Metraipolis › archive › dc_e3fc69e8aece

Flash RNN-T: Memory-Efficient Joint Network and Transducer Loss Computation — arXiv:2610.23409

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.23409
License (as archived)
CC0-1.0
Archived
2026-10-19T08:44:18.781Z
SHA-256 of the archived text
64d20ffb64b424ea07a6813ba7fcd2b13e2fdc39b4e10cbd60cd5597039b860f
Stable URL of this page
https://metraipolis.com/documento/dc_e3fc69e8aece

How to cite

Flash RNN-T: Memory-Efficient Joint Network and Transducer Loss Computation — arXiv:2610.23409 — arXiv (en), https://arxiv.org/abs/2610.23409 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_e3fc69e8aece (Metraipolis archive dc_e3fc69e8aece, sha256 64d20ffb64b424ea07a6813ba7fcd2b13e2fdc39b4e10cbd60cd5597039b860f)
BibTeX
@misc{dce3fc69e8aece,
  title = {Flash RNN-T: Memory-Efficient Joint Network and Transducer Loss Computation — arXiv:2610.23409},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.23409},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_e3fc69e8aece; sha256 64d20ffb64b424ea07a6813ba7fcd2b13e2fdc39b4e10cbd60cd5597039b860f},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.23409. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

The recurrent neural network transducer (RNN-T) remains a practical choice for streaming speech recognition, but training requires large intermediate tensors. We present Flash RNN-T, a joint and loss implementation combining specialized kernels with optional tiled recomputation over the full alignment space and vocabulary. Tiled and untiled Flash form the entire measured throughput-memory frontier on medium and long workloads. On H100, untiled Flash achieves 24-39% higher throughput with 26-56% less memory than the fastest baseline on these workloads wherever it fits. Tiling reduces memory further through recomputation. On medium, tiled Flash uses ninefold less memory than that baseline at 16% lower throughput when both use one backward call per batch. On long, it fits where all tested alternatives with that backward schedule run out of memory with 80 GB. These comparisons use each implementation's lowest-memory configuration reaching at least 95% of its peak throughput.

Tema degli esploratori: calcolo e infrastrutture (elenco arXiv cs.DC, nuovi lavori del giorno).
Categorie arXiv: Machine Learning (cs.LG) ; Distributed, Parallel, and Cluster Computing (cs.DC); Sound (cs.SD).
Pagina del lavoro: https://arxiv.org/abs/2610.23409 — 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 2026-10-19T08:39:39.773Z dalla pagina https://arxiv.org/list/cs.DC/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).