{"id":6442,"date":"2025-10-22T13:56:01","date_gmt":"2025-10-22T05:56:01","guid":{"rendered":"https:\/\/nullthought.net\/?p=6442"},"modified":"2025-10-22T13:56:03","modified_gmt":"2025-10-22T05:56:03","slug":"bitdistill%ef%bc%9allm%e6%9d%83%e9%87%8d%e4%bb%8efp16%e9%87%8f%e5%8c%96%e5%88%b01-58-bit%ef%bc%8c%e7%b2%be%e5%ba%a6%e5%9f%ba%e6%9c%ac%e6%97%a0%e6%8d%9f%ef%bc%8c%e5%86%85%e5%ad%98%e6%b6%88%e8%80%97","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6442","title":{"rendered":"BitDistill\uff1aLLM\u6743\u91cd\u4eceFP16\u91cf\u5316\u52301.58-bit\uff0c\u7cbe\u5ea6\u57fa\u672c\u65e0\u635f\uff0c\u5185\u5b58\u6d88\u8017\u4e0b\u964d\u663e\u8457\uff0c\u63a8\u7406\u901f\u5ea6\u63d0\u5347"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2510.13998\" target=\"_blank\" rel=\"noreferrer noopener\">BitNet Distillation<\/a><\/strong>\u63d0\u51fa BitNet Distillation\uff08BitDistill\uff09\u6846\u67b6\uff0c\u5c06\u73b0\u6210\u5168\u7cbe\u5ea6\u5927\u6a21\u578b\uff08\u5982 Qwen \u7cfb\u5217\uff09\u4ee5\u6781\u4f4e\u6210\u672c\u5fae\u8c03\u4e3a <a href=\"https:\/\/nullthought.net\/?p=3533\" target=\"_blank\" rel=\"noreferrer noopener\">1.58-bit\uff08\u4e09\u503c {-1,0,1}\uff09\u6743\u91cd\u91cf\u5316\u6a21\u578b\uff0c<\/a>\u7528\u4e8e\u5177\u4f53\u4e0b\u6e38\u4efb\u52a1\uff0c\u540c\u65f6\u5728 CPU \u4e0a\u5b9e\u73b0\u7ea6 10\u00d7 \u5185\u5b58\u7f29\u51cf\u4e0e 2.65\u00d7 \u63a8\u7406\u52a0\u901f\uff0c\u5e76\u628a\u7cbe\u5ea6\u7ef4\u6301\u5728\u63a5\u8fd1 FP16 \u7684\u6c34\u5e73\u3002\u6838\u5fc3\u505a\u6cd5\u5305\u542b\u4e09\u9636\u6bb5\uff1a\u5728 Transformer \u5c42\u5185\u65b0\u589e SubLN \u7a33\u5b9a\u8bad\u7ec3\uff1b\u4f7f\u7528\u5c11\u91cf\u901a\u7528\u8bed\u6599\u505a\u7ee7\u7eed\u9884\u8bad\u7ec3\u4ee5\u6d88\u9664\u968f\u6a21\u578b\u89c4\u6a21\u6269\u5927\u7684\u6027\u80fd\u5dee\u8ddd\uff1b\u6700\u540e\u8fdb\u884c\u201clogits <a href=\"https:\/\/nullthought.net\/?p=5098\" target=\"_blank\" rel=\"noreferrer noopener\">\u84b8\u998f<\/a> + \u591a\u5934\u6ce8\u610f\u529b\u5173\u7cfb\u84b8\u998f\uff08MiniLM \u98ce\u683c\uff09\u201d\u3002\u5728\u5206\u7c7b\uff08MNLI\/QNLI\/SST-2\uff09\u4e0e\u6458\u8981\uff08CNN\/DailyMail\uff09\u4e0a\uff0cBitDistill \u7684 1.58-bit \u6a21\u578b\u4e0e\u540c\u89c4\u6a21 FP16 \u5fae\u8c03\u7ed3\u679c\u57fa\u672c\u6301\u5e73\u4f46\u663e\u8457\u66f4\u7701\u8d44\u6e90\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aXun Wu, Shaohan Huang, Wenhui Wang, Ting Song, Li Dong, Yan Xia, Furu Wei\uff0c\u6765\u81ea\u5fae\u8f6f\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"770\" height=\"495\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-15.png\" alt=\"\" class=\"wp-image-6443\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-15.png 770w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-15-300x193.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-15-768x494.png 768w\" sizes=\"auto, (max-width: 770px) 100vw, 770px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2510.13998\" target=\"_blank\" rel=\"noreferrer noopener\">BitNet Distillation<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u4e00\u3001\u7814\u7a76\u52a8\u673a\u4e0e\u95ee\u9898\u754c\u5b9a<br>\u6781\u4f4e\u6bd4\u7279\uff08\u5982 1.58-bit\uff09BitNet \u5df2\u8bc1\u660e\u80fd\u5927\u5e45\u964d\u4f4e\u5185\u5b58\u4e0e\u7b97\u529b\uff0c\u4f46\u4ee5\u5f80\u8981\u60f3\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\u6709\u597d\u6548\u679c\uff0c\u5e38\u9700\u201c\u4ece\u96f6\u9884\u8bad\u7ec3\u201d\u5de8\u91cf\u8bed\u6599\uff08\u4f8b\u5982\u6570\u4e07\u4ebf tokens\uff09\uff0c\u4ee3\u4ef7\u9ad8\u6602\u3002\u53e6\u4e00\u8def\u5f84\u662f\u76f4\u63a5\u5bf9\u73b0\u6210\u5168\u7cbe\u5ea6 LLM \u505a 1.58-bit \u7684 QAT \u518d\u5728\u4efb\u52a1\u4e0a\u5fae\u8c03\uff0c\u4f46\u4f5c\u8005\u89c2\u5bdf\u5230\uff1a\u8fd9\u79cd\u201c\u76f4\u8f6c\u76f4\u8bad\u201d\u7684\u7a33\u5b9a\u6027\u5dee\u3001\u4e0e FP16 \u7684\u7cbe\u5ea6\u5dee\u8ddd\u660e\u663e\u4e14\u201c\u8d8a\u5927\u6a21\u578b\u5dee\u8ddd\u8d8a\u5927\u201d\u7684\u53ef\u6269\u5c55\u6027\u95ee\u9898\u7a81\u51fa\u3002\u56e0\u6b64\u8bba\u6587\u76ee\u6807\u662f\uff1a\u4e0d\u4ece\u96f6\u8bad\u7ec3\uff0c\u53ea\u9760\u8f7b\u91cf\u6d41\u6c34\u7ebf\u628a\u5168\u7cbe\u5ea6 LLM \u53ef\u9760\u5730\u538b\u5230 1.58-bit\uff0c\u5e76\u5728\u4e0d\u540c\u6a21\u578b\u89c4\u6a21\u4e0e\u4efb\u52a1\u4e0a\u4fdd\u6301\u63a5\u8fd1 FP16 \u7684\u53ef\u6269\u5c55\u6027\u80fd\u3002<\/p>\n\n\n\n<p>\u4e8c\u3001\u65b9\u6cd5\u603b\u89c8\uff08BitDistill \u4e09\u9636\u6bb5\uff09<br>1\uff09\u5efa\u6a21\u6539\u9020\uff08Stage-1\uff09\uff1a\u5728\u6bcf\u4e2a Transformer block \u7684 MHSA \u8f93\u51fa\u6295\u5f71\u524d\u4e0e FFN \u8f93\u51fa\u6295\u5f71\u524d\u63d2\u5165 SubLN\uff0c\u4f7f\u91cf\u5316\u540e\u6fc0\u6d3b\u65b9\u5dee\u53d7\u63a7\u3001\u8bad\u7ec3\u66f4\u7a33\u5b9a\u3002<br>2\uff09\u7ee7\u7eed\u9884\u8bad\u7ec3\uff08Stage-2\uff09\uff1a\u7528\u4e00\u5c0f\u90e8\u5206\u901a\u7528\u8bed\u6599\uff08\u6587\u4e2d\u793a\u4f8b\u4e3a FALCON \u8bed\u6599\u62bd\u6837 10B tokens\uff09\u5bf9\u201c\u5df2\u63d2\u5165 SubLN \u5e76\u505a\u91cf\u5316\u201d\u7684\u6a21\u578b\u505a\u77ed\u6682\u8bed\u8a00\u5efa\u6a21\u8bad\u7ec3\uff0c\u5e2e\u52a9\u6743\u91cd\u5206\u5e03\u5411\u201c\u9002\u5408\u4e09\u503c\u5316\u201d\u7684\u5f62\u6001\u8fc1\u79fb\uff0c\u7f13\u89e3\u89c4\u6a21\u653e\u5927\u65f6\u7684\u7cbe\u5ea6\u9e3f\u6c9f\u3002<br>3\uff09\u84b8\u998f\u5f0f\u5fae\u8c03\uff08Stage-3\uff09\uff1a\u4ee5\u540c\u4efb\u52a1 FP16 \u6559\u5e08\u6307\u5bfc 1.58-bit \u5b66\u751f\uff0c\u8054\u5408\u4f7f\u7528 logits \u84b8\u998f\uff08KL\uff09\u4e0e\u6ce8\u610f\u529b\u5173\u7cfb\u84b8\u998f\uff08MiniLM \u601d\u8def\uff0c\u5bf9 Q\/K\/V \u7684\u5173\u7cfb\u77e9\u9635\u505a KL\uff09\uff0c\u914d\u5408\u4efb\u52a1\u4ea4\u53c9\u71b5\u5f62\u6210\u4e09\u9879\u635f\u5931\u7684\u52a0\u6743\u7ec4\u5408\u3002<\/p>\n\n\n\n<p>\u4e09\u3001\u91cf\u5316\u4e0e\u53cd\u4f20\u8fd1\u4f3c\u7684\u6280\u672f\u7ec6\u8282<br>\u6743\u91cd\u91cf\u5316\u91c7\u7528 per-tensor \u7684 absmean \u9608\u503c\u4e09\u503c\u5316\uff1a\u5148\u7528 \u0394=mean(|W|) \u5f52\u4e00\uff0c\u518d\u56db\u820d\u4e94\u5165\u5e76\u88c1\u526a\u5230 {-1,0,1}\uff1b\u6fc0\u6d3b\u91cf\u5316\u91c7\u7528 8-bit\u3001per-token \u7684 absmax\/absmean\uff1b\u7531\u4e8e RoundClip \u7b49\u64cd\u4f5c\u4e0d\u53ef\u5bfc\uff0c\u53cd\u5411\u4f20\u64ad\u7528\u76f4\u901a\u4f30\u8ba1\uff08STE\uff09\u3002\u8fd9\u4e9b\u9009\u62e9\u5728\u4fdd\u8bc1\u5b9e\u73b0\u7b80\u5355\u7684\u540c\u65f6\u517c\u987e\u4e86\u8bad\u7ec3\u53ef\u884c\u6027\u3002<\/p>\n\n\n\n<p>\u56db\u3001Stage-1\uff1aSubLN \u7684\u4f5c\u7528\u673a\u5236<br>\u4f4e\u6bd4\u7279\u6a21\u578b\u5e38\u89c1\u6fc0\u6d3b\u65b9\u5dee\u81a8\u80c0\u4e0e\u4f18\u5316\u4e0d\u7a33\u3002\u4f5c\u8005\u5728 MHSA \u4e0e FFN \u7684\u8f93\u51fa\u6295\u5f71\u524d\u5404\u52a0\u4e00\u5c42 SubLN\uff0c\u76f8\u5f53\u4e8e\u5728\u201c\u8fdb\u5165\u88ab\u91cf\u5316\u7684\u7ebf\u6027\u6295\u5f71\u4e4b\u524d\u201d\u5148\u628a\u8868\u5f81\u5c3a\u5ea6\u7a33\u4f4f\uff0c\u4ece\u800c\u51cf\u5c0f\u4e09\u503c\u5316\u5e26\u6765\u7684\u975e\u7ebf\u6027\u6270\u52a8\u5bf9\u6536\u655b\u7684\u7834\u574f\u3002\u5b9e\u9a8c\u8bc1\u660e\uff0c\u5f15\u5165 SubLN \u540e\u7684\u6536\u655b\u901f\u5ea6\u66f4\u5feb\u3001\u6700\u7ec8\u635f\u5931\u66f4\u4f4e\uff0c\u4e0b\u6e38\u6548\u679c\u66f4\u597d\u3002<\/p>\n\n\n\n<p>\u4e94\u3001Stage-2\uff1a\u4e3a\u4f55\u201c\u5c11\u91cf\u7ee7\u7eed\u9884\u8bad\u7ec3\u201d\u80fd\u89e3\u51b3\u53ef\u6269\u5c55\u6027\u96be\u9898<br>\u76f4\u63a5\u628a\u5168\u7cbe\u5ea6\u6743\u91cd\u538b\u5230 1.58-bit\uff0c\u6709\u9650\u7684\u4e0b\u6e38\u4efb\u52a1\u6570\u636e\u4e0d\u8db3\u4ee5\u628a\u6743\u91cd\u8c03\u6574\u5230\u201c\u91cf\u5316\u53cb\u597d\u201d\u7684\u5206\u5e03\uff0c\u89c4\u6a21\u8d8a\u5927\u8d8a\u96be\u5b66\u3002\u7ee7\u7eed\u9884\u8bad\u7ec3\u7528\u5c11\u91cf\u901a\u7528\u8bed\u6599\u8fdb\u884c\u8bed\u8a00\u5efa\u6a21\uff0c\u8ba9\u6743\u91cd\u5206\u5e03\u4ece\u63a5\u8fd1\u9ad8\u65af\uff0c\u8fc1\u79fb\u4e3a\u66f4\u50cf\u201c\u4ece\u96f6\u8bad\u7ec3 BitNet\u201d\u65f6\u51fa\u73b0\u7684\u96c6\u4e2d\u5206\u5e03\uff0c\u5c24\u5176\u5728 0\u2194\u00b11 \u7684\u9608\u503c\u9644\u8fd1\u805a\u96c6\u66f4\u591a\u6743\u91cd\uff0c\u4f7f\u91cf\u5316\u540e\u7684\u5c0f\u68af\u5ea6\u4e5f\u66f4\u5bb9\u6613\u5f15\u8d77\u7801\u5143\u7ffb\u8f6c\uff0c\u63d0\u5347\u53ef\u62df\u5408\u6027\u4e0e\u6536\u655b\u6027\u3002\u4f5c\u8005\u7684\u53ef\u89c6\u5316\u4e0e\u635f\u5931\u66f2\u7ebf\u5bf9\u8fd9\u4e00\u5047\u8bbe\u63d0\u4f9b\u4e86\u5b9e\u8bc1\u652f\u6301\u3002<\/p>\n\n\n\n<p>\u516d\u3001Stage-3\uff1a\u84b8\u998f\u8bbe\u8ba1\u4e0e\u635f\u5931\u51fd\u6570<br>1\uff09Logits \u84b8\u998f\uff1a\u5bf9\u6559\u5e08\u4e0e\u5b66\u751f\u7684\u6e29\u5ea6\u5316 softmax \u8f93\u51fa\u505a KL\uff08\u03c4=5\uff09\u3002<br>2\uff09\u6ce8\u610f\u529b\u5173\u7cfb\u84b8\u998f\uff1a\u6cbf MiniLM \u7cfb\u5217\uff0c\u53ea\u5728\u5355\u4e2a\u7cbe\u9009\u5c42\u8fdb\u884c\uff08\u800c\u975e\u5168\u5c42\uff09\uff0c\u5206\u522b\u5bf9 Q\/K\/V \u7684\u201c\u5f52\u4e00\u5316\u540e\u5173\u7cfb\u77e9\u9635\u201d\uff08scaled dot-product + softmax\uff09\u505a KL\uff1b\u5b9e\u9a8c\u8bc1\u660e\u201c\u4ec5\u4e00\u5c42\u4e14\u9760\u540e\u7684\u5c42\u201d\u66f4\u4f18\uff0c\u7ed9\u5b66\u751f\u7559\u51fa\u8db3\u591f\u7684\u81ea\u9002\u5e94\u81ea\u7531\u5ea6\u3002<br>3\uff09\u603b\u635f\u5931\u4e3a L = LCE + \u03bb\u00b7LLD + \u03b3\u00b7LAD\u3002\u5206\u7c7b\u4efb\u52a1\u5efa\u8bae \u03bb=10, \u03b3=1e5\uff1b\u6458\u8981\u4efb\u52a1 \u03bb=1, \u03b3=1e3\u3002<\/p>\n\n\n\n<p>\u4e03\u3001\u5b9e\u9a8c\u8bbe\u7f6e\u4e0e\u8bc4\u4ef7<br>\u57fa\u5ea7\u6a21\u578b\u4ee5 Qwen3\uff080.6B\/1.7B\/4B\uff09\u4e3a\u4e3b\uff0c\u5e76\u5728 Qwen2.5 \u4e0e Gemma \u4e0a\u9a8c\u8bc1\u666e\u9002\u6027\uff1b\u8bad\u7ec3\u5e8f\u5217\u957f 512\uff0cbatch 32\uff1b\u7ee7\u7eed\u9884\u8bad\u7ec3\u9636\u6bb5\u4f7f\u7528 10B tokens\uff08\u76f8\u5bf9\u4ece\u96f6\u8bad\u7ec3 BitNet \u9700\u7ea6 4T tokens \u7684\u6210\u672c\u51e0\u4e4e\u53ef\u5ffd\u7565\uff09\uff1b\u5fae\u8c03\u4e0e\u8bc4\u6d4b\u8986\u76d6 GLUE \u7684 MNLI\/QNLI\/SST-2 \u4e0e CNN\/DailyMail \u6458\u8981\uff0c\u751f\u6210\u8bc4\u6d4b\u7528 BLEU\/ROUGE\uff1b\u6548\u7387\u8bc4\u6d4b\u5728 16 \u7ebf\u7a0b CPU \u4e0a\u7edf\u8ba1 tokens\/s \u4e0e\u5185\u5b58\u5360\u7528\u3002<\/p>\n\n\n\n<p>\u516b\u3001\u4e3b\u8981\u7ed3\u679c\u4e0e\u6307\u6807\u89e3\u8bfb<br>1\uff09\u7cbe\u5ea6\uff1a\u5728\u4e09\u79cd\u89c4\u6a21\u4e0e\u4e24\u7c7b\u4efb\u52a1\u4e0a\uff0cBitDistill \u7684 1.58-bit \u5b66\u751f\u4e0e FP16 \u5fae\u8c03\u6559\u5e08\u7684\u5dee\u8ddd\u901a\u5e38\u5728\u7edf\u8ba1\u6ce2\u52a8\u8303\u56f4\u5185\uff08\u5982 MNLI\/QNLI\/SST-2 \u7684\u8868 1\uff0cCNNDM \u7684\u8868 2\uff09\uff0c\u5bf9\u6bd4\u201c\u76f4\u63a5 SFT \u7684 1.58-bit\u201d\u6709\u5927\u5e45\u63d0\u5347\uff0c\u4e14\u201c\u968f\u6a21\u578b\u53d8\u5927\u4ecd\u80fd\u8ddf\u4e0a FP16\u201d\u7684\u53ef\u6269\u5c55\u6027\u5f97\u4ee5\u4fdd\u6301\u3002<br>2\uff09\u6548\u7387\uff1aCPU \u63a8\u7406\u541e\u5410\u7ea6 2\u00d7 FP16\uff08\u793a\u4f8b\uff1a427 tokens\/s \u2192 1,135 tokens\/s\uff09\uff0c\u663e\u5b58\/\u5185\u5b58\u7ea6 10\u00d7 \u8282\u7701\uff081.20 GB \u2192 0.11 GB\uff09\u3002<br>3\uff09\u666e\u9002\u6027\uff1a\u66f4\u6362\u4e3a Gemma\u3001Qwen2.5 \u7b49\u57fa\u5ea7\u4ecd\u590d\u73b0\u201c\u63a5\u8fd1 FP16\u201d\u7684\u8d8b\u52bf\uff1b\u4e0e Block-Quant\u3001GPTQ\u3001AWQ \u7b49\u91cf\u5316\u6280\u672f\u7ec4\u5408\u4e5f\u80fd\u7a33\u5065\u589e\u76ca\u3002<\/p>\n\n\n\n<p>\u4e5d\u3001\u6d88\u878d\u4e0e\u8bca\u65ad\u6027\u5206\u6790<br>1\uff09\u9636\u6bb5\u6d88\u878d\uff1a\u53bb\u6389\u4efb\u4e00\u9636\u6bb5\uff08SubLN\u3001\u7ee7\u7eed\u9884\u8bad\u7ec3\u3001\u6ce8\u610f\u529b\/Logits \u84b8\u998f\uff09\u90fd\u4f1a\u663e\u8457\u6389\u70b9\uff0c\u4e09\u8005\u4e92\u8865\u3001\u5168\u94fe\u8def\u6700\u4f18\u3002<br>2\uff09\u84b8\u998f\u6210\u5206\uff1a\u5355\u7528 logits \u6216\u5355\u7528\u6ce8\u610f\u529b\u5173\u7cfb\u7686\u6709\u6548\uff0c\u4f46\u4e8c\u8005\u5e76\u7528\u6700\u7a33\u5065\u3002<br>3\uff09\u84b8\u998f\u5c42\u4f4d\uff1a\u5355\u5c42\u4f18\u4e8e\u5168\u5c42\uff0c\u4e14\u9760\u540e\u5c42\u6700\u4f73\u2014\u2014\u517c\u987e\u201c\u5bf9\u9f50\u80fd\u529b\u201d\u4e0e\u201c\u9002\u914d\u81ea\u7531\u5ea6\u201d\u3002<br>4\uff09\u6559\u5e08\u89c4\u6a21\uff1a\u66f4\u5f3a\u7684 FP16 \u6559\u5e08\u80fd\u628a\u540c\u89c4\u6a21 1.58-bit \u5b66\u751f\u8fdb\u4e00\u6b65\u201c\u5e26\u4e0a\u53bb\u201d\uff0c\u751a\u81f3\u5728\u5c11\u6570\u8bbe\u7f6e\u4e2d\u8d85\u8d8a\u540c\u5c3a\u5bf8 FP16\u3002<\/p>\n\n\n\n<p>\u5341\u3001\u5b9e\u8df5\u843d\u5730\u4e0e\u590d\u73b0\u5b9e\u64cd\u5efa\u8bae<br>1\uff09\u6a21\u5757\u6539\u9020\uff1a\u5728\u6bcf\u5c42 MHSA out-proj \u4e0e FFN down-proj \u524d\u63d2\u5165 SubLN\uff1b\u4fdd\u6301\u5176\u4f59\u7ed3\u6784\u4e0d\u53d8\uff0c\u4fbf\u4e8e\u590d\u7528\u6743\u91cd\u3002<br>2\uff09\u91cf\u5316\u7b56\u7565\uff1a\u6743\u91cd\u7528 per-tensor absmean \u4e09\u503c\u5316\uff0c\u6fc0\u6d3b\u7528 per-token 8-bit\uff08absmax\/absmean\uff09\uff0c\u53cd\u4f20\u7528 STE\u3002<br>3\uff09\u7ee7\u7eed\u9884\u8bad\u7ec3\uff1a\u62bd\u53d6 ~10B \u901a\u7528\u8bed\u6599\u505a LM \u8bad\u7ec3\uff0cepoch \u4e0d\u5fc5\u591a\uff0c\u89c2\u5bdf\u635f\u5931\u7a33\u5b9a\u4e0e\u6743\u91cd\u5206\u5e03\u7684\u201c\u9608\u503c\u9644\u8fd1\u805a\u96c6\u201d\u73b0\u8c61\u3002<br>4\uff09\u84b8\u998f\u5fae\u8c03\uff1a\u4e0e\u4efb\u52a1\u76d1\u7763\u4e00\u8d77\u8bad\u7ec3\uff1b\u9ed8\u8ba4 \u03c4=5\uff1b\u5206\u7c7b\u7528 \u03bb=10, \u03b3=1e5\uff0c\u6458\u8981\u7528 \u03bb=1, \u03b3=1e3\uff1b\u4ece\u9760\u540e\u67d0\u5c42\u6311\u4e00\u5c42\u505a\u6ce8\u610f\u529b\u5173\u7cfb\u84b8\u998f\uff0c\u5fc5\u8981\u65f6\u7f51\u683c\u641c\u7d22\u8be5\u5c42\uff1b\u82e5\u53ef\u7528\u66f4\u5f3a\u6559\u5e08\uff0c\u4f18\u5148\u7528\u5f3a\u6559\u5e08\u3002<br>5\uff09\u8bc4\u6d4b\u4e0e\u90e8\u7f72\uff1aCPU 16 \u7ebf\u7a0b\u4e0b\u53ef\u5f97 2\u00d7 \u541e\u5410\u300110\u00d7 \u5185\u5b58\u6536\u76ca\uff1b\u5bf9\u670d\u52a1\u4fa7\u63a8\u7406\u4e0e\u8fb9\u7f18\u8bbe\u5907\u843d\u5730\uff08\u5185\u5b58\/\u80fd\u8017\u654f\u611f\uff09\u5c24\u5176\u53cb\u597d\u3002<\/p>\n\n\n\n<p>\u5341\u4e00\u3001\u4e0e\u76f8\u5173\u5de5\u4f5c\u7684\u5173\u7cfb<br>BitDistill \u5e76\u975e\u4e0e GPTQ\/AWQ \u7b49 PTQ \u6216\u5176\u5b83 QAT \u4e92\u65a5\uff0c\u800c\u662f\u201c\u53ef\u53e0\u52a0\u201d\u7684\u8bad\u7ec3-\u84b8\u998f\u6d41\u7a0b\uff1b\u76f8\u8f83\u201c\u4ece\u96f6\u8bad\u7ec3 BitNet\u201d\u6781\u5927\u964d\u4f4e\u4e86\u9884\u8bad\u7ec3\u6210\u672c\uff1b\u76f8\u8f83\u4ec5\u505a QAT \u7684\u76f4\u63a5 SFT\uff0c\u89e3\u51b3\u4e86\u7a33\u5b9a\u6027\u5dee\u4e0e\u89c4\u6a21\u653e\u5927\u6389\u70b9\u7684\u95ee\u9898\uff1b\u76f8\u8f83\u4ec5\u505a logits \u84b8\u998f\u6216\u5c42\u5bf9\u9f50\u84b8\u998f\uff0c\u52a0\u5165\u201c\u6ce8\u610f\u529b\u5173\u7cfb\u84b8\u998f + \u5355\u5c42\u9009\u62e9\u201d\u662f\u5173\u952e\u7ec6\u8282\u3002<\/p>\n\n\n\n<p>\u5341\u4e8c\u3001\u5c40\u9650\u6027\u4e0e\u6f5c\u5728\u6539\u8fdb<br>1\uff09\u7ee7\u7eed\u9884\u8bad\u7ec3\u4ecd\u9700\u5c11\u91cf\u901a\u7528\u8bed\u6599\u4e0e\u82e5\u5e72\u7b97\u529b\uff0c\u5c3d\u7ba1\u8fdc\u4f4e\u4e8e\u4ece\u96f6\u8bad\u7ec3\uff1b2\uff09\u8bba\u6587\u4e3b\u8981\u9a8c\u8bc1\u5728\u4e2d\u5c0f\u89c4\u6a21\uff08\u22644B\uff09\u4e0e\u5178\u578b NLP \u4efb\u52a1\u4e0a\uff0c\u8de8\u6a21\u6001\u3001\u957f\u4e0a\u4e0b\u6587\u4e0e\u66f4\u5927\u89c4\u6a21\uff08\u22657B\/13B\uff09\u7684\u7cfb\u7edf\u6027\u8bc4\u4f30\u4ecd\u9700\u8865\u9f50\uff1b3\uff09\u5f53\u524d\u6ce8\u610f\u529b\u5173\u7cfb\u84b8\u998f\u7684\u5c42\u9009\u62e9\u4f9d\u8d56\u7ecf\u9a8c\/\u641c\u7d22\uff0c\u672a\u6765\u53ef\u63a2\u7d22\u5b66\u4e60\u5f0f\u6216\u81ea\u9002\u5e94\u7684\u5c42\u9009\u7b56\u7565\u3002<\/p>\n\n\n\n<p>\u5341\u4e09\u3001\u7ed3\u8bba\u4e0e\u542f\u793a<br>BitDistill \u7ed9\u51fa\u4e00\u6761\u201c\u8f7b\u91cf\u3001\u7a33\u5065\u3001\u53ef\u6269\u5c55\u201d\u7684\u9053\u8def\uff1a\u7528\u6700\u5c0f\u7684\u989d\u5916\u8bad\u7ec3\uff08SubLN+\u5c11\u91cf\u7ee7\u7eed\u9884\u8bad\u7ec3+\u53cc\u91cd\u84b8\u998f\uff09\u628a\u73b0\u6210 FP16 LLM \u538b\u5230 1.58-bit\uff0c\u800c\u4e14\u4e0d\u727a\u7272\u4efb\u52a1\u7cbe\u5ea6\uff0c\u8fd8\u80fd\u5728 CPU\/<a href=\"https:\/\/nullthought.net\/?p=3085\" target=\"_blank\" rel=\"noreferrer noopener\">\u8fb9\u7f18\u7aef<\/a>\u83b7\u5f97\u663e\u8457\u541e\u5410\u4e0e\u5185\u5b58\u4f18\u52bf\u3002\u5bf9\u4e8e\u8ffd\u6c42\u201c\u4efb\u52a1\u4e13\u7528 + \u7aef\u4fa7\u9ad8\u6548\u201d\u7684\u5e94\u7528\u56e2\u961f\uff0c\u8fd9\u662f\u5f53\u524d\u5de5\u7a0b\u4e0a\u6027\u4ef7\u6bd4\u6781\u9ad8\u7684\u65b9\u6848\u4e4b\u4e00\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>BitNet on GitHub: <a href=\"https:\/\/github.com\/microsoft\/BitNet\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/microsoft\/BitNet<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587BitNet Distillation\u63d0\u51fa BitNet Distillation\uff08BitDistill\uff09 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,95,86,79],"class_list":["post-6442","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-transformer","tag-llm","tag-microsoft"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587BitNet Distillation\u63d0\u51fa BitNet Distillation\uff08BitDistill\uff09&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6442","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6442"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6442\/revisions"}],"predecessor-version":[{"id":6444,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6442\/revisions\/6444"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6442"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6442"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6442"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}