling persistent kernels whose workgroups use queues ordered by consumer demand to switch between communication and computation roles as data become ready, with no scheduler, per-call cost model, or tuned split. KOSMOS fuses every TP+SP collective of a transformer layer with its GEMM, and runs an entire expert layer, from router to fold, in one launch per direction. Evaluated on AMD Instinct MI350X and AMD Instinct MI355X GPUs, KOSMOS speeds up TP+SP end-to-end training in Megatron-LM by 1.10-1.28x in BF16 and 1.10-1.25x in MXFP8. KOSMOS accelerates end-to-end training of DeepSeek-V3 and Qwen3-235B proxies by 1.29-1.83x over DeepEP.
| Comments: | 39 pages, 9 figures |
| Subjects: | Distributed, Parallel, and Cluster Computing (cs.DC); Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Performance (cs.PF) |
| Cite as: | arXiv:2611.07477 [cs.DC] |
| (or arXiv:2611.07477v1 [cs.DC] for this version) | |
| https://doi.org/10.48550/arXiv.2611.07477 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Alexander Magro [view email]
[v1]
Thu, 5 Nov 2026 05:36:08 UTC (2,476 KB)
来源：arXiv cs.AI · arxiv.org