{"id":94169,"date":"2026-08-13T21:25:09","date_gmt":"2026-08-13T13:25:09","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/94169.html"},"modified":"2026-08-13T21:25:09","modified_gmt":"2026-08-13T13:25:09","slug":"%e8%b6%85%e7%ae%97%e4%b8%ad%e5%bf%83%e6%b5%b7%e5%85%89%e5%bc%82%e6%9e%84%e5%8d%a1dcu-bw-64g%e6%98%be%e5%8d%a1%e6%8a%a5%e9%94%99-nccl%e9%80%9a%e8%ae%af%e9%94%99%e8%af%af%ef%bc%8c%e6%97%a0%e6%b3%95","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/94169.html","title":{"rendered":"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 nccl\u901a\u8baf\u9519\u8bef\uff0c\u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3\u7684\u529e\u6cd5\uff0c\u9700\u8981\u5b89\u88c5shca IB\u5361\u9a71\u52a8"},"content":{"rendered":"<h3>\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 \u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3&#xff0c;\u901a\u8fc7\u65b0\u589e\u670d\u52a1\u5668\u8df3\u8fc7\u663e\u5361<\/h3>\n<p>flash_attn: disabled  HOSTQUEUE &lt;0x1921b670&gt;: &gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt; DUMP KERNEL ARGS: size: 160 &lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;<br \/>\n00 40 83 00 00 00 00 00 00 00 90 01 00 00 00 00<br \/>\n00 40 83 00 00 00 00 00 00 00 b4 4e e1 7f 00 00<br \/>\n00 00 b4 4e e1 7f 00 00 00 00 60 d0 d6 7f 00 00<br \/>\n00 80 6f 31 e2 7f 00 00 00 00 80 3f 00 00 00 00<br \/>\n00 14 00 00 01 00 00 00 00 14 00 00 01 00 00 00<br \/>\n00 14 00 00 01 00 00 00 00 14 00 00 01 00 00 00<br \/>\n00 14 00 00 90 06 00 00 01 00 00 00 00 14 00 00<br \/>\n00 00 00 00 14 00 00 00 07 00 00 00 00 00 00 00<br \/>\n07 00 00 00 93 24 49 12 00 00 00 00 00 00 00 00<br \/>\n00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: &gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt; DUMP KERNEL ARGS PTR INFO &lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: ptr arg index: 3, ptr: 0x7fe14eb40000<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: origin ptr: 0x7fe14d200000, size byte: 113246208<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: ptr arg index: 4, ptr: 0x7fe14eb40000<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: origin ptr: 0x7fe14d200000, size byte: 113246208<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: ptr arg index: 5, ptr: 0x7fd6d0600000<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: origin ptr: 0x7fd6d0600000, size byte: 52428800<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: ptr arg index: 6, ptr: 0x7fe2316f8000<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: origin ptr: 0x7fe22ec00000, size byte: 88080384<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: &gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt; DUMP AQL 0x7fe93e825000 PACKET WITH INDEX 217545 &lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: &gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt; MATCH KERNEL COMMAND &lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: header: 0xb02<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: setup: 3<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: workgroup: x:256, y:1, z:1<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: grid: x:2150400, y:1, z:1<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: group_segment_size: 6144<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: private_segment_size: 0<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: completion_signal: 0x0<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: kernel_object: 0x7fe4a82424c0<br \/>\nHOSTQUEUE &lt;0x1921b670&gt;: device id: 0<br \/>\nInvalid address access: 0x7fde7654c000, Error code: 3.<\/p>\n<p>KERNEL VMFault !!! &lt;&lt;&lt;&lt;&lt;&lt;<br \/>\nPID: 6011, SIGNAL: 0 !!! &lt;&lt;&lt;&lt;&lt;&lt;<br \/>\n\u8fd8\u662f\u4e00\u6837<\/p>\n<h3>\u53cc\u5361\u63d0\u793ancll \u9519\u8bef<\/h3>\n<h4>\u539f\u56e0&#xff0c;\u95ee\u4e86\u5ba2\u670d<\/h4>\n<h6>\u95ee\u904d \u51e0\u4e4e\u6240\u6709ai ,\u90fd\u6ca1\u6709\u627e\u5230\u6b63\u786e\u7b54\u6848:<\/h6>\n<p>claude gmini gpt doubao  kimi<\/p>\n<h6>\u5ba2\u670d\u56de\u7b54<\/h6>\n<p>\u8001\u5e08\u60a8\u597d&#xff0c;\u770b\u5230\u60a8\u4f7f\u7528\u7684\u662f113\u7ec4&#xff0c;\u8be5\u7ec4\u8d44\u6e90\u4f7f\u7528\u7684\u662fshca IB\u5361 &#xff0c;\u9700\u8981\u5b89\u88c5\u4e0b\u63d2\u4ef6&#xff0c;\u76ee\u524d\u5728\u60a8\/public\/home\/aaaa\/IB \u8def\u5f84\u653e\u4e86\u811a\u672c&#xff0c;\u60a8\u540e\u7eed\u5728notebook\u4e2d\u4f7f\u7528root\u6743\u9650\u8fd0\u884c\u4e0bbash mlxtoshca.sh &#xff0c;\u7136\u540e\u8ba1\u7b97\u7684\u65f6\u5019\u518d\u6dfb\u52a0\u5982\u4e0b\u53c2\u6570\u6d4b\u8bd5\u4e0b<br \/>\nexport GLOO_SOCKET_IFNAME&#061;\u201ceth0\u201d<br \/>\nexport NCCL_SOCKET_IFNAME&#061;\u201ceth0\u201d<br \/>\nexport NCCL_IB_HCA&#061;shca_0:1,shca_1:1,shca_2:1,shca_3:1<br \/>\nexport HSA_FORCE_FINE_GRAIN_PCIE&#061;1<br \/>\nexport NCCL_IB_DISABLE&#061;0<br \/>\nexport NCCL_NET_PLUGIN&#061;shca<br \/>\nexport  NCCL_DEBUG&#061;INFO<\/p>\n<h4>\u89e3\u51b3\u65b9\u6cd51,\u7b28\u529e\u6cd5&#xff08;\u4e34\u65f6\u529e\u6cd5\u4e0d\u5efa\u8bae&#xff09;<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813132508-6a7dc5b41cd0d.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<h4>\u65b9\u6cd5\u4e8c &#xff0c;\u5b89\u88c5shca IB\u5361\u9a71\u52a8<\/h4>\n<p>\u8001\u5e08\u60a8\u597d&#xff0c;\u770b\u5230\u60a8\u4f7f\u7528\u7684\u662f113\u7ec4&#xff0c;\u8be5\u7ec4\u8d44\u6e90\u4f7f\u7528\u7684\u662fshca IB\u5361 &#xff0c;\u9700\u8981\u5b89\u88c5\u4e0b\u63d2\u4ef6&#xff0c;\u76ee\u524d\u5728\u60a8\/public\/home\/aaaa\/IB \u8def\u5f84\u653e\u4e86\u811a\u672c&#xff0c;\u60a8\u540e\u7eed\u5728notebook\u4e2d\u4f7f\u7528root\u6743\u9650\u8fd0\u884c\u4e0bbash mlxtoshca.sh &#xff0c;\u7136\u540e\u8ba1\u7b97\u7684\u65f6\u5019\u518d\u6dfb\u52a0\u5982\u4e0b\u53c2\u6570\u6d4b\u8bd5<\/p>\n<p>\u4e0b<\/p>\n<p><span class=\"token function\">apt-get<\/span> update <span class=\"token parameter variable\">-y<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span>  rdmacm-utils <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> ibacm <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span>  perftest <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> ibverbs-utils <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> ucx <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> libibverbs-dev <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> libibmad-dev <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> libibumad-dev <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> librdmacm1 <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> infiniband-diags <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> opensm <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> rdma-core <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> libibmad5 <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> libibumad3 <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> ibverbs-providers  <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> remove <span class=\"token parameter variable\">-y<\/span> libibverbs1 <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt<\/span> <span class=\"token function\">install<\/span> <span class=\"token parameter variable\">-y<\/span> libmosquitto1 <span class=\"token operator\">||<\/span> <span class=\"token boolean\">true<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n        <span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">apt-get<\/span> clean <span class=\"token punctuation\">\\\\<\/span><br \/>\n    <span class=\"token operator\">&amp;&amp;<\/span> <span class=\"token function\">rm<\/span> <span class=\"token parameter variable\">-rf<\/span> \/var\/lib\/apt\/lists\/* <span class=\"token punctuation\">\\\\<\/span><br \/>\n    <span class=\"token operator\">&amp;&amp;<\/span> dpkg <span class=\"token parameter variable\">-i<\/span> shca-tools_2.500.4.B059-Ubuntu22.04_amd64.deb <\/p>\n<h5>\u8ba1\u7b97\u7684\u65f6\u5019\u518d\u6dfb\u52a0\u5982\u4e0b\u53c2\u6570\u6d4b\u8bd5<\/h5>\n<p><span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">GLOO_SOCKET_IFNAME<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;eth0&#034;<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_SOCKET_IFNAME<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;eth0&#034;<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_IB_HCA<\/span><span class=\"token operator\">&#061;<\/span>shca_0:1,shca_1:1,shca_2:1,shca_3:1<br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">HSA_FORCE_FINE_GRAIN_PCIE<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_IB_DISABLE<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">0<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_NET_PLUGIN<\/span><span class=\"token operator\">&#061;<\/span>shca<br \/>\n<span class=\"token builtin class-name\">export<\/span>  <span class=\"token assign-left variable\">NCCL_DEBUG<\/span><span class=\"token operator\">&#061;<\/span>INFO<\/p>\n<h4>shca ib\u5361\u80cc\u666f\u77e5\u8bc6<\/h4>\n<p>\u6211\u7528\u5927\u767d\u8bdd\u7ed9\u4f60\u62c6\u5f00\u8bb2&#xff0c;\u4e0d\u641e\u6280\u672f\u9ed1\u8bdd\u3002<\/p>\n<hr \/>\n<h3>\u4e00\u3001\u4ec0\u4e48\u662f\u300c\u6d77\u5149\u5f02\u6784\u5361BW 64G\u300d&#xff1f;<\/h3>\n<ul>\n<li>\u6d77\u5149&#xff1a;\u56fd\u4ea7\u82af\u7247\u516c\u53f8&#xff0c;\u505a\u56fd\u4ea7GPU\/AI\u52a0\u901f\u5361&#xff0c;\u5bf9\u6807\u82f1\u4f1f\u8fbeA100\/H100\u3002<\/li>\n<li>\u5f02\u6784\u5361&#xff08;DCU\/BW&#xff09;&#xff1a;\u5c31\u662f\u56fd\u4ea7GPU&#xff0c;\u4e13\u95e8\u7528\u6765\u8dd1AI\u8bad\u7ec3\u3001\u5927\u6a21\u578b\u3001\u79d1\u5b66\u8ba1\u7b97\u3002<\/li>\n<li>BW1000&#xff1a;\u6d77\u5149\u6df1\u7b97\u4e09\u53f7&#xff0c;\u578b\u53f7\u53ebBW&#xff0c;64G\u662f\u6307\u5b83\u81ea\u5e2664GB \u9ad8\u901f\u663e\u5b58&#xff08;HBM2e&#xff09;\u3002<\/li>\n<li>\u4e00\u53e5\u8bdd&#xff1a;\u4f60\u7528\u7684\u662f\u56fd\u4ea7\u9ad8\u7aefAI\u663e\u5361&#xff0c;64G\u663e\u5b58&#xff0c;\u6027\u80fd\u63a5\u8fd1\u82f1\u4f1f\u8fbeA100\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e8c\u3001\u4ec0\u4e48\u662f\u300cshca IB\u5361\u300d&#xff1f;<\/h3>\n<h4>1. IB \u662f\u4ec0\u4e48&#xff1f;<\/h4>\n<ul>\n<li>IB &#061; InfiniBand&#xff1a;\u4e00\u79cd\u8d85\u9ad8\u901f\u3001\u4f4e\u5ef6\u8fdf\u7684\u4e13\u7528\u7f51\u7edc&#xff0c;\u4e13\u95e8\u7528\u5728\u8d85\u7b97\u3001AI\u96c6\u7fa4\u91cc&#xff0c;\u8ba9\u5f88\u591a\u5f20GPU\u5361\/\u670d\u52a1\u5668\u4e4b\u95f4\u98de\u5feb\u5730\u4e92\u76f8\u4f20\u6570\u636e\u3002<\/li>\n<li>\u666e\u901a\u7f51\u5361&#xff08;\u4ee5\u592a\u7f51&#xff09;\u50cf\u666e\u901a\u516c\u8def&#xff1b;IB\u5361\u50cf\u9ad8\u94c1\u4e13\u7ebf&#xff0c;\u901f\u5ea6\u5feb\u51e0\u5341\u500d\u3001\u5ef6\u8fdf\u6781\u4f4e\u3002<\/li>\n<\/ul>\n<h4>2. shca \u662f\u4ec0\u4e48&#xff1f;<\/h4>\n<ul>\n<li>shca&#xff1a;\u662f\u66d9\u5149\u81ea\u7814\u7684\u56fd\u4ea7IB\u9a71\u52a8\/\u63d2\u4ef6\u7684\u540d\u5b57&#xff08;\u53ef\u4ee5\u7406\u89e3\u4e3a&#xff1a;\u56fd\u4ea7\u7248IB\u9a71\u52a8&#xff09;\u3002<\/li>\n<li>\u56e0\u4e3a\u7f8e\u56fd\u9650\u5236&#xff0c;\u4e0d\u80fd\u7528\u82f1\u4f1f\u8fbe&#xff08;Mellanox&#xff09;\u7684IB\u5361&#xff0c;\u66d9\u5149\u5c31\u81ea\u5df1\u505a\u4e86\u56fd\u4ea7IB\u7f51\u5361&#043;\u9a71\u52a8&#xff0c;\u540d\u5b57\u53ebshca\u3002<\/li>\n<\/ul>\n<h4>3. \u5408\u8d77\u6765&#xff1a;shca IB\u5361<\/h4>\n<ul>\n<li>\u5c31\u662f&#xff1a;\u66d9\u5149\u56fd\u4ea7\u9ad8\u901f\u4e92\u8054\u7f51\u5361&#xff08;\u66ff\u4ee3\u82f1\u4f1f\u8fbeIB\u5361&#xff09;&#xff0c;\u4e13\u95e8\u7ed9\u6d77\u5149BW\u5361\u96c6\u7fa4\u505a\u9ad8\u901f\u901a\u4fe1\u7528\u3002<\/li>\n<li>\u4f5c\u7528&#xff1a;\u8ba9\u591a\u5f20\u6d77\u5149BW\u5361\u4e4b\u95f4\u9ad8\u901f\u4e92\u8054&#xff0c;\u8dd1\u5927\u6a21\u578b\u591a\u5361\u8bad\u7ec3\u65f6\u66f4\u5feb\u3001\u4e0d\u5361\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e09\u3001\u4ed6\u8ba9\u4f60\u505a\u7684\u4e8b&#xff0c;\u672c\u8d28\u662f\u5565&#xff1f;<\/h3>\n<p>\u8001\u5e08\u7684\u610f\u601d\u662f&#xff1a;<\/p>\n<p>\u4f60\u7528\u7684113\u7ec4\u673a\u5668&#xff0c;\u914d\u7684\u662f\u66d9\u5149\u56fd\u4ea7shca IB\u9ad8\u901f\u7f51\u5361&#xff1b;<br \/>\n\u8981\u7528\u597d\u8fd9\u5f20\u9ad8\u901f\u7f51\u5361\u3001\u53d1\u6325\u591a\u5361\u6027\u80fd&#xff0c;\u5fc5\u987b&#xff1a;<\/p>\n<li>\u5148\u88c5shca\u9a71\u52a8\u63d2\u4ef6&#xff08;\u4ed6\u7ed9\u7684\u811a\u672c&#xff09;&#xff1b;<\/li>\n<li>\u518d\u8bbe\u51e0\u4e2a\u73af\u5883\u53d8\u91cf&#xff0c;\u544a\u8bc9\u7cfb\u7edf&#xff1a;\u201c\u7528shca IB\u5361\u901a\u4fe1&#xff0c;\u522b\u7528\u666e\u901a\u7f51\u5361\u201d\u3002<\/li>\n<hr \/>\n<h3>\u56db\u3001\u6781\u7b80\u603b\u7ed3&#xff08;\u8bb0\u8fd9\u51e0\u53e5\u5c31\u884c&#xff09;<\/h3>\n<ul>\n<li>\u6d77\u5149BW 64G&#xff1a;\u56fd\u4ea7AI\u663e\u5361&#xff0c;64G\u663e\u5b58&#xff0c;\u7528\u6765\u8dd1\u5927\u6a21\u578b\u3002<\/li>\n<li>IB\u5361&#xff1a;\u9ad8\u901f\u4e13\u7528\u7f51\u5361&#xff08;\u50cf\u9ad8\u94c1&#xff09;&#xff0c;\u591a\u5361\/\u591a\u673a\u901a\u4fe1\u7528\u3002<\/li>\n<li>shca&#xff1a;\u66d9\u5149\u56fd\u4ea7IB\u9a71\u52a8&#xff08;\u66ff\u4ee3\u7f8e\u56fd\u8d27&#xff09;\u3002<\/li>\n<li>\u88c5\u811a\u672c&#043;\u8bbe\u53d8\u91cf&#xff1a;\u5c31\u662f\u542f\u7528\u56fd\u4ea7\u9ad8\u901f\u7f51\u5361&#xff0c;\u8ba9\u591a\u5361\u8bad\u7ec3\u66f4\u5feb\u3002<\/li>\n<\/ul>\n<h4>\u539f\u6765\u7684\u6b63\u786e\u7684\u73af\u5883\u53d8\u91cf<\/h4>\n<p><span class=\"token comment\"># &#061;&#061;&#061;&#061;&#061;&#061;&#061;\u66d9\u5149BW 2\u5361DCU 100\u5c0f\u65f6\u7a33\u5b9a\u8bad\u7ec3\u3010\u65e0\u62a5\u9519\u7248\u3011 &#061;&#061;&#061;&#061;&#061;&#061;&#061;<\/span><br \/>\n<span class=\"token comment\"># &#061;&#061;&#061;&#061;&#061;  \u8d85\u7b97\u4e2d\u5fc3  \u8001\u662f\u62a5\u9519 \u53ef\u4ee5 &#xff0c;\u63622\u4e2a\u663e\u5361 &#061;&#061;&#061;&#061;<\/span><br \/>\n<span class=\"token comment\"># NCCL \u5185\u90e8\u901a\u4fe1\u9519\u8bef<\/span><br \/>\n<span class=\"token comment\"># export NCCL_IGNORE_DISABLED_P2P&#061;1<\/span><br \/>\n<span class=\"token comment\"># export NCCL_P2P_DISABLE&#061;1<\/span><\/p>\n<p><span class=\"token comment\"># 1. \u7981\u7528\u53ef\u80fd\u5b58\u5728 Bug \u7684\u52a0\u901f\u7b97\u5b50\u5e93<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">ROC_BLAS_USE_HIPBLASLT<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">0<\/span><br \/>\n<span class=\"token comment\"># 2. \u5f3a\u5236\u540c\u6b65\u6267\u884c&#xff0c;\u9632\u6b62\u5f02\u6b65\u5185\u5b58\u8bbf\u95ee\u8d8a\u754c&#xff08;\u867d\u7136\u6162&#xff0c;\u4f46\u80fd\u4fdd\u547d&#xff09;<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">HIP_LAUNCH_BLOCKING<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><br \/>\n<span class=\"token comment\"># 3. \u9650\u5236 MIOpen \u7b97\u5b50&#xff0c;\u9632\u6b62\u5728\u5377\u79ef\u6216\u77e9\u9635\u8fd0\u7b97\u65f6\u51fa\u73b0\u6bb5\u9519\u8bef<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">MIOPEN_DEBUG_DISABLE_CUDNN_COMPAT<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><\/p>\n<p><span class=\"token comment\"># \u4e0d\u5fc5\u5220\u9664 checkpoint-2500\u7684global_step&#xff0c;\u5426\u5219\u6211\u542f\u52a8\u4e0d\u4e86<\/span><\/p>\n<p><span class=\"token comment\"># gfx936 \u4e0d\u4e00\u5b9a\u9700\u8981<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">HSA_OVERRIDE_GFX_VERSION<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;9.3.6&#034;<\/span><\/p>\n<p><span class=\"token comment\"># unset HSA_OVERRIDE_GFX_VERSION<\/span><\/p>\n<p><span class=\"token comment\"># \u8bbe\u7f6eNCCL\u73af\u5883\u53d8\u91cf<\/span><br \/>\n<span class=\"token comment\"># 4. \u5f7b\u5e95\u7981\u7528 P2P \u901a\u4fe1&#xff0c;\u9632\u6b62\u5355\u5361\u6a21\u5f0f\u4e0b\u8bef\u89e6\u901a\u4fe1 Bug<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_P2P_DISABLE<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span>          <span class=\"token comment\"># \u6d77\u5149DCU\u5fc5\u987b\u5f00&#xff01;\u4e0d\u5f00\u76f4\u63a5\u5361<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_BLOCKING_WAIT<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span>         <span class=\"token comment\"># \u9632\u6b62\u5361\u6b7b<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_ASYNC_ERROR_HANDLING<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span>  <span class=\"token comment\"># \u9632\u6b62\u901a\u4fe1\u5d29\u6e83<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">RCCL_P2P_DISABLE<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><\/p>\n<p><span class=\"token comment\"># \u8d85\u7b97BW\u7981\u7528IB\u8fdc\u7a0b\u901a\u4fe1&#xff0c;\u907f\u514d\u5206\u5e03\u5f0f\u8bad\u7ec3\u7f51\u7edc\u62a5\u9519<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_IB_DISABLE<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><\/p>\n<p><span class=\"token comment\"># \u6307\u5b9aNCCL\u4f7f\u7528\u672c\u5730\u56de\u73af\u7f51\u5361lo&#xff0c;\u53cc\u5361\u672c\u673a\u901a\u4fe1\u6700\u7a33\u5b9a<\/span><br \/>\n<span class=\"token comment\"># \u672c\u5730\u53cc\u5361 \u4e0d\u80fd\u7528\u7f51\u5361<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_SOCKET_IFNAME<\/span><span class=\"token operator\">&#061;<\/span>lo<\/p>\n<p><span class=\"token comment\"># \u5f00\u542fNCCL\u5f02\u6b65\u9519\u8bef\u5904\u7406&#xff0c;\u9632\u6b62\u901a\u4fe1\u5361\u6b7b\u5bfc\u81f4\u8282\u70b9\u5931\u8d25<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_ASYNC_ERROR_HANDLING<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><\/p>\n<p><span class=\"token comment\"># NCCL\u8d85\u957f\u8d85\u65f6(100\u5c0f\u65f6)&#xff0c;\u907f\u514d\u6162\u8282\u70b9\u4e2d\u65ad<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_TIMEOUT<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">3600000<\/span><\/p>\n<p><span class=\"token comment\"># \u4fee\u590d ROCm \u663e\u5b58\u788e\u7247<\/span><br \/>\n<span class=\"token comment\"># \u6211\u76f4\u63a5\u7ed9\u4f60\u6df1\u7b97 3 \u53f7 Z100 (gfx928) \u4e13\u7528\u3001100% \u80fd\u8dd1\u7684\u6700\u7ec8\u7248<\/span><br \/>\n<span class=\"token comment\"># \u6d77\u5149DCU\u4e0d\u652f\u6301expandable_segments&#xff0c;\u53ea\u4fdd\u7559max_split_size_mb\u9632\u6b62\u663e\u5b58\u788e\u7247<\/span><br \/>\n<span class=\"token comment\"># \u4f46\u9700\u8981\u6ce8\u610f&#xff0c;\u6d77\u5149\u73af\u5883\u6709\u65f6\u5bf9 expandable_segments \u652f\u6301\u6709 Bug\u3002\u5982\u679c\u663e\u5b58\u7aef\u4e5f\u51fa\u73b0\u65e0\u6545 OOM&#xff0c;\u53ef\u4ee5\u5c1d\u8bd5\u5c06\u5176\u7b80\u5316\u4e3a export PYTORCH_HIP_ALLOC_CONF&#061;&#034;garbage_collection_threshold:0.8&#034;<\/span><br \/>\n<span class=\"token comment\"># export PYTORCH_HIP_ALLOC_CONF&#061;&#034;expandable_segments:True,garbage_collection_threshold:0.8&#034;<\/span><br \/>\n<span class=\"token comment\"># export PYTORCH_HIP_ALLOC_CONF&#061;&#034;garbage_collection_threshold:0.8&#034;<\/span><\/p>\n<p><span class=\"token comment\"># \u5efa\u8bae\u4fee\u6539\u8fd9\u4e00\u884c<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">PYTORCH_HIP_ALLOC_CONF<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;max_split_size_mb:128,garbage_collection_threshold:0.8&#034;<\/span><\/p>\n<p><span class=\"token comment\"># \u5173\u95edMIOPEN\u786e\u5b9a\u6027\u8ba1\u7b97&#xff0c;\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3001\u63d0\u901f<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">0<\/span><\/p>\n<p><span class=\"token comment\"># \u5173\u95edMIOPEN\u5197\u4f59\u65e5\u5fd7&#xff0c;\u907f\u514d\u7f13\u51b2\u533a\u5360\u6ee1\u5185\u5b58<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">MIOPEN_LOG_LEVEL<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><\/p>\n<p><span class=\"token comment\"># \u6307\u5b9a\u4f7f\u75282\u5f20DCU\u663e\u5361<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">HIP_VISIBLE_DEVICES<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">0,1<\/span><\/p>\n<p><span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">ROCM_VISIBLE_DEVICES<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">0,1<\/span><\/p>\n<p><span class=\"token builtin class-name\">unset<\/span> CUDA_VISIBLE_DEVICES<\/p>\n<p><span class=\"token comment\"># DCU\u517c\u5bb9\u6027\u5fc5\u987b\u8bbe\u7f6e&#xff0c;\u89e3\u51b3\u7b97\u5b50\u5d29\u6e83\/\u663e\u5b58\u6cc4\u6f0f<\/span><br \/>\n<span class=\"token comment\"># export HSA_OVERRIDE_GFX_VERSION&#061;10.3.0 pytorch\u4e0d\u652f\u6301<\/span><\/p>\n<p><span class=\"token comment\"># \u5173\u95edCuDNN&#xff0c;\u9632\u6b62DCU\u73af\u5883\u4e0d\u517c\u5bb9<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">Torch_USE_CUDNN<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">0<\/span><\/p>\n<h4>\u67e5\u770bib\u7f51\u7edc<\/h4>\n<p>\u4e0b\u9762\u7ed9\u4f60\u4e00\u5957\u76f4\u63a5\u590d\u5236\u5c31\u80fd\u7528\u3001\u6700\u5e38\u7528\u3001\u6700\u51c6\u7684\u6392\u67e5\u65b9\u6cd5&#xff08;\u5728\u4f60\u8fd9\u53f0\u6d77\u5149BW\u670d\u52a1\u5668\u4e0a&#xff0c;root \u6216 sudo \u6267\u884c&#xff09;\u3002<\/p>\n<hr \/>\n<h3>\u4e00\u3001\u5148\u770b&#xff1a;\u670d\u52a1\u5668\u6709\u6ca1\u6709 IB \u5361&#xff08;shca\/InfiniBand&#xff09;<\/h3>\n<h4>1. \u770b\u786c\u4ef6\u662f\u5426\u8bc6\u522b&#xff08;\u6700\u57fa\u7840&#xff09;<\/h4>\n<p>lspci <span class=\"token operator\">|<\/span> <span class=\"token function\">grep<\/span> <span class=\"token parameter variable\">-i<\/span> infiniband<br \/>\nlspci <span class=\"token operator\">|<\/span> <span class=\"token function\">grep<\/span> <span class=\"token parameter variable\">-i<\/span> shca<\/p>\n<ul>\n<li>\u6709\u8f93\u51fa &#061; \u8bc6\u522b\u5230 IB \u5361&#xff08;shca \u4e5f\u662f IB \u63a7\u5236\u5668&#xff09;<\/li>\n<li>\u65e0\u8f93\u51fa &#061; \u6ca1\u8bc6\u522b\u5230 IB \u5361<\/li>\n<\/ul>\n<h4>2. \u770b IB \u8bbe\u5907\u5217\u8868&#xff08;\u66f4\u4e13\u4e1a&#xff09;<\/h4>\n<p>ibv_devices<\/p>\n<ul>\n<li>\u8f93\u51fa\u6709 shca_0\u3001mlx5_0 \u4e4b\u7c7b &#061; \u6709 IB \u5361<\/li>\n<li>\u63d0\u793a No devices found &#061; \u65e0 IB \u5361<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e8c\u3001\u518d\u770b&#xff1a;IB \u5361\u72b6\u6001 &amp; \u662f\u5426\u771f\u7684\u8fde\u4e0a IB \u7f51\u7edc<\/h3>\n<h4>1. \u6700\u7b80\u547d\u4ee4&#xff1a;ibstat&#xff08;\u5fc5\u7528&#xff09;<\/h4>\n<p>ibstat<\/p>\n<p>\u770b\u5173\u952e\u884c&#xff1a;<\/p>\n<p>State: Active<br \/>\nPhysical state: LinkUp<br \/>\nRate: 100 or 200 Gb\/s<\/p>\n<ul>\n<li>State: Active &#043; LinkUp &#061; IB \u5361\u6b63\u5e38\u3001\u7f51\u7ebf\u8fde\u4e0a\u3001IB\u7f51\u7edc\u6b63\u5e38<\/li>\n<li>State: Down \/ Polling &#061; \u6ca1\u8fde\u4e0a IB \u7f51\u7edc&#xff08;\u7ebf\u6ca1\u63d2\/\u4ea4\u6362\u673a\u6ca1\u8d77\u6765\/\u6ca1\u6743\u9650&#xff09;<\/li>\n<li>Rate: 100G\/200G &#061; \u9ad8\u901f IB&#xff1b;\u666e\u901a\u5149\u7ea4\u4ee5\u592a\u7f51\u4e0d\u4f1a\u6709\u8fd9\u4e2a<\/li>\n<\/ul>\n<h4>2. \u770b\u8be6\u7ec6\u4fe1\u606f&#xff08;\u542b shca \u9a71\u52a8&#xff09;<\/h4>\n<p>ibv_devinfo<\/p>\n<p>\u91cd\u70b9\u770b&#xff1a;<\/p>\n<ul>\n<li>hca_id: shca_0 \u2192 \u5c31\u662f\u4f60\u7528\u7684\u66d9\u5149\u56fd\u4ea7 shca IB \u5361<\/li>\n<li>transport: InfiniBand \u2192 \u786e\u8ba4\u662f IB \u6a21\u5f0f<\/li>\n<li>port_state: ACTIVE \u2192 \u7aef\u53e3\u6b63\u5e38<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e09\u3001\u770b&#xff1a;IB \u7f51\u7edc\u62d3\u6251 &amp; \u540c\u7f51\u7edc\u6709\u54ea\u4e9b\u673a\u5668<\/h3>\n<p>ibnetdiscover<br \/>\nibnodes<\/p>\n<ul>\n<li>\u80fd\u770b\u5230\u5176\u4ed6\u670d\u52a1\u5668\u7684 GUID\/\u540d\u5b57 \u2192 IB \u7f51\u7edc\u662f\u901a\u7684<\/li>\n<li>\u53ea\u6709\u81ea\u5df1 \u2192 \u53ef\u80fd\u662f\u5b64\u7acb\/\u6ca1\u52a0\u5165 IB \u5b50\u7f51<\/li>\n<\/ul>\n<hr \/>\n<h3>\u56db\u3001\u770b&#xff1a;shca \u9a71\u52a8\u662f\u5426\u52a0\u8f7d&#xff08;\u5bf9\u5e94\u4f60\u8fd9\u8fb9\u7684\u73af\u5883&#xff09;<\/h3>\n<p>lsmod <span class=\"token operator\">|<\/span> <span class=\"token function\">grep<\/span> shca<\/p>\n<ul>\n<li>\u6709\u8f93\u51fa \u2192 shca \u9a71\u52a8\u5df2\u52a0\u8f7d&#xff08;\u8001\u5e08\u8ba9\u4f60\u88c5\u7684\u90a3\u4e2a&#xff09;<\/li>\n<li>\u65e0\u8f93\u51fa \u2192 \u6ca1\u88c5\/\u6ca1\u52a0\u8f7d&#xff0c;\u9700\u8981\u8dd1 mlxtoshca.sh<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e94\u3001\u6700\u7cbe\u7b80\u603b\u7ed3&#xff08;\u4f60\u53ea\u8981\u8bb0\u8fd9\u51e0\u6761&#xff09;<\/h3>\n<li>\u6709\u6ca1\u6709 IB \u5361&#xff1a;lspci | grep -i infiniband<\/li>\n<li>\u662f\u4e0d\u662f shca IB \u5361&#xff1a;ibv_devices \u770b\u6709\u6ca1\u6709 shca_0<\/li>\n<li>IB \u5361\u662f\u5426\u6b63\u5e38\u8054\u7f51&#xff1a;ibstat \u770b Active &#043; LinkUp<\/li>\n<li>shca \u9a71\u52a8\u662f\u5426\u88c5\u597d&#xff1a;lsmod | grep shca<\/li>\n<h2>\u90a3\u5982\u679c\u6ca1\u6709\u5b89\u88c5\u9a71\u52a8\u7684\u60c5\u51b5\u4e0b\u5462<\/h2>\n<p>\u6211\u7528\u6700\u7b80\u5355\u3001\u6700\u76f4\u767d\u7684\u65b9\u5f0f\u544a\u8bc9\u4f60&#xff1a;<br \/>\n\u6ca1\u88c5 IB \/ shca \u9a71\u52a8\u65f6&#xff0c;\u4f60\u770b\u5230\u7684\u7ed3\u679c\u662f\u4ec0\u4e48\u6837&#xff1f;<br \/>\n\u4ee5\u53ca \u4e3a\u4ec0\u4e48\u5fc5\u987b\u88c5\u9a71\u52a8\u624d\u80fd\u7528 IB \u9ad8\u901f\u7f51\u7edc\u3002<\/p>\n<hr \/>\n<h3>\u4e00\u3001\u6ca1\u88c5\u9a71\u52a8\u65f6&#xff0c;\u4f60\u6267\u884c\u547d\u4ee4\u4f1a\u770b\u5230\u4ec0\u4e48&#xff1f;<\/h3>\n<h3>1. ibv_devices<\/h3>\n<p>No IB devices found<\/p>\n<p>\u610f\u601d&#xff1a;\u7cfb\u7edf\u770b\u4e0d\u5230 IB \u5361&#xff08;\u56e0\u4e3a\u9a71\u52a8\u6ca1\u88c5&#xff0c;\u4e0d\u8ba4\u786c\u4ef6&#xff09;<\/p>\n<h3>2. ibstat<\/h3>\n<p>Command not found<\/p>\n<p>\u6216\u8005<\/p>\n<p>no device found<\/p>\n<h3>3. lspci | grep -i infiniband<\/h3>\n<p>\u53ef\u80fd\u80fd\u770b\u5230\u5361&#xff0c;\u4f46\u7cfb\u7edf\u4e0d\u77e5\u9053\u5b83\u662f IB \u5361<br \/>\n\u5c31\u50cf\u4f60\u63d2\u4e86\u663e\u5361&#xff0c;\u4f46\u6ca1\u88c5\u663e\u5361\u9a71\u52a8&#xff0c;\u53ea\u80fd\u5f53\u4eae\u673a\u5361\u7528\u3002<\/p>\n<h3>4. lsmod | grep shca<\/h3>\n<p>\u4ec0\u4e48\u90fd\u4e0d\u8f93\u51fa<br \/>\n&#061; \u9a71\u52a8\u6ca1\u52a0\u8f7d\u3001\u6ca1\u5b89\u88c5<\/p>\n<hr \/>\n<h3>\u4e8c\u3001\u6ca1\u88c5\u9a71\u52a8 &#061; IB \u5361\u5b8c\u5168\u4e0d\u80fd\u7528<\/h3>\n<p>\u91cd\u70b9\u6765\u4e86&#xff1a;<\/p>\n<h4>IB \u5361 \u2260 \u666e\u901a\u7f51\u5361<\/h4>\n<p>\u666e\u901a\u4ee5\u592a\u7f51&#xff08;\u5149\u7ea4\/\u7535\u53e3&#xff09;\u64cd\u4f5c\u7cfb\u7edf\u81ea\u5e26\u9a71\u52a8&#xff0c;\u63d2\u4e0a\u5c31\u80fd\u7528\u3002<\/p>\n<p>\u4f46&#xff1a;<\/p>\n<h3>IB \u5361&#xff08;\u5305\u62ec shca \u56fd\u4ea7 IB&#xff09;\u5fc5\u987b\u88c5\u4e13\u7528\u9a71\u52a8\u624d\u80fd\u5de5\u4f5c<\/h3>\n<p>\u6ca1\u9a71\u52a8 &#061; IB \u5361\u662f\u6b7b\u7684&#xff0c;\u4e0d\u4f1a\u5de5\u4f5c&#xff0c;\u4e0d\u4f1a\u8054\u7f51&#xff0c;\u4e0d\u4f1a\u52a0\u901f<\/p>\n<hr \/>\n<h3>\u4e09\u3001\u6ca1\u88c5\u9a71\u52a8\u65f6&#xff0c;\u4f60\u7684\u591a\u5361\u8bad\u7ec3\u4f1a\u600e\u6837&#xff1f;<\/h3>\n<p>\u4f60\u7528\u7684\u662f \u6d77\u5149BW 64G \u5f02\u6784\u5361<br \/>\n\u5982\u679c \u6ca1\u88c5 shca \u9a71\u52a8&#xff1a;<\/p>\n<ul>\n<li>\u7cfb\u7edf\u4e0d\u4f1a\u4f7f\u7528 IB \u9ad8\u901f\u7f51\u7edc<\/li>\n<li>\u81ea\u52a8 fallback \u5230 \u666e\u901a\u4ee5\u592a\u7f51&#xff08;\u5149\u7ea4&#xff09;<\/li>\n<li>\u591a\u5361\u8bad\u7ec3\u901f\u5ea6 \u6162 5&#xff5e;20 \u500d<\/li>\n<li>\u5927\u6a21\u578b\u6839\u672c\u8dd1\u4e0d\u8d77\u6765<\/li>\n<li>\u51fa\u73b0 \u6162\u3001\u5361\u3001\u901a\u4fe1\u8d85\u65f6<\/li>\n<\/ul>\n<p>\u4e00\u53e5\u8bdd&#xff1a;<\/p>\n<h3>\u6ca1\u9a71\u52a8 &#061; IB \u9ad8\u901f\u7f51\u5361\u5f62\u540c\u865a\u8bbe<\/h3>\n<p>\u2013<\/p>\n<h3>\u56db\u3001\u4f60\u73b0\u5728\u5fc5\u987b\u505a\u7684&#xff08;\u5c31\u4e24\u6b65&#xff09;<\/h3>\n<p>\u8001\u5e08\u7ed9\u4f60\u7684\u64cd\u4f5c&#xff0c;\u672c\u8d28\u5c31\u662f\u88c5 shca IB \u9a71\u52a8<\/p>\n<h3>\u7b2c\u4e00\u6b65&#xff08;\u88c5\u9a71\u52a8&#xff09;<\/h3>\n<p><span class=\"token function\">sudo<\/span> <span class=\"token function\">bash<\/span> mlxtoshca.sh<\/p>\n<h3>\u7b2c\u4e8c\u6b65&#xff08;\u542f\u7528 IB \u901a\u4fe1&#xff09;<\/h3>\n<p><span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">GLOO_SOCKET_IFNAME<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;eth0&#034;<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_SOCKET_IFNAME<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;eth0&#034;<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_IB_HCA<\/span><span class=\"token operator\">&#061;<\/span>shca_0:1,shca_1:1,shca_2:1,shca_3:1<br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">HSA_FORCE_FINE_GRAIN_PCIE<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_IB_DISABLE<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">0<\/span><br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_NET_PLUGIN<\/span><span class=\"token operator\">&#061;<\/span>shca<br \/>\n<span class=\"token builtin class-name\">export<\/span> <span class=\"token assign-left variable\">NCCL_DEBUG<\/span><span class=\"token operator\">&#061;<\/span>INFO<\/p>\n<hr \/>\n<h2>\u4e94\u3001\u88c5\u5b8c\u9a71\u52a8\u540e&#xff0c;\u4f60\u4f1a\u770b\u5230\u4ec0\u4e48&#xff08;\u6210\u529f\u6807\u5fd7&#xff09;<\/h2>\n<h3>1. ibv_devices<\/h3>\n<p>shca_0<br \/>\nshca_1<br \/>\nshca_2<br \/>\nshca_3<\/p>\n<h3>2. ibstat<\/h3>\n<p>State: Active<br \/>\nPhysical state: LinkUp<\/p>\n<h3>3. lsmod | grep shca<\/h3>\n<p>\u80fd\u770b\u5230\u9a71\u52a8\u6b63\u5728\u8fd0\u884c<\/p>\n<p>\u770b\u5230\u8fd9\u4e9b &#061; IB \u5361\u6b63\u5e38\u5de5\u4f5c&#xff0c;\u53ef\u4ee5\u8dd1\u9ad8\u901f\u8bad\u7ec3<\/p>\n<hr \/>\n<h2>\u8d85\u7ea7\u5927\u767d\u8bdd\u603b\u7ed3<\/h2>\n<ul>\n<li>\u6ca1\u88c5\u9a71\u52a8&#xff1a;IB \u5361\u770b\u4e0d\u89c1\u3001\u7528\u4e0d\u4e86\u3001\u8bad\u7ec3\u5f88\u6162<\/li>\n<li>\u88c5\u4e86\u9a71\u52a8&#xff1a;IB \u5361\u53d8\u6210\u9ad8\u94c1&#xff0c;\u591a\u5361\u8bad\u7ec3\u98de\u5feb<\/li>\n<li>\u4f60\u73b0\u5728\u7684\u670d\u52a1\u5668\u5fc5\u987b\u88c5 shca \u9a71\u52a8<\/li>\n<\/ul>\n<hr \/>\n","protected":false},"excerpt":{"rendered":"<p>\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 \u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3&#xff0c;\u901a\u8fc7\u65b0\u589e\u670d\u52a1\u5668\u8df3\u8fc7\u663e\u5361<br \/>\nflash_attn: disabled  HOSTQUEUE : &gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt; DUMP KERNEL ARGS: size: 160 &lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;<br \/>\n00 40 83 00 00 00 00 00 00 00 90 01 00 00 00 00<br \/>\n00 40 83 00 00 00 00 00 00 00 b4 4e e1 7f 00 0<\/p>\n","protected":false},"author":2,"featured_media":94168,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50],"topic":[],"class_list":["post-94169","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-50"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 nccl\u901a\u8baf\u9519\u8bef\uff0c\u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3\u7684\u529e\u6cd5\uff0c\u9700\u8981\u5b89\u88c5shca IB\u5361\u9a71\u52a8 - \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\/94169.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 nccl\u901a\u8baf\u9519\u8bef\uff0c\u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3\u7684\u529e\u6cd5\uff0c\u9700\u8981\u5b89\u88c5shca IB\u5361\u9a71\u52a8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 \u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3&#xff0c;\u901a\u8fc7\u65b0\u589e\u670d\u52a1\u5668\u8df3\u8fc7\u663e\u5361 flash_attn: disabled HOSTQUEUE : &gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt; DUMP KERNEL ARGS: size: 160 &lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt; 00 40 83 00 00 00 00 00 00 00 90 01 00 00 00 00 00 40 83 00 00 00 00 00 00 00 b4 4e e1 7f 00 0\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/94169.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-13T13:25:09+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813132508-6a7dc5b41cd0d.png\" \/>\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=\"7 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/94169.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/94169.html\",\"name\":\"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 nccl\u901a\u8baf\u9519\u8bef\uff0c\u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3\u7684\u529e\u6cd5\uff0c\u9700\u8981\u5b89\u88c5shca IB\u5361\u9a71\u52a8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-13T13:25:09+00:00\",\"dateModified\":\"2026-08-13T13:25:09+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/94169.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/94169.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/94169.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 nccl\u901a\u8baf\u9519\u8bef\uff0c\u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3\u7684\u529e\u6cd5\uff0c\u9700\u8981\u5b89\u88c5shca IB\u5361\u9a71\u52a8\"}]},{\"@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":"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 nccl\u901a\u8baf\u9519\u8bef\uff0c\u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3\u7684\u529e\u6cd5\uff0c\u9700\u8981\u5b89\u88c5shca IB\u5361\u9a71\u52a8 - \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\/94169.html","og_locale":"zh_CN","og_type":"article","og_title":"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 nccl\u901a\u8baf\u9519\u8bef\uff0c\u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3\u7684\u529e\u6cd5\uff0c\u9700\u8981\u5b89\u88c5shca IB\u5361\u9a71\u52a8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 \u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3&#xff0c;\u901a\u8fc7\u65b0\u589e\u670d\u52a1\u5668\u8df3\u8fc7\u663e\u5361 flash_attn: disabled HOSTQUEUE : &gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt; DUMP KERNEL ARGS: size: 160 &lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt; 00 40 83 00 00 00 00 00 00 00 90 01 00 00 00 00 00 40 83 00 00 00 00 00 00 00 b4 4e e1 7f 00 0","og_url":"https:\/\/www.wsisp.com\/helps\/94169.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-13T13:25:09+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260813132508-6a7dc5b41cd0d.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"7 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/94169.html","url":"https:\/\/www.wsisp.com\/helps\/94169.html","name":"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 nccl\u901a\u8baf\u9519\u8bef\uff0c\u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3\u7684\u529e\u6cd5\uff0c\u9700\u8981\u5b89\u88c5shca IB\u5361\u9a71\u52a8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-13T13:25:09+00:00","dateModified":"2026-08-13T13:25:09+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/94169.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/94169.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/94169.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u8d85\u7b97\u4e2d\u5fc3\u6d77\u5149\u5f02\u6784\u5361dcu bw 64G\u663e\u5361\u62a5\u9519 nccl\u901a\u8baf\u9519\u8bef\uff0c\u65e0\u6cd5\u901a\u8fc7\u8bbe\u7f6e\u6765\u89e3\u51b3\u7684\u529e\u6cd5\uff0c\u9700\u8981\u5b89\u88c5shca IB\u5361\u9a71\u52a8"}]},{"@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\/94169","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=94169"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/94169\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/94168"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=94169"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=94169"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=94169"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=94169"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}