{"id":122715,"date":"2026-11-05T14:24:58","date_gmt":"2026-11-05T06:24:58","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/122715.html"},"modified":"2026-11-05T14:24:58","modified_gmt":"2026-11-05T06:24:58","slug":"openash-v3-5-%e6%9e%b6%e6%9e%84%e8%bf%ad%e4%bb%a3%e6%8a%a5%e5%91%8a%ef%bc%9a%e8%a1%b0%e5%87%8f%e6%9c%80%e5%a4%a7%e5%80%bc%e4%b8%89%e9%80%9a%e9%81%93%e6%b7%b7%e5%90%88%e6%9e%b6%e6%9e%84","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/122715.html","title":{"rendered":"OpenASH v3.5 \u67b6\u6784\u8fed\u4ee3\u62a5\u544a\uff1a\u8870\u51cf\u6700\u5927\u503c\u4e09\u901a\u9053\u6df7\u5408\u67b6\u6784"},"content":{"rendered":"<p>        \/* MD \/ \u5bcc\u6587\u672c\u4e2d\u7684 .toc\uff08\u542b\u535a\u5ba2\u56ed\u642c\u5bb6\u7b49\u5d4c\u5957\u7ed3\u6784\uff09\uff1b.toc-box \u5728\u4fa7\u680f\uff0c\u4e0d\u53d7\u5f71\u54cd *\/<br \/>\n        #content_views .toc,<br \/>\n        \/* \u7f16\u8f91\u5668\u5e38\u5728\u76ee\u5f55\u524d\u540e\u63d2\u5165\u7a7a p\uff08:empty \u4ecd\u5360 20px\uff09\uff0c\u4e00\u5e76\u53bb\u6389\u907f\u514d\u9876\u7a7a\u9699 *\/<br \/>\n        #content_views.markdown_views &gt; p:empty:has(+ .toc),<br \/>\n        #content_views.markdown_views &gt; .toc + p:empty,<br \/>\n        \/* \u5bcc\u6587\u672c\u65e7\u7248\u76ee\u5f55\u6807\u8bb0 *\/<br \/>\n        #content_views.htmledit_views #main-toc,<br \/>\n        #content_views.htmledit_views #hr-toc,<br \/>\n        #content_views.htmledit_views p[id*=\\&#8221;-toc\\&#8221;] {<br \/>\n          display: none !important;<br \/>\n        }<br \/>\n        \/* \u76ee\u5f55\u53bb\u6389\u540e\uff0c\u7d27\u8ddf\u7684\u9996\u4e2a\u6807\u9898\u4e0d\u518d\u591a\u51fa\u4e00\u5757\u4e0a\u8fb9\u8ddd *\/<br \/>\n        #content_views.markdown_views &gt; .toc + h1,<br \/>\n        #content_views.markdown_views &gt; .toc + h2,<br \/>\n        #content_views.markdown_views &gt; .toc + h3,<br \/>\n        #content_views.markdown_views &gt; .toc + h4,<br \/>\n        #content_views.markdown_views &gt; .toc + p:empty + h1,<br \/>\n        #content_views.markdown_views &gt; .toc + p:empty + h2,<br \/>\n        #content_views.markdown_views &gt; .toc + p:empty + h3,<br \/>\n        #content_views.markdown_views &gt; .toc + p:empty + h4 {<br \/>\n          margin-top: 0 !important;<br \/>\n        }<\/p>\n<p>\u7ed3\u8bba&#xff1a;\u540c\u9884\u7b97&#xff08;15000 \u6b65 \u00d7 184M tokens\u300195-105M \u53c2\u6570&#xff09;\u4e0b&#xff0c;v3.5 \u5728\u5168\u90e8 7 \u4e2a\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0a\u8d85\u8fc7\u540c\u9884\u7b97\u73b0\u4ee3 Transformer \u57fa\u7ebf&#xff08;RoPE &#043; RMSNorm &#043; SwiGLU &#043; FlashAttention&#xff09;&#xff0c;\u603b\u4f53 PPL \u9886\u5148 10.7%&#xff0c;2048 \u957f\u5ea6\u533a\u9886\u5148 58.5%&#xff0c;\u540c\u65f6\u4fdd\u6301 O(S) \u8bad\u7ec3\u590d\u6742\u5ea6\u4e0e O(1) \u63a8\u7406\u72b6\u6001\u3002<\/p>\n<hr \/>\n<h3>1. \u67b6\u6784\u6982\u8981<\/h3>\n<p>v3.5 \u7684\u6bcf\u5c42\u6ce8\u610f\u529b\u7531\u4e09\u6761 O(1) \u72b6\u6001\u901a\u9053\u6df7\u5408\u6784\u6210&#xff0c;\u5168\u90e8\u57fa\u4e8e\u6bcf\u5c42\u4e13\u5c5e\u7684\u4f4e\u79e9\u72b6\u6001\u5d4c\u5165 state_em(23004\u00d764) &#064; state_proj(64\u2192768)&#xff08;c \u6d41&#xff0c;\u4e00\u6b21\u67e5\u8868\u4e09\u5904\u590d\u7528&#xff09;&#xff1a;<\/p>\n<table>\n<tr>\u901a\u9053\u9012\u63a8\u68c0\u7d22\u7c7b\u578b\u8bb0\u5fc6\u5c3a\u5ea6&#xff08;\u5b9e\u6d4b&#xff09;\u4f5c\u7528<\/tr>\n<tbody>\n<tr>\n<td>\u5feb max-scan<\/td>\n<td>e_j &#061; max(\u03b3_j\u00b7e_{j-1}, c_j)&#xff0c;\u03b3 &#061; exp(\u2212softplus(c\u00b7v&#043;b))<\/td>\n<td>query \u65e0\u5173&#xff08;\u65b0\u9c9c\u5cf0\u503c\u8ddf\u8e2a&#xff09;<\/td>\n<td>\u534a\u8870\u671f ~4 tok<\/td>\n<td>\u5c40\u90e8\u4fe1\u53f7\u3001\u6574\u4f53\u8d28\u91cf<\/td>\n<\/tr>\n<tr>\n<td>\u6162 max-scan<\/td>\n<td>\u540c\u4e0a&#xff0c;\u03b3 \u2208 [0.995, 0.9995]&#xff08;\u5c42 6-11&#xff09;<\/td>\n<td>query \u65e0\u5173&#xff08;\u957f\u65f6\u4fdd\u6301&#xff09;<\/td>\n<td>\u534a\u8870\u671f ~245 tok<\/td>\n<td>\u8fdc\u7a0b\u4fe1\u606f\u4fdd\u6301\u3001\u957f\u5ea6\u5916\u63a8<\/td>\n<\/tr>\n<tr>\n<td>KV \u68c0\u7d22<\/td>\n<td>S_j &#061; a_j\u00b7S_{j-1} &#043; \u03b2\u00b7k\u0302_j\u2297c_j&#xff0c;\u8bfb\u51fa r_j &#061; q_j\u00b7S_j<\/td>\n<td>query \u76f8\u5173&#xff08;k\u0302 L2 \u5f52\u4e00\u5316&#xff09;<\/td>\n<td>a\u22480.98 \u53ef\u5b66\u4e60<\/td>\n<td>\u5f52\u7eb3\u5934\/\u7cbe\u786e\u952e\u503c\u53ec\u56de<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4e09\u901a\u9053\u8f93\u51fa\u5728 gen_model \u591a\u9879\u5f0f\u878d\u5408&#xff08;Triton kernel&#xff09;&#xff1a;head(cat[a,b,c,d,e])\u00b7e &#043; a\u00b7b &#043; \u03b1\u2081\u00b7b &#043; \u03b1\u2082\u00b7d &#043; a\u00b7(\u03b1\u2083\u00b7e&#043;d) &#043; b\u00b7(c&#043;e) &#043; c\u00b7e &#043; kv_gate\u00b7r\u3002<\/p>\n<p>\u914d\u5957\u73b0\u4ee3\u5316&#xff1a;pre-norm RMSNorm \u00d72\u3001SiLU-GLU FFN\u3001final RMSNorm\u3001max_seq_len&#061;2048\u3001\u5206\u5c42\u591a\u5c3a\u5ea6 \u03b3 \u521d\u59cb\u5316&#xff08;v3.3&#xff09;\u3002<\/p>\n<h3>2. \u6f14\u8fdb\u94fe\u4e0e\u6d88\u878d\u6570\u636e&#xff08;\u5168\u90e8\u540c\u9884\u7b97 A\/B&#xff0c;seed\/\u6570\u636e\u987a\u5e8f\u5bf9\u9f50&#xff09;<\/h3>\n<table>\n<tr>\u7248\u672c\u5173\u952e\u6539\u52a8val PPL\u76f8\u5bf9\u4e0a\u4e00\u7248<\/tr>\n<tbody>\n<tr>\n<td>v3.0 pure cummax<\/td>\n<td>\u65e0\u8870\u51cf<\/td>\n<td>84.53<\/td>\n<td>\u2014<\/td>\n<\/tr>\n<tr>\n<td>v3.1 decay-max<\/td>\n<td>\u8870\u51cf\u626b\u63cf &#043; Triton \u8bad\u7ec3\/\u63a8\u7406 kernel<\/td>\n<td>59.07<\/td>\n<td>\u221230.1%<\/td>\n<\/tr>\n<tr>\n<td>v3.2 \u73b0\u4ee3\u5316<\/td>\n<td>pre-norm RMSNorm &#043; SiLU &#043; \u03b3 \u6162\u5c3e &#043; \u8f93\u5165\u76f8\u5173\u8870\u51cf &#043; pos 2048<\/td>\n<td>46.90<\/td>\n<td>\u221220.6%<\/td>\n<\/tr>\n<tr>\n<td>v3.3 \u5206\u5c42\u6162\u901a\u9053<\/td>\n<td>\u5c42 6-11 \u6162 \u03b3&#xff08;245 tok \u534a\u8870\u671f\u5b58\u6d3b&#xff09;<\/td>\n<td>46.80<\/td>\n<td>\u6301\u5e73&#xff08;\u957f\u7a0b &#043;&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>v3.4 &#043;KV \u901a\u9053<\/td>\n<td>query \u76f8\u5173\u68c0\u7d22&#xff08;sum \u5199\u5165&#xff09;<\/td>\n<td>41.17<\/td>\n<td>\u221212.0%<\/td>\n<\/tr>\n<tr>\n<td>v3.5 \u5f52\u4e00\u5316\u952e<\/td>\n<td>k\u0302 L2 \u5f52\u4e00\u5316 &#043; per-head \u5199\u5165\u95e8<\/td>\n<td>40.27<\/td>\n<td>\u22122.2%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>3. \u4e3b\u5b9e\u9a8c\u4e00&#xff1a;2500 \u6b65\u5c0f\u6570\u636e&#xff08;11.5M tokens&#xff09;\u5bf9\u51b3<\/h3>\n<p>v3.5 37.05 vs tf 37.03 \u2014\u2014 \u603b\u4f53\u8ffd\u5e73&#xff08;\u00b10.05%&#xff09;\u3002\u957f\u5ea6\u6269\u5c55\u4e0a v3.5 \u4ece 512 \u8d77\u53cd\u8d85&#xff08;512: 37.01 vs 39.32&#xff09;\u3002<\/p>\n<h3>4. \u4e3b\u5b9e\u9a8c\u4e8c&#xff1a;15000 \u6b65\u5927\u6570\u636e&#xff08;488M tokens \u6570\u636e\u96c6\u3001184M \u8bad\u7ec3 tokens&#xff09;\u2014\u2014\u6700\u7ec8\u88c1\u51b3<\/h3>\n<table>\n<tr>\u4e0a\u4e0b\u6587 Lv3.5 PPLtf PPLv3.5 \u9886\u5148<\/tr>\n<tbody>\n<tr>\n<td>64<\/td>\n<td>18.56<\/td>\n<td>19.47<\/td>\n<td>&#043;4.7%<\/td>\n<\/tr>\n<tr>\n<td>128<\/td>\n<td>17.03<\/td>\n<td>18.05<\/td>\n<td>&#043;5.7%<\/td>\n<\/tr>\n<tr>\n<td>256<\/td>\n<td>16.36<\/td>\n<td>17.69<\/td>\n<td>&#043;7.5%<\/td>\n<\/tr>\n<tr>\n<td>384<\/td>\n<td>16.01<\/td>\n<td>17.93<\/td>\n<td>&#043;10.7%<\/td>\n<\/tr>\n<tr>\n<td>512<\/td>\n<td>15.92<\/td>\n<td>19.01<\/td>\n<td>&#043;16.3%<\/td>\n<\/tr>\n<tr>\n<td>1024<\/td>\n<td>15.71<\/td>\n<td>25.55<\/td>\n<td>&#043;38.5%<\/td>\n<\/tr>\n<tr>\n<td>2048<\/td>\n<td>15.64<\/td>\n<td>37.68<\/td>\n<td>&#043;58.5%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<ul>\n<li>\u4e03\u4e2a\u957f\u5ea6\u5168\u90e8\u9886\u5148&#xff1b;v3.5 \u66f2\u7ebf\u968f\u957f\u5ea6\u6301\u7eed\u4e0b\u964d&#xff08;42.2\u219215.6&#xff09;&#xff0c;tf \u4ece 512 \u8d77\u5d29\u6e83&#xff08;\u65e0\u4f4d\u7f6e\u5916\u63a8\u9002\u914d&#xff09;<\/li>\n<li>\u8bad\u7ec3\u8f68\u8ff9\u5448\u6559\u79d1\u4e66\u5f0f\u4ea4\u53c9&#xff1a;\u65e9\u671f tf \u7565\u9886\u5148&#xff0c;step ~8000 \u8d77\u4ea4\u53c9&#xff0c;\u4e4b\u540e v3.5 \u6301\u7eed\u62c9\u5f00\u4e14\u7ec8\u70b9\u672a\u9971\u548c<\/li>\n<li>\u8fdc\u7a0b\u590d\u5236&#xff08;\u91cd\u590d span \u63a2\u9488&#xff09;&#xff1a;\u7edd\u5bf9 CE \u5168\u7a0b\u66f4\u4f4e&#xff08;3.13~4.12 vs 3.32~5.11&#xff09;&#xff1b;gain \u53e3\u5f84 gap&#061;128 \u53cd\u8d85&#xff08;&#043;0.74 vs &#043;0.58&#xff09;&#xff0c;gap&#061;32 tf \u4fdd\u7559\u4f18\u52bf&#xff08;&#043;1.28 vs &#043;0.89&#xff09;<\/li>\n<li>\u4e0a\u4e0b\u6587\u5229\u7528\u66f2\u7ebf&#xff1a;\u5168\u6876\u9886\u5148&#xff0c;tf \u5c3e\u6876\u6076\u5316&#xff08;3.01&#xff09;\u800c v3.5 \u4fdd\u6301&#xff08;2.79&#xff09;<\/li>\n<\/ul>\n<h3>5. \u957f\u7a0b\u80fd\u529b\u63a2\u9488\u5957\u4ef6&#xff08;bench_longrange.py&#xff09;<\/h3>\n<li>\u957f\u5ea6\u6269\u5c55 PPL&#xff1a;64\u21922048 \u4e03\u6863&#xff0c;&gt;512 \u4e3a\u4f4d\u7f6e\u5916\u63a8\u533a&#xff08;\u53ef\u5b66\u4e60\u4f4d\u7f6e\u7f16\u7801 clamp \u9971\u548c&#xff0c;\u4e13\u6d4b scan \u72b6\u6001\u4fdd\u6301&#xff09;<\/li>\n<li>\u8fdc\u7a0b\u590d\u5236\u53ec\u56de&#xff1a;[prefix][span A][gap G][span A] \u91cd\u590d\u6bb5 CE \u4e0b\u964d\u91cf&#xff0c;gap \u5373\u53d7\u63a7\u8bb0\u5fc6\u8ddd\u79bb<\/li>\n<li>\u5206\u4f4d\u7f6e loss \u66f2\u7ebf&#xff1a;8 \u6876\u4e0a\u4e0b\u6587\u5229\u7528\u66f2\u7ebf<\/li>\n<li>\u03b3 \u6709\u6548\u8bb0\u5fc6\u957f\u5ea6&#xff1a;\u771f\u5b9e\u8bed\u6599\u4e0a \u03b3 \u5206\u5e03\u7edf\u8ba1&#xff08;v3.5 \u5b9e\u6d4b&#xff1a;\u5feb\u5c42 4.1 tok \/ \u6162\u5c42 245 tok&#xff0c;\u5feb\u6162\u5206\u5de5\u88ab\u8bad\u7ec3\u4fdd\u7559&#xff09;<\/li>\n<h3>6. \u5de5\u7a0b\u5b9e\u73b0\u8981\u70b9<\/h3>\n<ul>\n<li>\u8bad\u7ec3 297 ms\/step&#xff08;bf16 autocast&#xff0c;bs&#061;32\u00d7S&#061;384&#xff1b;KV Triton \u524d\u5411 0.42 ms\/\u5c42 &#043; torch \u91cd\u7b97\u53cd\u5411 4.3 ms\/\u5c42&#xff09;<\/li>\n<li>O(1) \u53cc\u72b6\u6001\u589e\u91cf\u63a8\u7406&#xff1a;state \u5143\u7ec4\u643a\u5e26 (max-scan e \u5411\u91cf, KV \u77e9\u9635 S)&#xff1b;fp32 \u4e0b\u9010 token \u89e3\u7801 vs \u5168\u5e8f\u5217\u524d\u5411 diff &#061; 3.7e-06<\/li>\n<li>\u663e\u5b58&#xff1a;\u8bad\u7ec3\u5cf0\u503c 12.4 GiB&#xff08;bs&#061;32&#xff09;&#xff0c;DecayScanFn \u7cfb\u6570\u53cd\u5411\u91cd\u7b97\u7701 1.8 GB<\/li>\n<li>\u7a33\u5b9a\u6027\u53d1\u73b0&#xff08;\u91cd\u8981&#xff09;&#xff1a;\u672c\u73af\u5883 Triton \u5bf9&#034;\u5faa\u73af\u643a\u5e26\u7d2f\u52a0\u5bc4\u5b58\u5668 &#043; \u5faa\u73af\u51fa\u53e3 store&#034;\u5b58\u5728 codegen \u975e\u786e\u5b9a\u6027&#xff08;miscompilation \u5acc\u7591&#xff09;&#xff1b;\u5faa\u73af\u5185\u9010\u6b65 store \u6a21\u5f0f\u5b8c\u5168\u7a33\u5b9a\u2014\u2014\u6240\u6709 kernel \u9075\u5faa\u8be5\u6a21\u5f0f\u8bbe\u8ba1&#xff0c;\u591a\u5f62\u72b6 \u00d7 \u591a\u6b21\u53d1\u5c04\u96f6\u6f02\u79fb\u9a8c\u8bc1<\/li>\n<li>\u6570\u503c\u5b89\u5168&#xff1a;max-scan \u65e0\u9664\u6cd5&#xff1b;KV \u5206\u5757\u6743\u91cd C[j,i] &#061; exp(l_j \u2212 l_i) \u56e0\u679c\u63a9\u7801\u540e \u2264 1&#xff0c;\u65e0\u9664\u6cd5\u65e0\u4e0b\u6ea2<\/li>\n<\/ul>\n<h3>7. \u590d\u73b0\u6307\u5357<\/h3>\n<p><span class=\"token comment\"># \u540c\u9884\u7b97 A\/B&#xff08;\u67b6\u6784\u53ef\u9009 v30\/v31\/v32\/v33\/v34\/v35\/v35d\/tf&#xff09;<\/span><br \/>\npython train_ab.py <span class=\"token parameter variable\">&#8211;arch<\/span> v35 <span class=\"token parameter variable\">&#8211;steps<\/span> <span class=\"token number\">3000<\/span> <span class=\"token parameter variable\">&#8211;out<\/span> ab_v35_3k<br \/>\n<span class=\"token comment\"># \u5927\u6570\u636e\u7f29\u653e<\/span><br \/>\npython train_ab.py <span class=\"token parameter variable\">&#8211;arch<\/span> v35 <span class=\"token parameter variable\">&#8211;steps<\/span> <span class=\"token number\">15000<\/span> <span class=\"token parameter variable\">&#8211;warmup<\/span> <span class=\"token number\">500<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n    <span class=\"token parameter variable\">&#8211;data<\/span> minimind_data\/pretrain_cached_1270238_384.pt <span class=\"token parameter variable\">&#8211;out<\/span> big_v35<br \/>\n<span class=\"token comment\"># \u957f\u7a0b\u56db\u63a2\u9488&#xff08;\u7f16\u8f91 MODEL_KEYS \u9009\u62e9\u5bf9\u6bd4\u6a21\u578b&#xff09;<\/span><br \/>\npython bench_longrange.py<br \/>\n<span class=\"token comment\"># \u6b63\u786e\u6027\u56de\u5f52&#xff08;58 \u9879&#xff1a;max-scan \u524d\u5411\/\u53cd\u5411\u3001gen_model\u3001fusedCE\u3001KV scan&#xff09;<\/span><br \/>\npython bench_v3_triton.py<br \/>\n<span class=\"token comment\"># \u589e\u91cf\u63a8\u7406\u7b49\u4ef7\u6027<\/span><br \/>\npython <span class=\"token operator\">&lt;<\/span>\u89c1 test_incremental.py \u6a21\u5f0f<span class=\"token operator\">&gt;<\/span>  <span class=\"token comment\"># fp32 \u5168\u5e8f\u5217 vs \u9010 token&#xff1a;3.7e-06<\/span><\/p>\n<p>\u4ea7\u7269&#xff1a;longrange_results.json&#xff08;\u56db\u63a2\u9488\u539f\u59cb\u6570\u636e&#xff09;\u3001\u5404 ab_*\/big_*\/train_log.json&#xff08;\u8bad\u7ec3\u66f2\u7ebf&#xff09;\u3002<\/p>\n<h3>8. \u8fb9\u754c\u4e0e\u540e\u7eed\u65b9\u5411<\/h3>\n<li>\u9a8c\u8bc1\u5c3a\u5ea6\u4e3a 488M tokens \/ 96-105M \u53c2\u6570\u2014\u2014\u76f8\u5bf9\u771f\u5b9e LLM \u9884\u8bad\u7ec3\u4ecd\u5c0f&#xff1b;\u66f4\u5927\u89c4\u6a21\u9700\u6269\u5c55\u7b97\u529b<\/li>\n<li>delta-rule \u5206\u5757&#xff08;UT\/WY \u53d8\u6362&#xff09;\u4e0e KV scan \u624b\u5de5 Triton \u53cd\u5411\u662f\u660e\u786e\u7684\u4e0b\u4e00\u7ea7\u4f18\u5316&#xff08;\u5f53\u524d delta \u9010\u6b65\u7248\u4e0d\u53ef\u8bad\u7ec3\u3001\u91cd\u7b97\u53cd\u5411 ~4.3 ms\/\u5c42&#xff09;<\/li>\n<li>gap&#061;128 \u4e2d\u8ddd\u68c0\u7d22\u4ecd\u843d\u540e tf&#xff08;&#043;0.25 vs &#043;0.62 &#064;3000 \u6b65\u7248&#xff1b;\u5927\u6570\u636e\u7248 gap&#061;128 \u5df2\u53cd\u8d85&#xff09;\u2014\u2014delta \u5199\u5165\u9884\u671f\u8fdb\u4e00\u6b65\u6269\u5927\u4f18\u52bf<\/li>\n<li>tf \u57fa\u7ebf\u672a\u505a\u4f4d\u7f6e\u5916\u63a8\u9002\u914d&#xff08;YaRN\/\u63d2\u503c&#xff09;&#xff1b;v3.5 \u7684\u957f\u5ea6\u65e0\u5173\u6027\u4e3a\u67b6\u6784\u81ea\u5e26<\/li>\n<li>\u901f\u5ea6\u4e0a KV \u901a\u9053\u4e3a max-scan \u7684 ~15\u00d7 \u5355 token \u6210\u672c&#xff0c;Triton \u524d\u5411\u5df2\u56de\u6536\u5927\u5934&#xff1b;\u8fdb\u4e00\u6b65\u53ef\u505a KV \u7626\u8eab&#xff08;dh \u51cf\u534a&#xff09;\u6d88\u878d<\/li>\n<h3>9. \u540e\u7eed\u4efb\u52a1\u6267\u884c\u7ed3\u679c&#xff08;2026-11-05 \u66f4\u65b0&#xff09;<\/h3>\n<table>\n<tr>\u4efb\u52a1\u7ed3\u679c<\/tr>\n<tbody>\n<tr>\n<td>tf \u4f4d\u7f6e\u63d2\u503c\u516c\u5e73\u6027&#xff08;PI \u00d74&#xff09;<\/td>\n<td>tf&#064;2048 \u6062\u590d 37.68\u219235.20&#xff0c;v3.5 \u4ecd\u9886\u5148 22.8%&#xff08;15.64&#xff09;&#xff1b;\u4f18\u52bf\u5bf9\u516c\u5e73\u6027\u8865\u5f3a\u7a33\u5065<\/td>\n<\/tr>\n<tr>\n<td>\u589e\u91cf O(1) \u63a8\u7406\u5151\u73b0<\/td>\n<td>state \u5143\u7ec4\u643a\u5e26 (e, S)&#xff1b;fp32 \u9010 token vs \u5168\u5e8f\u5217 diff 3.7e-06&#xff1b;demo_o1_decode.py \u7aef\u5230\u7aef\u6f14\u793a&#xff08;\u5e38\u6570\u72b6\u6001 894k \u6570\u503c\/12 \u5c42&#xff0c;\u751f\u6210\u8d28\u91cf\u8fde\u8d2f&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>KV \u7626\u8eab\u6d88\u878d&#xff08;kv_dh 96\u219248&#xff09;<\/td>\n<td>v35s PPL 40.87 vs \u5b8c\u6574\u7248 41.34&#xff08;3000 \u6b65&#xff09;\u2014\u2014\u53c2\u6570 -3.5M\u3001KV \u72b6\u6001\/\u8ba1\u7b97\u51cf\u534a\u4e14\u8d28\u91cf\u7565\u4f18&#xff1b;48 \u7ef4\u952e\u7a7a\u95f4\u68c0\u7d22\u66f4\u9510\u5229\u3002v35s \u6210\u4e3a\u65b0\u63a8\u8350\u914d\u7f6e<\/td>\n<\/tr>\n<tr>\n<td>delta-rule \u5199\u5165<\/td>\n<td>naive \u7248\u6027\u8d28\u9a8c\u8bc1 PASS&#xff08;\u91cd\u590d\u952e\u7cbe\u786e\u8986\u76d6&#xff09;&#xff1b;\u5206\u5757 GPU \u7248\u5b58\u5728\u53cc\u5411\u72b6\u6001\u4f9d\u8d56&#xff08;\u53cd\u5411 gq_j &#061; \u03a3_v \u03bb_j\u00b7S_j \u9700 \u03bb \u53cd\u5411\u9012\u63a8\u4e0e S \u6b63\u5411\u91cd\u7b97\u540c\u573a&#xff09;\u2014\u2014\u7cbe\u786e\u5f62\u5f0f\u9700 Gated DeltaNet&#xff08;arXiv 2412.06464&#xff09;\u8bba\u6587\u7ea7 UT\/WY \u5b9e\u73b0&#xff1b;\u5df2\u4ea4\u4ed8 naive \u53c2\u8003\u5b9e\u73b0&#xff08;\u6027\u8d28\u9a8c\u8bc1 PASS&#xff09;\u4e0e TBPTT \u8bda\u5b9e\u8bc4\u4f30&#xff08;384 \u6b65 Python \u5faa\u73af 18s\/step \u4e0d\u53ef\u8bad\u7ec3&#xff09;\u3002\u5927\u6570\u636e\u7248 gap&#061;128 \u68c0\u7d22\u5df2\u53cd\u8d85 tf&#xff0c;\u7d27\u8feb\u6027\u4e0b\u964d<\/td>\n<\/tr>\n<tr>\n<td>KV scan \u624b\u5de5 Triton \u53cd\u5411<\/td>\n<td>\u964d\u7ea7&#xff1a;\u91cd\u7b97\u53cd\u5411&#xff08;4.3 ms\/\u5c42&#xff09;\u5df2\u591f\u5feb&#xff0c;\u624b\u5de5\u63a8\u5bfc\u6536\u76ca ~2 ms\/\u5c42\u3001\u98ce\u9669\u9ad8<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>v35s \u589e\u91cf\u53d1\u73b0<\/h4>\n<p>KV \u901a\u9053 48 \u7ef4\u952e\u7a7a\u95f4\u76f8\u5bf9 96 \u7ef4\u662f\u8d28\u91cf-\u6548\u7387\u53cc\u4f18&#xff08;\u66f4\u5c11\u53c2\u6570\u3001\u66f4\u9510\u68c0\u7d22\u3001\u66f4\u5c0f\u72b6\u6001 36,864\u219218,432 \u6570\u503c\/\u5c42&#xff09;\u3002\u67b6\u6784\u53d9\u8ff0\u66f4\u65b0\u4e3a&#xff1a;\u5feb\/\u6162 max-scan &#043; 48 \u7ef4\u5f52\u4e00\u5316 KV \u68c0\u7d22\u3002<\/p>\n<p>https:\/\/github.com\/dfytensor\/OpenAshAnyWhere\/tree\/master<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\/* MD \/ \u5bcc\u6587\u672c\u4e2d\u7684 .toc\uff08\u542b\u535a\u5ba2\u56ed\u642c\u5bb6\u7b49\u5d4c\u5957\u7ed3\u6784\uff09\uff1b.toc-box \u5728\u4fa7\u680f\uff0c\u4e0d\u53d7\u5f71\u54cd *\/#content_views .toc,\/* \u7f16\u8f91\u5668\u5e38\u5728\u76ee\u5f55\u524d\u540e\u63d2\u5165\u7a7a p\uff08:empty \u4ecd\u5360 20px\uff09\uff0c\u4e00\u5e76\u53bb\u6389\u907f\u514d\u9876\u7a7a\u9699 *\/#content_views.markdown_views &gt; p:empty:has(+ .toc),#content_views.markdown_views &gt; .toc + p:empty,\/* \u5bcc\u6587\u672c\u65e7\u7248\u76ee\u5f55\u6807\u8bb0 *\/#content<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[188],"topic":[],"class_list":["post-122715","post","type-post","status-publish","format-standard","hentry","category-server","tag-188"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>OpenASH v3.5 \u67b6\u6784\u8fed\u4ee3\u62a5\u544a\uff1a\u8870\u51cf\u6700\u5927\u503c\u4e09\u901a\u9053\u6df7\u5408\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/122715.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"OpenASH v3.5 \u67b6\u6784\u8fed\u4ee3\u62a5\u544a\uff1a\u8870\u51cf\u6700\u5927\u503c\u4e09\u901a\u9053\u6df7\u5408\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\/* MD \/ \u5bcc\u6587\u672c\u4e2d\u7684 .toc\uff08\u542b\u535a\u5ba2\u56ed\u642c\u5bb6\u7b49\u5d4c\u5957\u7ed3\u6784\uff09\uff1b.toc-box \u5728\u4fa7\u680f\uff0c\u4e0d\u53d7\u5f71\u54cd *\/#content_views .toc,\/* \u7f16\u8f91\u5668\u5e38\u5728\u76ee\u5f55\u524d\u540e\u63d2\u5165\u7a7a p\uff08:empty \u4ecd\u5360 20px\uff09\uff0c\u4e00\u5e76\u53bb\u6389\u907f\u514d\u9876\u7a7a\u9699 *\/#content_views.markdown_views &gt; p:empty:has(+ .toc),#content_views.markdown_views &gt; .toc + p:empty,\/* \u5bcc\u6587\u672c\u65e7\u7248\u76ee\u5f55\u6807\u8bb0 *\/#content\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/122715.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-11-05T06:24:58+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"4 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/122715.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/122715.html\",\"name\":\"OpenASH v3.5 \u67b6\u6784\u8fed\u4ee3\u62a5\u544a\uff1a\u8870\u51cf\u6700\u5927\u503c\u4e09\u901a\u9053\u6df7\u5408\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-11-05T06:24:58+00:00\",\"dateModified\":\"2026-11-05T06:24:58+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/122715.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/122715.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/122715.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"OpenASH v3.5 \u67b6\u6784\u8fed\u4ee3\u62a5\u544a\uff1a\u8870\u51cf\u6700\u5927\u503c\u4e09\u901a\u9053\u6df7\u5408\u67b6\u6784\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"OpenASH v3.5 \u67b6\u6784\u8fed\u4ee3\u62a5\u544a\uff1a\u8870\u51cf\u6700\u5927\u503c\u4e09\u901a\u9053\u6df7\u5408\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/122715.html","og_locale":"zh_CN","og_type":"article","og_title":"OpenASH v3.5 \u67b6\u6784\u8fed\u4ee3\u62a5\u544a\uff1a\u8870\u51cf\u6700\u5927\u503c\u4e09\u901a\u9053\u6df7\u5408\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\/* MD \/ \u5bcc\u6587\u672c\u4e2d\u7684 .toc\uff08\u542b\u535a\u5ba2\u56ed\u642c\u5bb6\u7b49\u5d4c\u5957\u7ed3\u6784\uff09\uff1b.toc-box \u5728\u4fa7\u680f\uff0c\u4e0d\u53d7\u5f71\u54cd *\/#content_views .toc,\/* \u7f16\u8f91\u5668\u5e38\u5728\u76ee\u5f55\u524d\u540e\u63d2\u5165\u7a7a p\uff08:empty \u4ecd\u5360 20px\uff09\uff0c\u4e00\u5e76\u53bb\u6389\u907f\u514d\u9876\u7a7a\u9699 *\/#content_views.markdown_views &gt; p:empty:has(+ .toc),#content_views.markdown_views &gt; .toc + p:empty,\/* \u5bcc\u6587\u672c\u65e7\u7248\u76ee\u5f55\u6807\u8bb0 *\/#content","og_url":"https:\/\/www.wsisp.com\/helps\/122715.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-11-05T06:24:58+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"4 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/122715.html","url":"https:\/\/www.wsisp.com\/helps\/122715.html","name":"OpenASH v3.5 \u67b6\u6784\u8fed\u4ee3\u62a5\u544a\uff1a\u8870\u51cf\u6700\u5927\u503c\u4e09\u901a\u9053\u6df7\u5408\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-11-05T06:24:58+00:00","dateModified":"2026-11-05T06:24:58+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/122715.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/122715.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/122715.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"OpenASH v3.5 \u67b6\u6784\u8fed\u4ee3\u62a5\u544a\uff1a\u8870\u51cf\u6700\u5927\u503c\u4e09\u901a\u9053\u6df7\u5408\u67b6\u6784"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/122715","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=122715"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/122715\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=122715"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=122715"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=122715"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=122715"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}