{"id":14775,"date":"2026-11-11T03:53:24","date_gmt":"2026-11-10T19:53:24","guid":{"rendered":"https:\/\/infernews.com\/blog\/infrarsi\/"},"modified":"2026-11-11T03:57:26","modified_gmt":"2026-11-10T19:57:26","slug":"infrarsi","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/infrarsi\/","title":{"rendered":"InfraRSI\uff1a\u81ea\u52d5\u52a0\u901f LLM \u8a13\u7df4\u7684 AI Agent"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/11\/InfraRSI-LLM-AI-Agent.jpg\" alt=\"InfraRSI Logo Banner\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u50b3\u7d71\u7684\u8abf\u512a\u5de5\u5177\u591a\u534a\u805a\u7126\u5728\u8d85\u53c3\u6578\u641c\u5c0b\uff08Knob Tuning\uff09\u2014 \u4f8b\u5982\u5617\u8a66\u8abf\u6574 Batch Size\u3001Gradient Accumulation Steps \u6216\u662f\u5e73\u884c\u5ea6\u8a2d\u5b9a\u3002\u7136\u800c\u5728\u73fe\u4ee3 GPU \u53e2\u96c6\u4e0a\uff0c\u771f\u6b63\u7684\u6548\u80fd\u74f6\u9838\u5f80\u5f80\u6df1\u690d\u65bc\u7a0b\u5f0f\u78bc\u5c64\u7d1a\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f55\u6642\u8a72\u5c0e\u5165\u878d\u5408\u7b97\u5b50 (Fused Kernels)\uff1f\u5982\u4f55\u6b63\u78ba\u555f\u7528 Flash Attention \u8207 torch.compile \u7de8\u8b6f\u6ce8\u610f\u529b\u6a5f\u5236\uff1f\u5982\u4f55\u5728 Activation Checkpointing\u3001Offload \u8207\u904b\u7b97\u901f\u5ea6\u4e4b\u9593\u53d6\u5f97\u5e73\u8861\uff0c\u907f\u514d\u9867\u4e86\u986f\u5b58\u537b\u72a7\u7272\u541e\u5410\u91cf\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">InfraRSI \u7a81\u7834\u50b3\u7d71\u7d14\u8abf\u53c3\uff08Knob Tuning\uff09\u7684\u5c40\u9650\uff0c\u81ea\u52d5\u5f15\u5165\u878d\u5408\u7b97\u5b50\uff08Fused Kernels\uff09\u3001\u7de8\u8b6f\u6ce8\u610f\u529b\u6a5f\u5236\uff08torch.compile\uff09\u53ca\u986f\u5b58\u5e73\u8861\u7b56\u7565\u3002\u5b98\u65b9\u5be6\u6e2c\u5e73\u5747\u63d0\u901f\u9054 1.83\u00d7\uff0cMFU \u7531 17.4% \u5927\u5e45\u62c9\u5347\u81f3 30.0%\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">RSI \u7d93\u56b4\u683c\u9a57\u8b49\u7684\u6210\u529f\u65b9\u6848\u6703\u5beb\u5165\u77e5\u8b58\u5eab\u3002\u8de8 GPU \u786c\u4ef6\u9077\u79fb\u6642\u80fd\u76f4\u63a5\u6cbf\u7528\u6b77\u53f2\u7d93\u9a57\uff0c\u5927\u5e45\u7e2e\u77ed\u6478\u7d22\u81f3\u6548\u80fd\u5cf0\u503c\u7684\u6642\u9593\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56b4\u8b39\u9a57\u8b49\u7684\u6a5f\u5236\u5305\u62ec\u5167\u7f6e 16 \u9805\u4e0d\u8b8a\u91cf\u6aa2\u9a57\uff0c\u78ba\u4fdd\u6578\u64da\u91cf\u3001Optimizer Steps \u5b8c\u5168\u4e00\u81f4\u4e14 Loss \u504f\u5dee\u5728\u5bb9\u8a31\u7bc4\u570d\u5167\uff1b\u914d\u5408\u96d9\u6a21\u578b\u4ea4\u53c9\u5be9\u67e5\uff08\u4e00\u500b\u5beb Patch\u3001\u53e6\u4e00\u500b Code Review\uff09\u53ca\u80cc\u9760\u80cc\u8907\u6e2c\uff0c\u675c\u7d55\u8a13\u7df4\u7570\u5e38\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5ee3\u6cdb\u652f\u63f4\u4e26\u76f8\u5bb9 Megatron-Core\u3001FSDP2\u3001DeepSpeed \u7b49\u4e3b\u6d41\u6846\u67b6\uff1b\u5f8c\u7aef\u53ef\u76f4\u63a5\u4e32\u63a5\u672c\u6a5f\u5df2\u767b\u5165\u7684 Claude Code\u3001Codex CLI\uff0c\u6216\u5404\u985e OpenAI \u76f8\u5bb9 API \u8207\u81ea\u5efa vLLM\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5feb\u901f\u4e0a\u624b\uff0c\u57f7\u884c<mark>\u55ae\u4e00 Bash \u6307\u4ee4\u5373\u53ef\u958b\u59cb\u904b\u4f5c<\/mark>\uff1a<\/p>\n\n\n\n<div class=\"codecopy-container\">\n                <div class=\"codecopy-header\">\n                    \n                    <button class=\"codecopy-btn\" data-target=\"codecopy-6af39a7404cf7\" title=\"Copy code to clipboard\">\n                <svg width=\"16\" height=\"16\" viewBox=\"0 0 24 24\" fill=\"none\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\">\n                    <path d=\"M16 1H4C2.9 1 2 1.9 2 3V17H4V3H16V1ZM19 5H8C6.9 5 6 5.9 6 7V21C6 22.1 6.9 23 8 23H19C20.1 23 21 22.1 21 21V7C21 5.9 20.1 5 19 5ZM19 21H8V7H19V21Z\" fill=\"currentColor\"\/>\n                <\/svg>\n                <span class=\"codecopy-text\">Copy<\/span>\n            <\/button>\n                <\/div>\n                <pre class=\"line-numbers\" id=\"codecopy-6af39a7404cf7\"><code class=\"language-text\">infrarsi optimize .\/train.sh --llm codex-cli<\/code><\/pre>\n            <\/div>\n\n\n\n<p class=\"wp-block-paragraph\">\u555f\u52d5\u6642\u6703\u900f\u904e\u4e92\u52d5\u5f0f\u554f\u7b54\u78ba\u8a8d\u8a2d\u5b9a\uff0c\u4e26\u652f\u63f4\u4ee5\u81ea\u7136\u8a9e\u8a00\u4e0b\u9054\u9650\u5236\uff08\u4f8b\u5982\u300c\u4e0d\u80fd\u6539\u52d5 optimizer\u300d\u3001\u300cbatch size \u4e0d\u53ef\u5927\u904e 4\u300d\uff09\u3002\u7cfb\u7d71\u9a57\u8b49\u5b8c\u6210\u5f8c\uff0c\u6703\u76f4\u63a5\u8f38\u51fa\u6700\u4f73\u5316\u8173\u672c infrarsi_optimized.sh \u8207\u5be9\u6838\u5b8c\u7562\u7684\u7a0b\u5f0f\u78bc\u88dc\u4e01\uff0c\u8b93\u5718\u968a\u4ee5\u6975\u4f4e\u9580\u6abb\u69a8\u76e1 GPU \u7b97\u529b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/infrarsi.github.io\/\" rel=\"noopener\" target=\"_blank\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/TokenRhythm\/InfraRSI\" rel=\"noopener\" target=\"_blank\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>InfraRSI \u662f\u4e00\u500b\u5c08\u6ce8\u65bc\u5927\u8a9e\u8a00\u6a21\u578b\uff08LLM\uff09\u8a13\u7df4\u57fa\u790e\u8a2d\u65bd\u512a\u5316\u7684\u81ea\u4e3b Agent\uff08Harness\uff09\u3002\u5176\u6838\u5fc3\u7406\u5ff5\u662f\u5229\u7528\u905e\u6b78\u81ea\u6211\u6539\u9032\uff08Recursive Self-Improvement, RSI\uff09\u6280\u8853\uff0c\u81ea\u52d5\u70ba\u65e2\u6709\u7684\u8a13\u7df4\u4efb\u52d9\u9032\u884c\u6548\u80fd\u8abf\u512a\u8207\u7a0b\u5f0f\u78bc\u5c64\u7d1a\u7684\u6700\u4f73\u5316\u3002<\/p>\n","protected":false},"author":8,"featured_media":14774,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,116,127,185,153,147,38,152,150,191],"tags":[],"class_list":["post-14775","post","type-post","status-publish","format-standard","hentry","category-133","category-agentic","category-127","category-qwen","category-openai","category-deepseek","category-38","category-python","category-150","category-anthropic"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/14775","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=14775"}],"version-history":[{"count":3,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/14775\/revisions"}],"predecessor-version":[{"id":14781,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/14775\/revisions\/14781"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/14774"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=14775"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=14775"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=14775"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}