
{"id":9786,"date":"2026-06-30T03:54:19","date_gmt":"2026-06-29T19:54:19","guid":{"rendered":"https:\/\/infernews.com\/blog\/mhf\/"},"modified":"2026-06-30T03:55:14","modified_gmt":"2026-06-29T19:55:14","slug":"mhf","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/mhf\/","title":{"rendered":"MultiHashFormer\uff1a\u7528\u96dc\u6e4a\u91cd\u5beb\u8a9e\u8a00\u6a21\u578b\u8a5e\u8868"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/pasted-e915a610f92d.jpg\" alt=\"Repository image for HUIYINXUE\/MHF\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">MultiHashFormer \u662f\u4e00\u500b\u751f\u6210\u5f0f\u8a9e\u8a00\u6a21\u578b\u7814\u7a76\u9805\u76ee\uff0c\u540c\u6642\u63d0\u4f9b Qwen3 \u76f8\u95dc\u5be6\u4f5c\u3001\u8a13\u7df4\u8173\u672c\u8207\u8a5e\u5f59\u64f4\u5145\u6d41\u7a0b\u3002\u5b83\u8981\u89e3\u6c7a\u7684\u662f\u50b3\u7d71 embedding matrix \u6703\u96a8 vocabulary size \u7dda\u6027\u81a8\u8139\uff0c\u4ee4\u6a21\u578b\u96e3\u4ee5\u7528\u56fa\u5b9a\u53c3\u6578\u91cf\u5438\u6536\u66f4\u591a\u8a5e\u5f59\u3001\u8a9e\u7a2e\u6216\u65b0\u9818\u57df\u5167\u5bb9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709 hash-based \u505a\u6cd5\u591a\u6578\u63a1\u7528 many-to-one mappings\uff0c\u628a\u591a\u500b token \u58d3\u5230\u540c\u4e00\u500b hash index\uff0c\u9019\u5728 encoder-only \u6a21\u578b\u5c1a\u53ef\u904b\u4f5c\uff0c\u4f46\u653e\u5230 causal LMs \u5c31\u6703\u51fa\u73fe\u89e3\u78bc\u6b67\u7fa9\uff1a\u6a21\u578b\u9810\u6e2c\u5230\u5171\u4eab index\uff0c\u672a\u5fc5\u80fd\u6e96\u78ba\u9084\u539f\u539f\u4f86\u90a3\u500b token\u3002MultiHashFormer \u7684\u505a\u6cd5\u662f\u70ba\u6bcf\u500b token \u5efa\u7acb unique hash signature\uff0c\u7528\u591a\u500b\u7368\u7acb hash functions \u7522\u751f\u4e00\u4e32\u96e2\u6563 hash IDs\uff0c\u518d\u4ea4\u7531 Hash Encoder \u58d3\u6210 latent vector\uff0c\u6700\u5f8c\u7531 Hash Decoder \u751f\u6210\u4e0b\u4e00\u500b token \u7684 hash signature\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u8a2d\u8a08\u7684\u53d6\u5411\u5f88\u660e\u78ba\uff1a\u5b83\u4e0d\u662f\u55ae\u7d14\u7e2e\u5c0f embedding\uff0c\u800c\u662f\u91cd\u7d44\u300ctoken \u5982\u4f55\u8868\u793a\u3001\u5982\u4f55\u751f\u6210\u300d\u9019\u689d\u8def\u5f91\uff0c\u76ee\u6a19\u662f\u5728\u4fdd\u6301\u53c3\u6578 footprint \u56fa\u5b9a\u7684\u524d\u63d0\u4e0b\uff0c\u4ecd\u53ef\u505a autoregression\u3002\u4f86\u6e90\u8cc7\u6599\u4ea6\u986f\u793a\u4f5c\u8005\u628a\u5b83\u653e\u5230 100M\u30011B\u30013B \u898f\u6a21\uff0c\u4e26\u63d0\u4f9b standard \u8207 MHF \u5169\u7d44\u8a13\u7df4\u8173\u672c\uff0c\u65b9\u4fbf\u76f4\u63a5\u5c0d\u7167 baseline\uff0c\u4e0d\u904e README \u672a\u5b8c\u6574\u5217\u51fa\u6240\u6709 benchmark \u6578\u5b57\uff0c\u95b1\u8b80\u6642\u61c9\u4ee5\u8ad6\u6587\u7d50\u679c\u70ba\u6e96\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u90e8\u7f72\u7406\u89e3\u4e0a\uff0c\u9019\u500b\u9805\u76ee\u6bd4\u8f03\u63a5\u8fd1\u7814\u7a76\u4ee3\u78bc\u800c\u975e\u5373\u88dd\u5373\u7528\u7522\u54c1\uff1apreprocessing \u5167\u6709\u82f1\u6587\u9810\u8a13\u7df4\u8cc7\u6599\u8655\u7406\u8173\u672c\uff0ctraining \u5167\u5206 standard \u8207 MHF \u8a13\u7df4\u6d41\u7a0b\uff0cvocab_expansion \u5247\u6db5\u84cb tokenizer \u8a13\u7df4\u3001\u8cc7\u6599\u6e96\u5099\u3001continual pretraining \u8207 expanded_tokenizer\u3002\u4f9d\u8cf4\u5305\u62ec transformers\u3001flash_attn\u3001tokenizers\u3001lm_eval \u8207 mmh3\uff0c\u4ee3\u8868\u5b83\u9762\u5411\u7684\u662f\u5df2\u6709 Python \u6df1\u5ea6\u5b78\u7fd2\u74b0\u5883\u3001\u60f3\u91cd\u73fe\u8ad6\u6587\u6216\u6e2c\u8a66\u8a5e\u5f59\u64f4\u5145\u7684\u4eba\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u9805\u76ee\u985e\u578b<\/strong>\uff1a\u7814\u7a76\u539f\u578b\u517c\u6a21\u578b\u8a13\u7df4\u4ee3\u78bc\uff0c\u6838\u5fc3\u662f hash-based autoregressive language modeling\u3002<\/li>\n<li><strong>\u4e3b\u8981\u5dee\u7570<\/strong>\uff1a\u4e0d\u518d\u7528 many-to-one hashing \u76f4\u63a5\u4ee3\u8868 token\uff0c\u800c\u662f\u751f\u6210\u53ef\u9084\u539f\u7684 unique hash signature\u3002<\/li>\n<li><strong>\u9069\u5408\u60c5\u5883<\/strong>\uff1a\u6bd4\u8f03\u6a19\u6e96 Transformer \u8207 MHF\u3001\u7814\u7a76 vocabulary expansion\u3001\u6e2c\u8a66\u56fa\u5b9a\u53c3\u6578\u91cf\u4e0b\u7684\u591a\u8a9e\u8a5e\u8868\u5ef6\u5c55\u3002<\/li>\n<li><strong>\u76f8\u95dc\u6a21\u578b<\/strong>\uff1aQwen3 \u6a19\u6e96\u7248\u3001qwen3_ori\u3001qwen3_hashformer\uff0c\u4ee5\u53ca 100M\uff0f1B\uff0f3B \u591a\u500b HuggingFace checkpoints\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u9ad4\u4f86\u770b\uff0c\u9019\u500b\u9805\u76ee\u7684\u50f9\u503c\u5728\u65bc\u5b83\u4e0d\u53ea\u63d0\u51fa\u4e00\u500b\u66f4\u7701\u53c3\u6578\u7684\u8868\u793a\u6cd5\uff0c\u9084\u8a66\u5716\u4fee\u88dc hash \u65b9\u6cd5\u9577\u671f\u7121\u6cd5\u81ea\u7136\u7528\u65bc\u751f\u6210\u5f0f\u6a21\u578b\u7684\u7f3a\u53e3\u3002\u5c0d\u7814\u7a76\u8a9e\u8a00\u6a21\u578b\u67b6\u69cb\u3001\u8a5e\u8868\u64f4\u5c55\u8207\u53c3\u6578\u6548\u7387\u7684\u5718\u968a\u4f86\u8aaa\uff0c\u5b83\u6bd4\u4e00\u822c\u300c\u63db\u500b tokenizer\u300d\u66f4\u503c\u5f97\u7d30\u770b\uff0c\u56e0\u70ba\u9023\u8f38\u5165\u8868\u793a\u8207\u4e0b\u4e00 token \u751f\u6210\u6a5f\u5236\u90fd\u4e00\u4f75\u6539\u5beb\u4e86\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/HUIYINXUE\/MHF\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2606.28057\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9019\u500b\u9805\u76ee\u628a\u8a9e\u8a00\u6a21\u578b\u7684\u8a5e\u8868\u554f\u984c\u6539\u5beb\u6210 hash signature \u751f\u6210\u3002\u91cd\u9ede\u4e0d\u53ea\u7701\u53c3\u6578\uff0c\u9084\u60f3\u8b93\u64f4\u5145\u8a5e\u5f59\u6642\u7dad\u6301\u56fa\u5b9a\u6a21\u578b\u9ad4\u7a4d\u3002<\/p>\n","protected":false},"author":8,"featured_media":9785,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[185,131,152,127,128],"tags":[],"class_list":["post-9786","post","type-post","status-publish","format-standard","hentry","category-qwen","category-embedding","category-python","category-127","category-128"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9786","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=9786"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9786\/revisions"}],"predecessor-version":[{"id":9788,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9786\/revisions\/9788"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9785"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9786"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9786"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9786"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}