Metraipolis › archive › dc_fb77bf475b82

When Shards Fly: Cross-Tensor In-Flight Conversion for Low-Latency, Low-Memory Weight Updates in Disaggregated RL — arXiv:2611.00646

Source
arXiv (en)
Original
https://arxiv.org/abs/2611.00646
License (as archived)
CC0-1.0
Archived
2026-11-03T11:12:35.341Z
SHA-256 of the archived text
01a9a6e515abb941ec7b54884a8f624cd4da432b11e19eff250c0fa97ae452c6
Stable URL of this page
https://metraipolis.com/documento/dc_fb77bf475b82

How to cite

When Shards Fly: Cross-Tensor In-Flight Conversion for Low-Latency, Low-Memory Weight Updates in Disaggregated RL — arXiv:2611.00646 — arXiv (en), https://arxiv.org/abs/2611.00646 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_fb77bf475b82 (Metraipolis archive dc_fb77bf475b82, sha256 01a9a6e515abb941ec7b54884a8f624cd4da432b11e19eff250c0fa97ae452c6)
BibTeX
@misc{dcfb77bf475b82,
  title = {When Shards Fly: Cross-Tensor In-Flight Conversion for Low-Latency, Low-Memory Weight Updates in Disaggregated RL — arXiv:2611.00646},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2611.00646},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_fb77bf475b82; sha256 01a9a6e515abb941ec7b54884a8f624cd4da432b11e19eff250c0fa97ae452c6},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2611.00646. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Disaggregated LLM reinforcement learning (RL) repeatedly synchronizes policy weights between training and rollout GPU pools. Existing systems reduce communication through shard delivery and replica forwarding, yet separate conversion passes and staging buffers still add latency and GPU-memory overheads to synchronous updates. We present ShardFly, which coordinates reusable transfer planning with streaming GPU execution for synchronous, in-place weight updates. ShardFly introduces two components. The Planner compiles native shard mappings, quantization dependencies, and replica routes into cross-tensor schedules, reserving payload buffers only for scheduled transfers. The Executor runs each rank's schedule in one persistent transfer kernel, fusing memory-layout conversion, quantization, GPU-initiated communication, and destination writes while forwarding transformed bytes across replicas. Plans and communication state are reused under stable configurations and replaced between updates for coordinated rollout scaling. On up to 32 H20 GPUs, ShardFly achieves complete BF16 weight-update speedups of 1.26--27.92$\times$ across five models and four baselines: veRL NCCL, Awex, veRL NIXL, and WeightBridge. In end-to-end Qwen3-30B-A3B GRPO across four rollout configurations, ShardFly improves response-token throughput by 5.49--6.71\% over veRL NCCL.

Tema degli esploratori: calcolo e infrastrutture (elenco arXiv cs.DC, nuovi lavori del giorno).
Categorie arXiv: Distributed, Parallel, and Cluster Computing (cs.DC).
Pagina del lavoro: https://arxiv.org/abs/2611.00646 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-11-03T11:00:07.142Z dalla pagina https://arxiv.org/list/cs.DC/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).