{"id":5392,"date":"2024-09-02T06:45:58","date_gmt":"2024-09-01T21:45:58","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=5392"},"modified":"2024-09-02T06:45:58","modified_gmt":"2024-09-01T21:45:58","slug":"dolphin-remamba-the-mamba-in-the-llama-distilling-and-accelerating-hybrid-models","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=5392","title":{"rendered":"Dolphin, ReMamba, The Mamba in the Llama: Distilling and Accelerating Hybrid Models"},"content":{"rendered":"\n<p>\uff08\u4e3b\u3068\u3057\u3066RAG\u3092\u60f3\u5b9a\u3057\u305f\uff09\u9ad8\u901f\u5316\u3001Transformer\u306b\u4ee3\u308f\u308b\u304b\u3082\u3057\u308c\u306a\u3044Mamba\uff08<a href=\"https:\/\/devneko.jp\/wordpress\/?tag=mamba\">Mamba \u2013 arXiv\u6700\u65b0\u8ad6\u6587\u306e\u7d39\u4ecb (devneko.jp)<\/a>\uff09\u306e\u6539\u5584\u63d0\u6848\u3001\u65e2\u5b58LLM\u306eMamba\u3078\u306e\u5909\u63db\u306a\u3069\u8a08\u7b97\u52b9\u7387\u3092\u9ad8\u3081\u308b\u7814\u7a76\u306f\u591a\u304f\u884c\u308f\u308c\u3066\u3044\u308b\u3002<\/p>\n\n\n\n<p><a href=\"https:\/\/mistral.ai\/news\/codestral-mamba\/\">Codestral Mamba | Mistral AI | Frontier AI in your hands<\/a>\uff08<a href=\"https:\/\/huggingface.co\/mistralai\/Mamba-Codestral-7B-v0.1\">mistralai\/Mamba-Codestral-7B-v0.1 \u00b7 Hugging Face<\/a>\uff09\u306a\u3069\u5b9f\u7528\u6027\u306e\u3042\u308b\u30e2\u30c7\u30eb\u3082\u51fa\u3066\u304d\u3066\u3044\u3066\u8208\u5473\u6df1\u3044\u3002\u3053\u306e\u3042\u305f\u308a\u306e\u30d6\u30ec\u30a4\u30af\u30b9\u30eb\u30fc\u304c\u3042\u308b\u3068\u5206\u91ce\u5168\u4f53\u304c\u9762\u767d\u304f\u306a\u308b\u306e\u3067\u8981\u6ce8\u76ee\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Dolphin: Long Context as a New Modality for Energy-Efficient On-Device Language Models&nbsp;<\/strong>[9.9]<br>Dolphin\u306f\u3001\u8a00\u8a9e\u30e2\u30c7\u30eb\u306b\u304a\u3051\u308b\u9577\u3044\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u306e\u30a8\u30cd\u30eb\u30ae\u30fc\u52b9\u7387\u306e\u9ad8\u3044\u51e6\u7406\u306e\u305f\u3081\u306e\u65b0\u3057\u3044\u30c7\u30b3\u30fc\u30c0\u30c7\u30b3\u30fc\u30c0\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u3067\u3042\u308b\u3002 \u5f0a\u793e\u306e\u30a2\u30d7\u30ed\u30fc\u30c1\u306f\u3001\u30c7\u30d0\u30a4\u30b9\u4e0a\u306e\u30e2\u30c7\u30eb\u306b\u56fa\u6709\u306e\u3001\u30a8\u30cd\u30eb\u30ae\u30fc\u6d88\u8cbb\u3068\u30ec\u30a4\u30c6\u30f3\u30b7\u306e\u554f\u984c\u306b\u5bfe\u51e6\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2408.15518v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2408.15518v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Wed, 28 Aug 2024 04:06:14 GMT)<\/li>\n\n\n\n<li>\u300cBy treating extended context as a distinct modality, Dolphin utilizes a compact 0.5B parameter decoder to distill contextual information into memory tokens, which are then processed by a larger 7B parameter decoder.\u300d\u3068\u5c0f\u898f\u6a21\u306a\u30d1\u30e9\u30e1\u30fc\u30bf\u3068\u5927\u898f\u6a21\u306a\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u69cb\u9020\u3092\u7d44\u307f\u5408\u308f\u305b\u305f\u624b\u6cd5\u306e\u63d0\u6848\u3002\u5c0f\u898f\u6a21\u90e8\u5206\u3067\uff08RAG\u306b\u3042\u308a\u304c\u3061\u306a\uff09\u9577\u3044\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3092\u51e6\u7406\u3055\u305b\u308b\u3053\u3068\u3092\u60f3\u5b9a\u3057\u3066\u3044\u308b\u3002MLLM\u306eProjector\u3068\u8fd1\u3057\u3044\u69cb\u6210\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/huggingface.co\/NexaAIDev\/Dolphin\">NexaAIDev\/Dolphin \u00b7 Hugging Face<\/a><\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>ReMamba: Equip Mamba with Effective Long-Sequence Modeling&nbsp;<\/strong>[50.5]<br>\u672c\u7814\u7a76\u3067\u306f,\u9577\u3044\u6587\u8108\u306e\u7406\u89e3\u80fd\u529b\u3092\u9ad8\u3081\u308bReMamba\u3092\u63d0\u6848\u3059\u308b\u3002 ReMamba\u306f2\u6bb5\u968e\u306e\u30d7\u30ed\u30bb\u30b9\u3067\u9078\u629e\u7684\u5727\u7e2e\u3068\u9069\u5fdc\u306e\u30c6\u30af\u30cb\u30c3\u30af\u3092\u53d6\u308a\u5165\u308c\u3066\u3044\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2408.15496v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2408.15496v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Wed, 28 Aug 2024 02:47:27 GMT)<\/li>\n\n\n\n<li>Mamba\u306e\u6539\u5584\u3001\u9577\u6587\u306b\u304a\u3044\u3066\u8a18\u61b6\u3059\u3079\u304d\u60c5\u5831\u3092\u9078\u629e\u3057\u72b6\u614b\u7a7a\u9593\u306b\u4fdd\u5b58\u3059\u308b\u6a5f\u69cb\u3092\u8ffd\u52a0\u3001\u6027\u80fd\u3092\u5411\u4e0a\u3068\u306e\u3053\u3068\u3002Llama2\u3068\u304b\u306a\u308a\u826f\u3044\u52dd\u8ca0<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>The Mamba in the Llama: Distilling and Accelerating Hybrid Models&nbsp;<\/strong>[76.6]<br>\u6211\u3005\u306f,\u30a2\u30ab\u30c7\u30df\u30c3\u30afGPU\u30ea\u30bd\u30fc\u30b9\u3092\u7528\u3044\u3066\u30a2\u30c6\u30f3\u30b7\u30e7\u30f3\u5c64\u304b\u3089\u306e\u7dda\u5f62\u6295\u5f71\u91cd\u307f\u3092\u518d\u5229\u7528\u3059\u308b\u3053\u3068\u306b\u3088\u308a,\u5927\u304d\u306a\u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc\u3092\u7dda\u5f62RNN\u306b\u84b8\u7559\u3067\u304d\u308b\u3053\u3068\u3092\u5b9f\u8a3c\u3059\u308b\u3002 \u305d\u306e\u7d50\u679c\u3001\u6ce8\u610f\u5c64\u30924\u5206\u306e1\u542b\u3080\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u30e2\u30c7\u30eb\u306f\u3001\u30c1\u30e3\u30c3\u30c8\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u306e\u30aa\u30ea\u30b8\u30ca\u30eb\u306eTransformer\u306b\u5339\u6575\u3059\u308b\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u9054\u6210\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2408.15237v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2408.15237v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Tue, 27 Aug 2024 17:56:11 GMT)<\/li>\n\n\n\n<li>LLM\u3092\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9Mamba\u306b\u5909\u63db\u3001\u30bc\u30ed\u304b\u3089\u5b66\u7fd2\u3057\u305f\u3082\u306e\u3088\u308a\u9ad8\u6027\u80fd\u3060\u3063\u305f\u3068\u3059\u308b\u8ad6\u6587\u3002\u6b8b\u5ff5\u306a\u304c\u3089\u300cThe distilled pure (0%) model does degrade significantly in accuracy.\u300d<\/li>\n\n\n\n<li>\u666e\u901a\u306b\u5c0f\u578b\u5316\u3059\u308b\u306e\u3068\u6bd4\u3079\u3066\u512a\u4f4d\u6027\u304c\u3042\u308b\u306e\u3060\u308d\u3046\u304b\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\uff08\u4e3b\u3068\u3057\u3066RAG\u3092\u60f3\u5b9a\u3057\u305f\uff09\u9ad8\u901f\u5316\u3001Transformer\u306b\u4ee3\u308f\u308b\u304b\u3082\u3057\u308c\u306a\u3044Mamba\uff08Mamba \u2013 arXiv\u6700\u65b0\u8ad6\u6587\u306e\u7d39\u4ecb (devneko.jp)\uff09\u306e\u6539\u5584\u63d0\u6848\u3001\u65e2\u5b58LLM\u306eMamba\u3078\u306e\u5909\u63db\u306a\u3069\u8a08\u7b97\u52b9\u7387\u3092\u9ad8\u3081\u308b &hellip; <a href=\"https:\/\/devneko.jp\/wordpress\/?p=5392\" class=\"more-link\"><span class=\"screen-reader-text\">&#8220;Dolphin, ReMamba, The Mamba in the Llama: Distilling and Accelerating Hybrid Models&#8221; \u306e<\/span>\u7d9a\u304d\u3092\u8aad\u3080<\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[118,235],"class_list":["post-5392","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-dolphin","tag-mamba"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/5392","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5392"}],"version-history":[{"count":0,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/5392\/revisions"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5392"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5392"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5392"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}