{"id":2564,"date":"2026-04-23T18:22:10","date_gmt":"2026-04-23T18:22:10","guid":{"rendered":"https:\/\/deepinsightai.io\/?p=2564"},"modified":"2026-04-23T18:22:12","modified_gmt":"2026-04-23T18:22:12","slug":"deepseek-starts-updating-frequently","status":"publish","type":"post","link":"https:\/\/deepinsightai.io\/fr\/deepseek-starts-updating-frequently\/","title":{"rendered":"DeepSeek commence \u00e0 publier des mises \u00e0 jour r\u00e9guli\u00e8res : Tile Kernels et DeepEP V2"},"content":{"rendered":"<p>\u00c0 l'instant, DeepSeek\u2019s <a href=\"https:\/\/deepinsightai.io\/fr\/the-fake-star-economy-on-github\/\">GitHub<\/a> a commenc\u00e9 \u00e0 publier des mises \u00e0 jour r\u00e9guli\u00e8res. Il a lanc\u00e9 et mis en open source un nouveau d\u00e9p\u00f4t, <strong>Noyaux de tuiles<\/strong>, et a par la m\u00eame occasion mis \u00e0 jour le <strong>DeepEP<\/strong> r\u00e9f\u00e9rentiel, apportant <strong>DeepEP V2<\/strong> en ligne. Cela fait moins d'une semaine que DeepSeek a discr\u00e8tement mis \u00e0 jour <strong>Mega MoE<\/strong> et <strong>Indexeur FP4<\/strong> la derni\u00e8re fois.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Noyaux de tuiles DeepSeek<\/h2>\n\n\n\n<figure data-spectra-id=\"spectra-mobt4mso-77si3j\" class=\"wp-block-image aligncenter size-full\"><img fetchpriority=\"high\" decoding=\"async\" width=\"889\" height=\"471\" src=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59.png\" alt=\"Noyaux de tuiles DeepSeek\" class=\"wp-image-2568\" srcset=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59.png 889w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59-300x159.png 300w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59-768x407.png 768w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59-18x10.png 18w\" sizes=\"(max-width: 889px) 100vw, 889px\" \/><\/figure>\n\n\n\n<p>Lien : <code>https:\/\/github.com\/deepseek-ai\/TileKernels<\/code><\/p>\n\n\n\n<p>D'apr\u00e8s l'introduction, <strong>Noyaux de tuiles<\/strong> sont des noyaux GPU optimis\u00e9s pour les op\u00e9rations LLM, d\u00e9velopp\u00e9s avec <strong>TileLang<\/strong>. TileLang est un langage sp\u00e9cifique \u00e0 un domaine permettant d'exprimer des noyaux GPU hautement performants en Python, qui se caract\u00e9rise notamment par une portabilit\u00e9 ais\u00e9e, un d\u00e9veloppement agile et une optimisation automatique.<\/p>\n\n\n\n<p>Les performances des \u201c Tile Kernels \u201d sont exceptionnelles. Comme l\u2019a \u00e9crit DeepSeek lui-m\u00eame : \u00ab La plupart des noyaux de ce projet fr\u00f4lent d\u00e9j\u00e0 les limites de performance du mat\u00e9riel en termes d\u2019intensit\u00e9 de calcul et de bande passante m\u00e9moire. Certains d\u2019entre eux ont d\u00e9j\u00e0 \u00e9t\u00e9 utilis\u00e9s en interne dans des sc\u00e9narios d\u2019entra\u00eenement et d\u2019inf\u00e9rence. Cependant, ils ne constituent pas encore des bonnes pratiques, et nous continuons \u00e0 am\u00e9liorer la qualit\u00e9 du code et la documentation. \u00bb<\/p>\n\n\n\n<p>Le r\u00e9f\u00e9rentiel ne contient pas beaucoup d'informations introductives, mais entre les lignes, il \u201c d\u00e9voile \u201d d\u00e9j\u00e0 la voie de l'innovation architecturale qui sous-tend les mod\u00e8les de nouvelle g\u00e9n\u00e9ration de DeepSeek, laissant entrevoir un bond en avant comparable \u00e0 celui r\u00e9cemment <a target=\"_blank\" rel=\"noreferrer noopener\" href=\"https:\/\/deepinsightai.io\/fr\/hy3-preview-launch\/\">Lancement en avant-premi\u00e8re de Hy3<\/a>.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Fonctionnalit\u00e9s des noyaux de tuiles DeepSeek<\/h3>\n\n\n\n<p>Voici quelques caract\u00e9ristiques sp\u00e9cifiques des noyaux de tuiles :<\/p>\n\n\n\n<p><strong>M\u00e9canisme de d\u00e9clenchement :<\/strong> S\u00e9lection des k meilleurs experts et notation pour le routage par MoE<\/p>\n\n\n\n<p><strong>Acheminement MoE :<\/strong> Mise en correspondance des jetons avec les experts, fusion des op\u00e9rations d'expansion et de r\u00e9duction, et normalisation des poids<\/p>\n\n\n\n<p><strong>Quantification :<\/strong> Prend en charge la conversion FP8\/FP4\/E5M6 en modes \u00ab par token \u00bb, \u00ab par bloc \u00bb et \u00ab par canal \u00bb, et combine les op\u00e9rations SwiGLU et de quantification<\/p>\n\n\n\n<p><strong>Transposer :<\/strong> Op\u00e9rations de transposition par lots<\/p>\n\n\n\n<p><strong>Engramme :<\/strong> Noyaux de d\u00e9clenchement d'engrammes, fusion de RMSNorm, propagation avant\/arri\u00e8re et r\u00e9duction du gradient des poids<\/p>\n\n\n\n<p><strong>Hyperconnexion multiple :<\/strong> Noyaux d'hyperconnexion, notamment la normalisation de Sinkhorn et la technique \u00ab split\/apply \u00bb pour le mixage<\/p>\n\n\n\n<p><strong>Mod\u00e9lisation :<\/strong> De haut niveau <code>torch.autograd.Function<\/code> enveloppes qui regroupent les noyaux sous-jacents en couches pouvant \u00eatre entra\u00een\u00e9es (engram gate, mHC pipeline)<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">DeepSeek EPv2 : une version plus rapide d'EP prenant en charge Engram, PP et CP<\/h2>\n\n\n\n<figure data-spectra-id=\"spectra-mobt54cl-th2xp4\" class=\"wp-block-image aligncenter size-large\"><img decoding=\"async\" width=\"1024\" height=\"656\" data-src=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-1024x656.png\" alt=\"DeepSeek EPv2 : une version plus rapide d&#039;EP prenant en charge Engram, PP et CP\" class=\"wp-image-2569 lazyload\" data-srcset=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-1024x656.png 1024w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-300x192.png 300w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-768x492.png 768w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-18x12.png 18w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60.png 1069w\" data-sizes=\"(max-width: 1024px) 100vw, 1024px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 1024px; --smush-placeholder-aspect-ratio: 1024\/656;\" \/><\/figure>\n\n\n\n<p>Lien EPv2 : <code>https:\/\/github.com\/deepseek-ai\/DeepEP\/pull\/605<\/code><\/p>\n\n\n\n<p>Plus t\u00f4t dans la journ\u00e9e, DeepSeek a \u00e9galement publi\u00e9 la derni\u00e8re version de <strong>EPv2<\/strong>, pour une livraison plus rapide <strong>parall\u00e9lisme expert (EP)<\/strong> et la prise en charge de <strong>Engramme \/ parall\u00e9lisme de pipeline (PP) \/ parall\u00e9lisme de contexte (CP)<\/strong>.<\/p>\n\n\n\n<p>\u00c0 mesure que le mat\u00e9riel, les r\u00e9seaux et les architectures de mod\u00e8les ont \u00e9volu\u00e9 parall\u00e8lement aux lancements rapides de nouveaux produits dans le secteur, tels que <a href=\"https:\/\/deepinsightai.io\/fr\/qwen-3-6\/\" target=\"_blank\" rel=\"noreferrer noopener\">Qwen 3.6<\/a>, La version pr\u00e9c\u00e9dente de DeepSeek, DeepEP V1, avait d\u00e9j\u00e0 accumul\u00e9 trop de probl\u00e8mes h\u00e9rit\u00e9s du pass\u00e9 et trop de probl\u00e8mes de performances.<\/p>\n\n\n\n<p>Cette mise \u00e0 jour restructure enti\u00e8rement <strong>Parall\u00e9lisme avanc\u00e9<\/strong>. Par rapport \u00e0 la version 1, elle ne n\u00e9cessite qu\u2019une fraction des ressources SM pour atteindre des performances exceptionnelles, tout en prenant en charge des applications \u00e0 plus grande \u00e9chelle <strong>mise \u00e0 l'\u00e9chelle<\/strong> (sur une seule machine) et <strong>\u00e9volutivit\u00e9 horizontale<\/strong> (sur diff\u00e9rents ordinateurs).<\/p>\n\n\n\n<p>Par ailleurs, DeepSeek a lanc\u00e9 une fonctionnalit\u00e9 exp\u00e9rimentale <strong>0 SM<\/strong> s\u00e9ries pr\u00e9sent\u00e9es dans cette mise \u00e0 jour, notamment <strong>0 SM Engram<\/strong>, <strong>0 Parall\u00e9lisme du pipeline SM (PP)<\/strong>, et <strong>0 Parall\u00e9lisme contextuel (CP) de SM<\/strong> Op\u00e9rateurs de regroupement. Parall\u00e8lement, le backend a \u00e9t\u00e9 remplac\u00e9 par <strong>NVSHMEM<\/strong> vers le briquet <strong>Gin NCCL<\/strong> backend.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Nouvelles fonctionnalit\u00e9s de DeepSeek DeepEP V2<\/h3>\n\n\n\n<p>Voici quelques-unes des nouvelles fonctionnalit\u00e9s de DeepEP V2 :<\/p>\n\n\n\n<p><strong>Enti\u00e8rement JIT<\/strong><\/p>\n\n\n\n<p><strong>Backend de NCCL Gin :<\/strong><\/p>\n\n\n\n<p>Uniquement l'en-t\u00eate, extr\u00eamement l\u00e9ger<\/p>\n\n\n\n<p>Permet de r\u00e9utiliser les communicateurs NCCL existants<\/p>\n\n\n\n<p><strong>EPv2 :<\/strong><\/p>\n\n\n\n<p>Il regroupe les API \u00e0 haut d\u00e9bit et \u00e0 faible latence au sein d'une interface unique, et adopte une toute nouvelle architecture GEMM.<\/p>\n\n\n\n<p>Prend en charge des domaines de mise \u00e0 l'\u00e9chelle plus vastes, jusqu'\u00e0 <strong>EP2048<\/strong><\/p>\n\n\n\n<p>Introduit un calcul analytique du nombre de SM et de QP, rendant ainsi le r\u00e9glage automatique inutile<\/p>\n\n\n\n<p>Continue \u00e0 prendre en charge les deux <strong>Hybride<\/strong> mode et <strong>Direct<\/strong> mode<\/p>\n\n\n\n<p>Pour les t\u00e2ches d'entra\u00eenement plus anciennes de type V3, l'utilisation des SM diminue, passant de <strong>24<\/strong> \u00e0 <strong>4\u20136<\/strong>, tout en conservant des performances identiques, voire sup\u00e9rieures<\/p>\n\n\n\n<p><strong>0 SM Engram<\/strong> (avec RDMA)<\/p>\n\n\n\n<p><strong>0 SM PP<\/strong> (avec RDMA)<\/p>\n\n\n\n<p><strong>0 SM CP<\/strong> (avec Copy Engine)<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Performances de DeepSeek DeepEP V2<\/h2>\n\n\n\n<p>Une fois la configuration de <strong>DeepSeek-V3<\/strong>, des tests ont \u00e9t\u00e9 effectu\u00e9s avec la nouvelle version, en utilisant les param\u00e8tres suivants : <strong>8 000 jetons par lot<\/strong>, <strong>7168 dimension cach\u00e9e<\/strong>, <strong>Les 8 meilleurs experts<\/strong>, <strong>Communiqu\u00e9 FP8<\/strong>, et <strong>Moissonneuse-batteuse BF16<\/strong>. Les r\u00e9sultats sont les suivants :<\/p>\n\n\n\n<figure data-spectra-id=\"spectra-mobt5q25-blwgn9\" class=\"wp-block-image aligncenter size-full\"><img decoding=\"async\" width=\"650\" height=\"289\" data-src=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-61.png\" alt=\"Performances de DeepSeek DeepEP V2\" class=\"wp-image-2570 lazyload\" data-srcset=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-61.png 650w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-61-300x133.png 300w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-61-18x8.png 18w\" data-sizes=\"(max-width: 650px) 100vw, 650px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 650px; --smush-placeholder-aspect-ratio: 650\/289;\" \/><\/figure>\n\n\n\n<p>Remarque : les r\u00e9sultats indiqu\u00e9s correspondent \u00e0 la bande passante logique. Par exemple, dans le cas de <strong>\u00c9pisode 8, partie 2<\/strong>, le <strong>90 Go\/s<\/strong> La bande passante inclut en effet le trafic entre les GPU locaux (rangs locaux).<\/p>\n\n\n\n<p>Par rapport \u00e0 la version V1, la version V2 atteint jusqu\u2019\u00e0 <strong>1,3 fois les performances maximales<\/strong>, tout en \u00e9conomisant jusqu'\u00e0 <strong>Utilisation des ressources SM \u00d7 4 <\/strong>\u2014 une optimisation essentielle pour rester comp\u00e9titif dans un secteur domin\u00e9 par des g\u00e9ants tels que <a href=\"https:\/\/deepinsightai.io\/fr\/claude-opus-4-7\/\" target=\"_blank\" rel=\"noreferrer noopener\">Claude Opus 4.7<\/a>.<\/p>\n\n\n\n<p>Pour finir, juste un petit conseil pour DeepSeek : d\u00e9p\u00eachez-vous de sortir la version <strong>V4<\/strong> D\u00e9j\u00e0. Tout le monde commence \u00e0 s'impatienter.<\/p>\n\n\n\n<p><\/p>","protected":false},"excerpt":{"rendered":"<p>Just now, DeepSeek\u2019s GitHub started updating frequently. It launched and open-sourced a new repository, Tile Kernels, and at the same time updated the DeepEP repository, bringing DeepEP V2 online. It has been less than a week since DeepSeek quietly updated Mega MoE and FP4 Indexer last time. DeepSeek Tile Kernels Link: https:\/\/github.com\/deepseek-ai\/TileKernels According to the [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":2567,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_seopress_robots_primary_cat":"none","_seopress_titles_title":"%%post_title%%","_seopress_titles_desc":"DeepSeek releases Tile Kernels and DeepEP V2 with faster expert parallelism, 1.3\u00d7 performance boost, and major GPU efficiency gains.","_seopress_robots_index":"","_uag_custom_page_level_css":"","site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[2,10],"tags":[],"class_list":["post-2564","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-news","category-llm"],"uagb_featured_image_src":{"full":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek.png",786,520,false],"thumbnail":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek-150x150.png",150,150,true],"medium":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek-300x198.png",300,198,true],"medium_large":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek-768x508.png",768,508,true],"large":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek.png",786,520,false],"1536x1536":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek.png",786,520,false],"2048x2048":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek.png",786,520,false],"trp-custom-language-flag":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek-18x12.png",18,12,true]},"uagb_author_info":{"display_name":"Claude Carter","author_link":"https:\/\/deepinsightai.io\/fr\/author\/cloud-han03gmail-com\/"},"uagb_comment_info":0,"uagb_excerpt":"Just now, DeepSeek\u2019s GitHub started updating frequently. It launched and open-sourced a new repository, Tile Kernels, and at the same time updated the DeepEP repository, bringing DeepEP V2 online. It has been less than a week since DeepSeek quietly updated Mega MoE and FP4 Indexer last time. DeepSeek Tile Kernels Link: https:\/\/github.com\/deepseek-ai\/TileKernels According to the\u2026","_links":{"self":[{"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/posts\/2564","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/comments?post=2564"}],"version-history":[{"count":1,"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/posts\/2564\/revisions"}],"predecessor-version":[{"id":2571,"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/posts\/2564\/revisions\/2571"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/media\/2567"}],"wp:attachment":[{"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/media?parent=2564"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/categories?post=2564"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/deepinsightai.io\/fr\/wp-json\/wp\/v2\/tags?post=2564"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}