{"id":14745,"date":"2026-11-10T02:49:34","date_gmt":"2026-11-09T18:49:34","guid":{"rendered":"https:\/\/infernews.com\/blog\/lowram\/"},"modified":"2026-11-10T03:14:32","modified_gmt":"2026-11-09T19:14:32","slug":"lowram","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/lowram\/","title":{"rendered":"LowRAM\uff1a\u7528\u9810\u6e2c\u58d3\u7e2e\u9577\u6587\u672c KV Cache"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/11\/LowRAM.jpg\" alt=\"LowRAM: up to 26.9\u00d7 KV-cache compression, retaining 94.5% of the uncompressed LongBench score on Mistral-7B prompts over\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u81ea\u56de\u6b78\u89e3\u78bc\u4f9d\u8cf4 KV \u5feb\u53d6\uff0c\u5176\u8a18\u61b6\u9ad4\u5360\u7528\u91cf\u6703\u96a8\u4e0a\u4e0b\u6587\u9577\u5ea6\u5448\u7dda\u6027\u6210\u9577\uff0c\u4e26\u9650\u5236\u6279\u6b21\u5927\u5c0f\u8207\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709\u7684\u5f8c\u8a13\u7df4\u65b9\u6cd5\u6703\u5229\u7528\u8a5e\u5143\u3001\u5c64\u8207\u7279\u5fb5\u7dad\u5ea6\u4e4b\u9593\u7684\u5197\u9918\uff0c\u6216\u964d\u4f4e\u6578\u503c\u7cbe\u78ba\u5ea6\u3002LowRAM \u627e\u5230\u4e00\u500b\u4e92\u88dc\u7684\u7dad\u5ea6\uff1a\u5728\u6bcf\u500b\u6ce8\u610f\u529b\u982d\u5167\uff0c\u503c\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53ef\u7531\u5176\u5c0d\u61c9\u7684\u9375\u9810\u6e2c\u3002\u6211\u5011\u5229\u7528\u9019\u4e00\u9ede\u63a1\u7528\u503c\u5feb\u53d6\u6524\u92b7\uff0c\u5728\u6e2c\u8a66\u6642\u50c5\u9700\u70ba\u6bcf\u500b\u6ce8\u610f\u529b\u982d\u914d\u5099\u8f15\u91cf\u7d1a\u9810\u6e2c\u5668\uff0c\u4e14\u53ea\u5132\u5b58\u5176\u91cd\u5efa\u6548\u679c\u6700\u5dee\u7684\u8a5e\u5143\u4e4b\u6b98\u5dee\uff0c\u5373\u53ef\u9054\u5230\u8fd1\u7121\u640d\u7684 4 \u500d\u503c\u58d3\u7e2e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9810\u6e2c\u5668\u53ef\u900f\u904e\u5143\u5b78\u7fd2\u521d\u59cb\u5316\uff0c\u4ee5\u5927\u5e45\u964d\u4f4e\u9810\u586b\u5145\u958b\u92b7\uff1b\u4e5f\u53ef\u5728\u5bb9\u8a31\u4e00\u6b21\u6027\u6210\u672c\u7684\u60c5\u5883\u4e0b\u5f9e\u982d\u64ec\u5408\uff0c\u4f8b\u5982\u591a\u67e5\u8a62\u63a8\u7406\u3002\u6211\u5011\u5c07\u503c\u6524\u92b7\u8207\u8de8\u5c64\u4f4e\u79e9\u9375\u58d3\u7e2e\u7d50\u5408\uff0c\u5f62\u6210 LowRAM\u2014\u2014\u4e00\u5957\u975e\u5c0d\u7a31\u58d3\u7e2e\u7ba1\u7dda\uff0c\u5728\u4e09\u500b\u6a21\u578b\u7cfb\u5217\u4e2d\u7686\u80fd\u7dad\u6301\u63a5\u8fd1\u7121\u640d\u7684\u8868\u73fe\u3002LowRAM \u5728\u6bcf\u500b\u89e3\u78bc\u6b65\u9a5f\u50c5\u91cd\u5efa\u9078\u5b9a\u5b50\u96c6\u7684 KV \u914d\u5c0d\uff0c\u56e0\u6b64\u80fd\u5c07\u89e3\u78bc\u958b\u92b7\u964d\u81f3\u6700\u4f4e\uff0c\u4e26\u5728 4 \u500d\u58d3\u7e2e\u4e0b\u5be6\u73fe\u6700\u9ad8 5 \u500d\u7684\u5cf0\u503c\u89e3\u78bc\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u7684 CompressedCache \u53ef\u63a5\u5165\u652f\u63f4 Transformers \u7684\u6a21\u578b\u6d41\u7a0b\uff0c<mark>keys \u8207 values \u80fd\u5206\u5225\u8a2d\u5b9a\u58d3\u7e2e\u65b9\u5f0f<\/mark>\uff0c\u4e0d\u9650\u65bc\u55ae\u7d14\u91cf\u5316\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8de8\u5c64\u58d3\u7e2e keys\uff0c\u6e1b\u5c11\u91cd\u8907\u5132\u5b58\u3002<\/li>\n\n\n\n<li>\u53ef\u914d\u5408\u91cf\u5316\u53ca token eviction\uff0c\u9032\u4e00\u6b65\u7e2e\u5c0f\u5feb\u53d6\u3002<\/li>\n\n\n\n<li>\u63d0\u4f9b\u9810\u8a13\u7df4\u9810\u6e2c\u5668\u53ca Llama\u3001Mistral \u7bc4\u4f8b\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6e2c\u8a66\u524d\u9808\u6e96\u5099 Python 3.12\u3001CUDA GPU\u3001Git LFS \u53ca\u6a21\u578b\u5b58\u53d6\u6b0a\u9650\uff1b\u9810\u8a2d Llama-3.1-8B-Instruct\u3002\u6587\u4ef6\u5efa\u8b70\u7531 4K-token \u7bc4\u4f8b\u958b\u59cb\uff0c\u4f46\u63d0\u4f9b\u7684\u7bc0\u9304\u672a\u5305\u542b\u5b8c\u6574\u5b89\u88dd\u53ca\u57f7\u884c\u6b65\u9a5f\uff1b\u7db2\u7ad9\u7684 128K \u5c55\u793a\u5247\u662f\u9304\u88fd\u7d50\u679c\uff0c\u4e26\u975e\u5373\u6642\u63a8\u7406\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Mistral-7B-Instruct-v0.3 \u5728\u8d85\u904e 10K tokens \u7684\u6e2c\u8a66\u4e2d\uff0c\u7d50\u5408\u91cf\u5316\u8207\u6dd8\u6c70\u53ef\u9054 26.9 \u500d\u5feb\u53d6\u58d3\u7e2e\uff0c\u4fdd\u7559\u672a\u58d3\u7e2e LongBench \u5206\u6578\u7684 94.5%\u3002\u9019\u4e0d\u662f\u6e96\u78ba\u7387\uff0c\u4e5f\u4e0d\u4ee3\u8868\u6574\u9ad4 GPU \u8a18\u61b6\u9ad4\u7e2e\u5c0f\u540c\u6a23\u500d\u6578\uff1b\u5718\u968a\u4ecd\u9808\u4ee5\u81ea\u5df1\u7684\u6587\u4ef6\u6aa2\u67e5\u7b54\u6848\u7d30\u7bc0\u8207\u901f\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/lowram-model.github.io\/\" rel=\"noopener\" target=\"_blank\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/lowram-model\/lowram\" rel=\"noopener\" target=\"_blank\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9577\u6587\u672c\u63a8\u7406\u60f3\u6173\u986f\u793a\u8a18\u61b6\u9ad4\uff0c\u4e0d\u4e00\u5b9a\u8981\u4fdd\u7559\u6bcf\u7b46\u5feb\u53d6\u3002LowRAM \u7528\u9810\u6e2c\u91cd\u5efa\u8cc7\u6599\uff0c\u4f46\u58d3\u7e2e\u5e45\u5ea6\u8207\u7b54\u6848\u8cea\u7d20\u4ecd\u8981\u53d6\u6368\u3002<\/p>\n","protected":false},"author":8,"featured_media":14744,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,76,127,179,38,40,152],"tags":[],"class_list":["post-14745","post","type-post","status-publish","format-standard","hentry","category-133","category-76","category-127","category-nvidia","category-38","category-llama","category-python"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/14745","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=14745"}],"version-history":[{"count":3,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/14745\/revisions"}],"predecessor-version":[{"id":14756,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/14745\/revisions\/14756"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/14744"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=14745"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=14745"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=14745"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}