{"id":13004,"date":"2026-10-07T05:17:44","date_gmt":"2026-10-06T21:17:44","guid":{"rendered":"https:\/\/infernews.com\/blog\/xllm-loop\/"},"modified":"2026-10-07T05:17:44","modified_gmt":"2026-10-06T21:17:44","slug":"xllm-loop","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/xllm-loop\/","title":{"rendered":"xllm-loop\uff1a\u5faa\u74b0\u8a9e\u8a00\u6a21\u578b"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/pasted-4f81623804a1.jpg\" alt=\"Repository image for ifm-ai\/xllm-loop\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u60f3\u7528\u8f03\u5c11\u6a21\u578b\u53c3\u6578\u548c KV cache\uff0c\u63db\u53d6\u66f4\u6df1\u7684\u8a08\u7b97\u80fd\u529b\uff0cxLLM-Loop \u63d0\u4f9b\u4e86\u4e00\u689d\u53ef\u76f4\u63a5\u7814\u7a76\u7684\u8def\u7dda\u3002\u5b83\u5c6c\u65bc\u958b\u6e90\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u6846\u67b6\u64f4\u5145\uff0c\u8655\u7406\u7684\u662f\u5faa\u74b0\u8a9e\u8a00\u6a21\u578b\u5982\u4f55\u91cd\u7528\u5c64\u3001\u63a7\u5236\u6df1\u5ea6\uff0c\u4ee5\u53ca\u8a55\u4f30\u56fa\u5b9a\u9ede\u5e36\u4f86\u7684\u6548\u7387\u53d6\u6368\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">LoopedTransformer \u6703\u91cd\u8907\u57f7\u884c\u6307\u5b9a\u5c64\u6bb5\uff0cHuginn \u5247\u628a\u6a21\u578b\u5206\u6210 prelude\u3001recurrent block \u548c coda\uff0c\u4e26\u652f\u63f4\u55ae\u4e00\u6216\u968e\u5c64\u5f0f recurrent state\u3002Part II \u9032\u4e00\u6b65\u7814\u7a76 fixed-point shortcuts\u3001learned depth prior \u548c orthogonal injection\uff0c\u76ee\u6a19\u662f\u5728\u589e\u52a0 FLOPs \u6642\u7dad\u6301\u8fd1\u4e4e\u56fa\u5b9a\u7684\u8a18\u61b6\u9ad4\u9700\u6c42\uff1b\u76f8\u95dc\u8aaa\u6cd5\u5305\u62ec\u8f03\u6a19\u6e96 Transformer \u5c11\u7d04\u4e09\u500d\u53c3\u6578\u548c KV cache\uff0c\u4f46\u4e0d\u80fd\u8996\u70ba\u6240\u6709\u4efb\u52d9\u90fd\u80fd\u9054\u6210\u7684\u4fdd\u8b49\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u5df2\u63a5\u5165 xLLM\uff0c\u63d0\u4f9b\u5169\u7bc7\u7814\u7a76\u5de5\u4f5c\u7684 training\u3001evaluation \u548c checkpoint \u7a0b\u5f0f\u78bc\uff0c\u4e26\u6709 49 \u500b Hugging Face checkpoints\u3002Part I \u6db5\u84cb Dense \u548c MoE \u5171 17 \u7a2e recipes\u300125 \u500b checkpoints\uff1bPart II \u6db5\u84cb depth priors\u3001input-injection variants \u548c distilled prefill\uff0c\u5171 24 \u500b checkpoints\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u9069\u5408\u7814\u7a76\u5faa\u74b0\u67b6\u69cb\u3001\u56fa\u5b9a\u9ede\u63a8\u7406\u548c\u53c3\u6578\u6548\u7387\u7684\u6a21\u578b\u5718\u968a<\/li>\n<li>\u53ef\u6bd4\u8f03\u56fa\u5b9a\u3001\u62bd\u6a23\u53ca learned depth \u7684\u8a13\u7df4\u65b9\u5f0f<\/li>\n<li>\u652f\u63f4 LoopedTransformer\u3001Huginn \u548c DepthControlledHuginn<\/li>\n<li>\u9700\u8981 PyTorch 2.11\u3001CUDA 12.8\u3001Python 3.12\uff0c\u5be6\u9a57\u66fe\u4f7f\u7528 NVIDIA H200<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b89\u88dd\u65b9\u5f0f\u6cbf\u7528 xLLM\uff0c\u4f46\u786c\u4ef6\u548c\u8edf\u4ef6\u7248\u672c\u8981\u6c42\u504f\u65b0\uff0c\u8a13\u7df4\u6210\u672c\u4ea6\u4e0d\u4f4e\uff0c\u8ad6\u6587\u5be6\u9a57\u662f\u5728 NVIDIA H200 \u4e0a\u5b8c\u6210\u3002\u5c0d\u60f3\u5728\u672c\u5730\u76f4\u63a5\u4f7f\u7528\u6210\u719f\u804a\u5929\u6a21\u578b\u7684\u4eba\uff0c\u5b83\u4e0d\u662f\u5373\u88dd\u5373\u7528\u7684\u7522\u54c1\uff1b\u5c0d\u9700\u8981\u91cd\u73fe\u7814\u7a76\u3001\u8f09\u5165 checkpoints \u6216\u6539\u9020\u8a13\u7df4\u6d41\u7a0b\u7684\u5718\u968a\uff0c\u7a0b\u5f0f\u78bc\u3001\u8a55\u4f30\u53ca checkpoint \u914d\u5957\u5c31\u66f4\u6709\u50f9\u503c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/x.com\/huskydogewoof\/status\/2106130737007091939\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/ifm-ai\/xllm-loop\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/collections\/IFM\/towards-looped-models-done-right\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>xLLM-Loop \u628a\u540c\u4e00\u7d44 Transformer \u5c64\u91cd\u8907\u904b\u7b97\uff0c\u63a2\u7d22\u4ee5\u8f03\u5c11\u53c3\u6578\u548c KV cache \u7dad\u6301\u8a9e\u8a00\u6a21\u578b\u80fd\u529b\u3002\u4ee5\u66f4\u5c11\u53c3\u6578\u63db\u53d6\u66f4\u5927\u63a8\u7406\u6df1\u5ea6\u3002<\/p>\n","protected":false},"author":8,"featured_media":13003,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,76,127,179,197,152],"tags":[],"class_list":["post-13004","post","type-post","status-publish","format-standard","hentry","category-133","category-76","category-127","category-nvidia","category-framework","category-python"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13004","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=13004"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13004\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13003"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13004"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13004"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13004"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}