{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/blockwise-parallel-decoding-for-deep","title":"Blockwise Parallel Decoding for Deep Autoregressive Models","arxiv_id":"1811.03115","date":"2018-11-07","proceeding":"NeurIPS 2018 12","authors":["Mitchell Stern","Noam Shazeer","Jakob Uszkoreit"],"abstract":"Deep autoregressive sequence-to-sequence models have demonstrated impressive\nperformance across a wide variety of tasks in recent years. While common\narchitecture classes such as recurrent, convolutional, and self-attention\nnetworks make different trade-offs between the amount of computation needed per\nlayer and the length of the critical path at training time, generation still\nremains an inherently sequential process. To overcome this limitation, we\npropose a novel blockwise parallel decoding scheme in which we make predictions\nfor multiple time steps in parallel then back off to the longest prefix\nvalidated by a scoring model. This allows for substantial theoretical\nimprovements in generation speed when applied to architectures that can process\noutput sequences in parallel. We verify our approach empirically through a\nseries of experiments using state-of-the-art self-attention models for machine\ntranslation and image super-resolution, achieving iteration reductions of up to\n2x over a baseline greedy decoder with no loss in quality, or up to 7x in\nexchange for a slight decrease in performance. In terms of wall-clock time, our\nfastest models exhibit real-time speedups of up to 4x over standard greedy\ndecoding.","url_abs":"http://arxiv.org/abs/1811.03115v1","url_pdf":"http://arxiv.org/pdf/1811.03115v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"blockwise-parallel-decoding-for-deep","repo_url":"https://github.com/aynetdia/mtp_curriculum","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"decoder","task_name":"Decoder"},{"task_slug":"image-super-resolution","task_name":"Image Super-Resolution"},{"task_slug":"machine-translation","task_name":"Machine Translation"},{"task_slug":"super-resolution","task_name":"Super-Resolution"},{"task_slug":"translation","task_name":"Translation"}],"methods":[{"method_slug":"speed","method_name":"SPEED"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1811.03115","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}