{"id":23884,"date":"2026-06-10T14:50:28","date_gmt":"2026-06-10T06:50:28","guid":{"rendered":"https:\/\/ab123.xyz\/?p=23884"},"modified":"2026-06-10T14:50:28","modified_gmt":"2026-06-10T06:50:28","slug":"amd-mi300x-instinct-%e5%8a%a0%e9%80%9f%e5%8d%a1-rocm-%e9%a9%b1%e5%8a%a8%e5%85%bc%e5%ae%b9%e6%80%a7%e6%8c%87%e5%8d%97","status":"publish","type":"post","link":"https:\/\/ab123.xyz\/?p=23884","title":{"rendered":"AMD MI300X Instinct \u52a0\u901f\u5361 ROCm \u9a71\u52a8\u517c\u5bb9\u6027\u6307\u5357"},"content":{"rendered":"<p>AMD MI300X Instinct \u52a0\u901f\u5361\u662f\u9762\u5411\u9ad8\u6027\u80fd\u8ba1\u7b97\u4e0e\u4eba\u5de5\u667a\u80fd\u8bad\u7ec3\u7684\u65b0\u4e00\u4ee3\u65d7\u8230\u4ea7\u54c1\uff0c\u5176\u5f3a\u5927\u7684\u786c\u4ef6\u6027\u80fd\u9700\u8981\u4e0e\u7a33\u5b9a\u7684\u8f6f\u4ef6\u6808\u914d\u5408\u624d\u80fd\u5145\u5206\u53d1\u6325\u3002\u672c\u6587\u63d0\u4f9b\u4e00\u4efd\u6743\u5a01\u7684 ROCm \u9a71\u52a8\u517c\u5bb9\u6027\u6307\u5357\uff0c\u5e2e\u52a9\u5f00\u53d1\u8005\u548c\u8fd0\u7ef4\u4eba\u5458\u5feb\u901f\u5b8c\u6210\u73af\u5883\u90e8\u7f72\u4e0e\u8c03\u4f18\u3002\u5b98\u65b9\u6587\u6863\u4e0e\u9a71\u52a8\u4e0b\u8f7d\u5165\u53e3\u8bf7\u8bbf\u95ee\uff1a<a href=\"https:\/\/rocm.docs.amd.com\" target=\"_blank\">\u5b98\u65b9\u7f51\u7ad9<\/a>\u3002<\/p>\n<h2>ROCm \u9a71\u52a8\u7248\u672c\u4e0e\u786c\u4ef6\u5339\u914d<\/h2>\n<p>MI300X \u57fa\u4e8e CDNA 3 \u67b6\u6784\uff0c\u5bf9 ROCm \u7248\u672c\u6709\u660e\u786e\u8981\u6c42\u3002\u5efa\u8bae\u4f7f\u7528 ROCm 6.0 \u53ca\u4ee5\u4e0a\u7248\u672c\uff0c\u8be5\u7248\u672c\u539f\u751f\u652f\u6301 MI300X \u7684\u77e9\u9635\u6838\u5fc3\u4e0e Infinity Fabric \u4e92\u8054\u7279\u6027\u3002\u65e7\u7248 ROCm 5.x \u867d\u53ef\u5b89\u88c5\uff0c\u4f46\u65e0\u6cd5\u5b8c\u6574\u8c03\u7528\u786c\u4ef6\u52a0\u901f\u5355\u5143\uff0c\u5bfc\u81f4\u6027\u80fd\u635f\u5931\u3002<\/p>\n<h3>\u63a8\u8350\u9a71\u52a8\u7248\u672c\u5bf9\u7167<\/h3>\n<ul>\n<li>ROCm 6.0.2\uff1a\u7a33\u5b9a\u7248\uff0c\u652f\u6301 PyTorch 2.1 \u4e0e TensorFlow 2.13\u3002<\/li>\n<li>ROCm 6.1.0\uff1a\u9884\u89c8\u7248\uff0c\u65b0\u589e\u5bf9 Flash Attention 2 \u7684\u786c\u4ef6\u4f18\u5316\u3002<\/li>\n<li>ROCm 5.7.x\uff1a\u4ec5\u7528\u4e8e\u9a8c\u8bc1\uff0c\u4e0d\u63a8\u8350\u751f\u4ea7\u73af\u5883\u3002<\/li>\n<\/ul>\n<h2>\u64cd\u4f5c\u7cfb\u7edf\u4e0e\u5185\u6838\u517c\u5bb9\u6027<\/h2>\n<p>ROCm \u9a71\u52a8\u4f9d\u8d56\u4e8e Linux \u5185\u6838\u7248\u672c\u53ca GPU \u56fa\u4ef6\u3002\u7ecf\u6d4b\u8bd5\uff0c\u4ee5\u4e0b\u7cfb\u7edf\u73af\u5883\u53ef\u786e\u4fdd\u517c\u5bb9\uff1a<\/p>\n<ul>\n<li>Ubuntu 22.04 LTS\uff08\u5185\u6838 6.2+\uff09\u2014\u2014\u5b98\u65b9\u9996\u9009\u3002<\/li>\n<li>RHEL 9.2 \u53ca\u4ee5\u4e0a\u2014\u2014\u9700\u5b89\u88c5\u989d\u5916\u7684 kernel-devel \u5305\u3002<\/li>\n<li>SLES 15 SP4\u2014\u2014\u652f\u6301\u4f46\u9700\u624b\u52a8\u52a0\u8f7d amdgpu \u6a21\u5757\u3002<\/li>\n<\/ul>\n<h3>\u5185\u6838\u53c2\u6570\u8c03\u6574<\/h3>\n<p>\u5b89\u88c5\u524d\u9700\u5728 grub \u4e2d\u6dfb\u52a0 <code>amdgpu.ppfeaturemask=0xffffffff<\/code> \u4ee5\u542f\u7528\u5b8c\u6574\u7535\u6e90\u7ba1\u7406\u529f\u80fd\uff0c\u5426\u5219\u53ef\u80fd\u5bfc\u81f4\u98ce\u6247\u8f6c\u901f\u5f02\u5e38\u3002\u540c\u65f6\u5efa\u8bae\u5173\u95ed secure boot\uff0c\u907f\u514d\u9a71\u52a8\u7b7e\u540d\u51b2\u7a81\u3002<\/p>\n<h2>\u5e38\u89c1\u517c\u5bb9\u6027\u95ee\u9898\u4e0e\u89e3\u51b3\u65b9\u6848<\/h2>\n<h3>\u9a71\u52a8\u5b89\u88c5\u5931\u8d25<\/h3>\n<p>\u82e5\u9047\u5230 amdgpu-dkms \u7f16\u8bd1\u9519\u8bef\uff0c\u8bf7\u68c0\u67e5 gcc \u7248\u672c\u662f\u5426\u4f4e\u4e8e 10\u3002\u9700\u5c06 gcc \u5347\u7ea7\u81f3 10.2+ \u5e76\u540c\u6b65\u66f4\u65b0 linux-headers\u3002<\/p>\n<h3>\u591a\u5361\u901a\u4fe1\u5f02\u5e38<\/h3>\n<p>MI300X \u901a\u8fc7 Infinity Fabric \u4e92\u8054\u65f6\uff0cNVLink \u7c7b\u4f3c\u529f\u80fd\u4f9d\u8d56\u4e8e ROCm \u7684 thunk \u5e93\u3002\u82e5\u51fa\u73b0\u65e0\u6cd5\u68c0\u6d4b\u6240\u6709 GPU\uff0c\u8bf7\u786e\u8ba4 BIOS \u4e2d Above 4G Decoding \u548c Resizable BAR \u5747\u5f00\u542f\u3002<\/p>\n<h3>PyTorch \u8fd0\u884c\u62a5\u9519<\/h3>\n<p>\u4f7f\u7528 torch.compile \u65f6\u53ef\u80fd\u9047\u5230 HIP \u4ee3\u7801\u751f\u6210\u9519\u8bef\uff0c\u964d\u7ea7 PyTorch \u81f3 2.0.1 \u6216\u7b49\u5f85\u5b98\u65b9\u70ed\u4fee\u590d\u8865\u4e01\u3002\u4e5f\u53ef\u901a\u8fc7\u8bbe\u7f6e\u73af\u5883\u53d8\u91cf <code>HSA_OVERRIDE_GFX_VERSION=11.0.0<\/code> \u4e34\u65f6\u89c4\u907f\u3002<\/p>\n<h2>\u6027\u80fd\u8c03\u4f18\u4e0e\u5de5\u5177\u94fe<\/h2>\n<p>ROCm \u63d0\u4f9b rocprof \u4e0e rocminfo \u5de5\u5177\u8fdb\u884c\u6027\u80fd\u5256\u6790\u3002\u9488\u5bf9 MI300X\uff0c\u4ee5\u4e0b\u53c2\u6570\u53ef\u663e\u8457\u63d0\u5347\u8bad\u7ec3\u541e\u5410\uff1a<\/p>\n<ul>\n<li>\u8bbe\u7f6e <code>GPU_MAX_HEAP_SIZE=100%<\/code> \u907f\u514d\u663e\u5b58\u788e\u7247\u3002<\/li>\n<li>\u4f7f\u7528 HIP_VISIBLE_DEVICES \u6309\u4efb\u52a1\u5206\u914d GPU \u6838\u5fc3\u3002<\/li>\n<li>\u5f00\u542f rocBLAS \u7684 TF32 \u6a21\u5f0f\u4ee5\u5e73\u8861\u7cbe\u5ea6\u4e0e\u901f\u5ea6\u3002<\/li>\n<\/ul>\n<p>\u66f4\u591a\u4f18\u5316\u7ec6\u8282\u53ef\u53c2\u8003 AMD \u5b98\u65b9\u6027\u80fd\u6307\u5357\u3002\u5efa\u8bae\u5b9a\u671f\u8bbf\u95ee <a href=\"https:\/\/rocm.docs.amd.com\" target=\"_blank\">\u5b98\u65b9\u7f51\u7ad9<\/a> \u83b7\u53d6\u6700\u65b0\u7684\u517c\u5bb9\u6027\u5217\u8868\u4e0e\u9a71\u52a8\u66f4\u65b0\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>AMD MI300X Instinct \u52a0\u901f\u5361\u662f\u9762\u5411\u9ad8\u6027\u80fd\u8ba1\u7b97\u4e0e\u4eba\u5de5\u667a\u80fd\u8bad\u7ec3\u7684\u65b0\u4e00\u4ee3\u65d7\u8230\u4ea7\u54c1\uff0c\u5176\u5f3a\u5927\u7684\u786c\u4ef6\u6027\u80fd [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[26833,26837,26835,26834,26836],"class_list":["post-23884","post","type-post","status-publish","format-standard","hentry","category-4","tag-amd-mi300x","tag-gpu","tag-instinct","tag-rocm","tag-26836"],"_links":{"self":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/23884","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=23884"}],"version-history":[{"count":1,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/23884\/revisions"}],"predecessor-version":[{"id":23885,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/23884\/revisions\/23885"}],"wp:attachment":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=23884"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=23884"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=23884"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}