An agent must speed up a language model's attention step on an NVIDIA B200 GPU while keeping results accurate.

DeepSeek has open-sourced six major infrastructure components for Huawei Ascend, including TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. Published benchmarks show some Ascend 950 k...

DeepSeek open-sourced an AI chip programming toolkit for Huawei Ascend on September 30, 2026 — TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect — plus a 128-chip Ascend 950 supernode ...

DeepSeek libera su infraestructura para Huawei Ascend: TileLang, DeepGEMM, FlashMLA y DeepEP. Código abierto que debilita el foso de CUDA y Nvidia.

DeepSeek released Ascend versions of TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect, while Huawei detailed a jointly defined 128-card SuperPoD Flex design.

Which is the best alternative to DualPipe? Based on common mentions it is: Profile-data, EPLB, DeepGEMM, DeepEP or FlashMLA

Which is the best alternative to DualPipe? Based on common mentions it is: Profile-data, EPLB, DeepGEMM, DeepEP or FlashMLA

Which is the best alternative to EPLB? Based on common mentions it is: DualPipe, DeepGEMM, Profile-data, DeepEP or FlashMLA

* * 來自中國的 DeepSeek（深度求索）自 24 日展開為期 5 天的「開源週」，計劃開源 5 個程式庫，希望以完全透明的方式分享該團隊的開發進展，為開源社群做出貢獻。 「開源週」來到第 2 天，DeepSeek 發表 DeepEP，是一套用於 MoE（Mixture of Experts Models，混合專家模型）訓練和推理的開源 EP（expert parallelism）通訊程式庫，它提供高吞吐量和低延遲的全互連 GPU 核心。它還支援低精度運算，包括 FP8。 為了與 DeepSeek-V3 論文提及的演算法一致，DeepEP 提供一組針對非對稱領域頻寬轉發進行最佳化的核心，例如從 NVLink 領域轉發數據到 RDMA 領域，這些核心具有很高的吞吐量，適合訓練和推理預填任務。對於延遲性較敏感的推理解碼，DeepEP 包括一組具有純 RDMA 的低延遲核心，以最大程度減少延遲。 DeepSeek 可說是是一項專門為多 GPU 系統實現高吞吐量、低延遲數據交換所設計的工具，以靈活控制 GPU 資源。 > 🚀 Day 2 of #OpenSourceWeek: DeepEP > > Excited to introduce DeepEP – the first open-source EP communication library for MoE model training and inference. > > ✅ Efficient and optimized all-to-all communication > ✅ Both intranode and internode support with NVLink and RDMA > ✅… > > — DeepSeek (@deepseek_ai) February 25, 2025 **▲ DeepSeek 發表用於 MoE 訓練和推理的開源 EP 通訊程式庫「DeepEP」。** 值得關注還有第 1 天所發表的 FlashMLA，是針對 NVIDIA Hopper 架構 GPU 進行最佳化的高效 MLA（Multi-Head Latent Attention，多頭潛在注意力）解碼核心，涵蓋 H100、H200、H800，不僅將 AI 推理的記憶體頻寬推升至 3000 GB/s 的驚人水準，更有高達 580 TFLOPS 的運算性能。 換個角度來看，FlashMLA 使 GPU 運算方式最佳化、推理速度提升，可使像是 H800 的推理性能提升 2~3 倍。 > 🚀 Day 1 of #OpenSourceWeek: FlashMLA > > Honored to share FlashMLA – our efficient MLA decoding kernel for Hopper GPUs, optimized for variable-length sequences and now in production. > > ✅ BF16 support > ✅ Paged KV cache (block size 64) > ⚡ 3000 GB/s memory-bound & 580 TFLOPS… > > — DeepSeek (@deepseek_ai) February 24, 2025 **▲ 針對 NVIDIA Hopper 架構 GPU 進行最佳化的「FlashMLA」。** （首圖來源：shutterstock） ### 延伸閱讀： * DeepSeek 上線 1 個月下載量破億，未來一週開源 5 個程式庫 文章看完覺得有幫助，何不給我們一個鼓勵 請我們喝杯咖啡 ## 想請我們喝幾杯咖啡？ ### 每杯咖啡 65 元 x 1 x 3 x 5 x 您的咖啡贊助將是讓我們持續走下去的動力 **總金額共新臺幣 0 元** 《關於請喝咖啡的 Q & A》 ### 留給我們的話 取消 確認 從這裡可透過《Google 新聞》追蹤 TechNews * * * 科技新知，時時更新 科技新報粉絲團 加入好友 訂閱免費電子報 * * * 關鍵字: AI , DeepSeek , DeepSeek Open Source Week , 人工智慧

<!-- Share Buttons --> <ul class="socialcount"> <!-- FB (recommend)--> <li class="insidefbcomm"> <a href="javascript:url_open_center('https://www.facebook.com/sharer/sharer.php?u=https://technews.tw/2025/02/24/deepseek-first-open-source-repository-is-flashmla/')"> <img src="https://img.technews.tw/wp-content/themes/twentytwelve/images/FB-share-btn.png"/></a> </li> <!-- Google (recommend)--> <li class="insidegooglecomm"> <a data-callback="gplusoneCallBack" href="https://plus.google.com/share?url=https://technews.tw/2025/02/24/deepseek-first-open-source-repository-is-flashmla/')" onclick="javascript:window.open(this.href,'', 'menubar=no,toolbar=no,resizable=yes,scrollbars=yes,height=600,width=600');return false;"> <img src="https://img.technews.tw/wp-content/themes/twentytwelve/images/Google-share-btn.png"/></a> </li> </ul> <!-- End of Share Buttons --> <div class="bigg"> <img alt="DeepSeek 啟動「開源週」 首個開源代碼庫為 FlashMLA" class="attachment-post-thumbnail size-post-thumbnail wp-post-image" height="533" itemprop="image" src="https://img.technews.tw/wp-content/uploads/2025/02/01113356/shutterstock_2577542989-800x533.jpg" width="800"/> </div> <div class="indent"> <div id="outlineList_top" style="display:none;"></div><p>據中媒報導，中國人工智慧（AI）新創公司 DeepSeek 24 日啟動「開源週」，首個開源資料庫為 FlashMLA， 針對 Hopper GPU 優化的高效 MLA 解碼內核，專為處理可變長度序列而設計。據悉，FlashMLA 的靈感來自 FlashAttention 2&amp;3 和 cutlass 專案。<span id="more-1347839"></span></p><div class="inside_AD no-ad-env" id="inside_AD" style="display:none"><!-- /95946079/(ON)TN_TW_PC_DTL_Content01_970x250 --> <div id="div-gpt-ad-1703223425197-0" style="min-width: 970px; min-height: 250px;"> <script> googletag.cmd.push(function() { googletag.display('div-gpt-ad-1703223425197-0'); }); </script> </div></div> <p>DeepSeek上週<a href="https://technews.tw/2025/02/21/deepseek-downloads/">宣布</a>，接下來一週將陸續開源5個資料庫，每日都有新內容解鎖，「毫無保留分享我們微小但真誠的進展」；當前線上服務的模組已經過測試、部署完備，可投入生產環境。</p> <p>今年1月11日，DeepSeek發布官方App。據數據分析平台QuestMobile最新數據顯示，從上線以來至2月9日，DeepSeek App的累計下載量已超過1.1億次，週活躍用戶規模最高近9,700萬。</p> <p>（本文由 <a href="https://www.moneydj.com//kmdj/news/newsviewer.aspx?a=7ec0b048-1640-48d1-aea5-c54633231425">MoneyDJ新聞</a> 授權轉載；首圖來源：shutterstock）</p> <h3>延伸閱讀：</h3> <ul> <li><a href="https://technews.tw/2025/02/21/deepseek-downloads/">DeepSeek 上線 1 個月下載量破億 未來一週開源 5 個代碼庫</a></li> <li><a href="https://technews.tw/2025/02/19/grok-3-vs-chatgpt-vs-deepseek/">馬斯克推地表最聰明 AI？到底誰最強三大模型比一比</a></li> </ul> </div> <!-- 20240612-請我們喝杯咖啡 btn 移植手機版按鈕--> <div class="coffee-btn-wrapper" style="display:none;"> <p>文章看完覺得有幫助，何不給我們一個鼓勵</p> <div class="coffee-btn-content"> <a class="coffee-btn" data-post-id="1347839" data-post-title="DeepSeek 啟動「開源週」 首個開源代碼庫為 FlashMLA" id="bmc-open-btn" rel="nofollow"> <span>請我們喝杯咖啡</span> <img alt="icon-coffee" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/bymeacoffee/icon-coffee01.png"/> </a> </div> </div> <!-- 請我們喝咖啡 popup model --> <div class="tn-modal" id="bmc-tn-modal"> <div class="tn-modal-content"> <div class="tn-modal-content-dash"> <div class="tn-modal-title coffee"> <h2>想請我們喝幾杯咖啡？</h2> </div> <div class="coffee-detail-text"> <img alt="icon-tag" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/bymeacoffee/icon-tag.png"/> <h3>每杯咖啡 65 元</h3> </div> <div class="coffee-list"> <div class="coffee-item" data-coffee-value="1"> <img alt="icon-coffee" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/bymeacoffee/icon-coffee01.png"/> <span class="mark">x</span> <span class="amount">1</span> </div> <div class="coffee-item" data-coffee-value="3"> <img alt="icon-coffee" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/bymeacoffee/icon-coffee01.png"/> <span class="mark">x</span> <span class="amount">3</span> </div> <div class="coffee-item" data-coffee-value="5"> <img alt="icon-coffee" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/bymeacoffee/icon-coffee01.png"/> <span class="mark">x</span> <span class="amount">5</span> </div> <div class="coffee-item tn-modal-form" data-coffee-value="0"> <img alt="icon-coffee" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/bymeacoffee/icon-coffee01.png"/> <span class="mark">x</span> <input class="tn-modal-form-control" id="bmc-custom-value" max="100" maxlength="3" min="1" type="number"/> </div> </div> <div class="coffee-detail-content-text"> <p>您的咖啡贊助將是讓我們持續走下去的動力</p> </div> <div class="coffee-detail-text"> <b>總金額共新臺幣 <span id="bmc-coffee-total">0</span> 元</b> </div> <a class="tn-modal-link" href="https://technews.tw/aboutus/buy-me-a-coffee-policy/" target="_blank">《關於請喝咖啡的 Q &amp; A》</a> </div> <div class="hr"></div> <div class="tn-modal-content-dash"> <div class="tn-modal-form"> <label for="message"><h3 class="tn-modal-title-sub">留給我們的話</h3></label> <textarea class="tn-modal-form-control" cols="30" id="bmc-message" maxlength="200" placeholder="歡迎留言給我們，給我們更多支持與鼓勵！" rows="5" style="min-width:500px;"></textarea> </div> </div> <div class="tn-modal-content-btn"> <a class="tn-modal-btn" href="javascript:void(0)" id="bmc-close-btn" rel="nofollow">取消</a> <a class="tn-modal-btn" href="javascript:void(0)" id="bmc-submit-btn" rel="nofollow">確認</a> </div> </div> </div> <!-- 20231204-請我們喝杯咖啡 JS --> <script> const bmcOpenBtn = document.querySelector("#bmc-open-btn"); const bmcCloseBtn = document.querySelector("#bmc-close-btn"); const bmcModal = document.querySelector("#bmc-tn-modal"); const bmcSubmitBtn = document.querySelector("#bmc-submit-btn"); const bmcCustomInput = document.getElementById('bmc-custom-value'); bmcOpenBtn.addEventListener('click',function(e){ e.preventDefault(); bmcModal.classList.add('open') document.body.classList.add('modal-open') }); bmcCloseBtn.addEventListener('click',function(e){ e.preventDefault(); bmcModal.classList.remove('open') document.body.classList.remove('modal-open') }); ['keyup', 'change'].forEach(function(event) { bmcCustomInput.addEventListener(event, function() { this.value = this.value.replace(/[^\d]/g, ''); tnChangeCoffeeTotal(); }); }); bmcSubmitBtn.addEventListener('click',function(e){ e.preventDefault(); const activeCoffeeItem = document.querySelector('.coffee-item.active'); if (!activeCoffeeItem) { alert('請選擇杯數'); return; } let coffeeValue = activeCoffeeItem.getAttribute('data-coffee-value'); if (coffeeValue == 0) { coffeeValue = bmcCustomInput.value; if (coffeeValue == 0 || coffeeValue == '') { alert('請輸入杯數'); return; } let isInteger = /^\d+$/.test(coffeeValue); if (!isInteger) { alert('杯數錯誤'); return; } } const bmcRequestParam = new URLSearchParams({ post_id: bmcOpenBtn.getAttribute("data-post-id"), post_title: bmcOpenBtn.getAttribute("data-post-title"), post_url: document.querySelector('link[rel="canonical"]').getAttribute('href'), quantity: coffeeValue, message: document.getElementById('bmc-message').value, }); location.href = 'https://member.technews.tw/buymeacoffee?' + bmcRequestParam }); let bmcItems = document.querySelectorAll('.coffee-item'); bmcItems.forEach(function(item){ item.addEventListener('click',function(e){ e.preventDefault(); if (item.dataset.coffeeValue == 0) { bmcCustomInput.focus(); } bmcItems.forEach(function(item){ item.classList.remove('active') }) item.classList.add('active') tnChangeCoffeeTotal(); }) }); function tnChangeCoffeeTotal() { const activeCoffeeItem = document.querySelector('.coffee-item.active'); let coffeeValue = activeCoffeeItem.getAttribute('data-coffee-value'); if (coffeeValue == 0) { coffeeValue = bmcCustomInput.value; } document.getElementById('bmc-coffee-total').innerText = coffeeValue * 65; } function tnDonateDisplay() { const tnUser = tn_get_user_cookie(); let dataSet = jQuery('#lockPost').data(); //文章資訊 if ('undefined' != typeof dataSet && 'freeMember' == dataSet.member_level) { if (tnUser.is_login && tnUser.permission < vipPermission) { jQuery('.coffee-btn-wrapper').show(); } } else { jQuery('.coffee-btn-wrapper').show(); } } jQuery(document).ready(function () { tnDonateDisplay(); }); </script> <!-- 增加Google News --> <div class="googlenews_Content"> <a href="https://news.google.com/publications/CAAqBwgKMJfS3gow8arWAQ" target="_blank"> <p>從這裡可透過《Google 新聞》追蹤 TechNews</p> <img alt="Google News" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/GoogleNews_banner.png"/> </a> </div> <div class="no-ad-env" style="display:none"> <script async="" data-cfasync="false" type="text/javascript"> var per = atob(getCookie('per')); if(per == null || per < 60){ var truvidScript = document.createElement('script'); truvidScript.async = true; truvidScript.setAttribute('data-cfasync','false'); truvidScript.type = 'text/javascript'; truvidScript.src = '//stg.truvidplayer.com/index.php?sub_user_id=597&widget_id=2933&playlist_id=2112&m=a&cb=' + (Math.random() * 10000000000000000); var currentScript = document.currentScript || document.scripts[document.scripts.length - 1]; currentScript.parentNode.insertBefore(truvidScript, currentScript.nextSibling); console.log('PC Truvid 廣告') } </script> </div> <!-- 增加 訂閱電子報內文欄位 by nick --> <div class="newsLetter_articleContent newsLetter_articleColumn"> <hr class="articleContent_hr"/> <div class="logo"><img src="https://img.technews.tw/wp-content/themes/twentytwelve/images/edm/articleContent_img01.png"/></div> <p class="articleContent_text">科技新知，時時更新</p> <div class="btnGroup"> <a class="articleContentBtn socialBtn" href="https://www.facebook.com/technewsinside"> <img alt="科技新報粉絲團" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/fbLike_icon.svg"/>科技新報粉絲團 </a> <a class="articleContentBtn lineBtn" data-v-4185e1d6="" href="https://lin.ee/nuCjuzt"> <img alt="加入好友" border="0" data-v-4185e1d6="" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/lineAt_icon.svg"/>加入好友 </a> <a class="articleContentBtn newsLetterBtn" href="https://member.technews.tw/user/newsPaperManager" id="contentSubscribe" target="_blank"> <img alt="訂閱免費電子報" src="https://img.technews.tw/wp-content/themes/twentytwelve/images/subscribeEDM_icon.svg"/>訂閱免費電子報 </a> </div> <hr class="articleContent_hr"/> </div> <!-- 增加 訂閱電子報內文欄位 --> <div> <!-- 廣告開關中的PC popin MaxwWu 修正註解敘述 20220727--> <!-- <div class="popIn"> <div class="wrapper rps_wrapper" id="rps_hcarousel"> </div> </div> --> <div id="_popIn_recommend"></div> <script type="text/javascript"> (function() { var pa = document.createElement('script'); pa.type = 'text/javascript'; pa.charset = "utf-8"; pa.async = true; pa.src = window.location.protocol + "//api.popin.cc/searchbox/technews_tw.js"; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(pa, s); })(); </script> </div> <div class="article-tags-area">關鍵字: <a href="https://technews.tw/tag/ai/" rel="tag">AI</a> , <a href="https://technews.tw/tag/deepseek/" rel="tag">DeepSeek</a> , <a href="https://technews.tw/tag/flashmla/" rel="tag">FlashMLA</a> , <a href="https://technews.tw/tag/%e9%96%8b%e6%ba%90%e8%b3%87%e6%96%99%e5%ba%ab/" rel="tag">開源資料庫</a></div><!-- .entry-meta -->

Contribute to deepseek-ai/FlashMLA development by creating an account on GitHub.

Contribute to deepseek-ai/FlashMLA development by creating an account on GitHub.

Contribute to deepseek-ai/FlashMLA development by creating an account on GitHub.