{"id":4486,"date":"2024-08-15T20:22:01","date_gmt":"2024-08-15T12:22:01","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=4486"},"modified":"2024-08-15T20:22:23","modified_gmt":"2024-08-15T12:22:23","slug":"%e5%9c%a8%e4%b8%80%e5%bc%a0-24-gb-%e7%9a%84%e6%b6%88%e8%b4%b9%e7%ba%a7%e6%98%be%e5%8d%a1%e4%b8%8a%e7%94%a8-rlhf-%e5%be%ae%e8%b0%83-20b-llms","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=4486","title":{"rendered":"\u5728\u4e00\u5f20 24 GB \u7684\u6d88\u8d39\u7ea7\u663e\u5361\u4e0a\u7528 RLHF \u5fae\u8c03 20B LLMs"},"content":{"rendered":"\n

\u6211\u4eec\u5f88\u9ad8\u5174\u6b63\u5f0f\u53d1\u5e03 trl<\/code> \u4e0e peft<\/code> \u7684\u96c6\u6210\uff0c\u4f7f\u4efb\u4f55\u4eba\u90fd\u53ef\u4ee5\u66f4\u8f7b\u677e\u5730\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60\u8fdb\u884c\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u5fae\u8c03\uff01\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u8fd9\u662f\u73b0\u6709\u5fae\u8c03\u65b9\u6cd5\u7684\u6709\u7ade\u4e89\u529b\u7684\u66ff\u4ee3\u65b9\u6848\u3002<\/p>\n\n\n\n

\u8bf7\u6ce8\u610f\uff0c peft<\/code> \u662f\u4e00\u79cd\u901a\u7528\u5de5\u5177\uff0c\u53ef\u4ee5\u5e94\u7528\u4e8e\u8bb8\u591a ML \u7528\u4f8b\uff0c\u4f46\u5b83\u5bf9 RLHF \u7279\u522b\u6709\u8da3\uff0c\u56e0\u4e3a\u8fd9\u79cd\u65b9\u6cd5\u7279\u522b\u9700\u8981\u5185\u5b58\uff01<\/p>\n\n\n\n

\u5982\u679c\u4f60\u60f3\u76f4\u63a5\u6df1\u5165\u7814\u7a76\u4ee3\u7801\uff0c\u8bf7\u76f4\u63a5\u5728 TRL \u7684\u6587\u6863\u9875\u9762<\/a> \u76f4\u63a5\u67e5\u770b\u793a\u4f8b\u811a\u672c\u3002<\/p>\n\n\n\n

<\/a>\u4ecb\u7ecd<\/h2>\n\n\n\n

<\/a>LLMs & RLHF<\/h3>\n\n\n\n

LLM \u7ed3\u5408 RLHF (\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60) \u4f3c\u4e4e\u662f\u6784\u5efa\u975e\u5e38\u5f3a\u5927\u7684 AI \u7cfb\u7edf (\u4f8b\u5982 ChatGPT) \u7684\u4e0b\u4e00\u4e2a\u9996\u9009\u65b9\u6cd5\u3002<\/p>\n\n\n\n

\u4f7f\u7528 RLHF \u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u901a\u5e38\u5305\u62ec\u4ee5\u4e0b\u4e09\u4e2a\u6b65\u9aa4:<\/p>\n\n\n\n

    \n
  1. \u5728\u7279\u5b9a\u9886\u57df\u6216\u6307\u4ee4\u548c\u4eba\u7c7b\u793a\u8303\u8bed\u6599\u5e93\u4e0a\u5fae\u8c03\u9884\u8bad\u7ec3\u7684 LLM;<\/li>\n\n\n\n
  2. \u6536\u96c6\u4eba\u7c7b\u6807\u6ce8\u7684\u6570\u636e\u96c6\uff0c\u8bad\u7ec3\u4e00\u4e2a\u5956\u52b1\u6a21\u578b;<\/li>\n\n\n\n
  3. \u4f7f\u7528 RL (\u4f8b\u5982 PPO)\uff0c\u7528\u6b64\u6570\u636e\u96c6\u548c\u5956\u52b1\u6a21\u578b\u8fdb\u4e00\u6b65\u5fae\u8c03\u6b65\u9aa4 1 \u4e2d\u7684 LLM\u3002<\/li>\n<\/ol>\n\n\n\n
    \"\"<\/figure>\n\n\n\n

    ChatGPT \u7684\u8bad\u7ec3\u534f\u8bae\u6982\u8ff0\uff0c\u4ece\u6570\u636e\u6536\u96c6\u5230 RL \u90e8\u5206\u3002 \u8d44\u6599\u6765\u6e90:\u00a0OpenAI \u7684 ChatGPT \u535a\u6587<\/a><\/strong><\/p>\n\n\n\n

    \u57fa\u7840 LLM \u7684\u9009\u62e9\u5728\u8fd9\u91cc\u662f\u81f3\u5173\u91cd\u8981\u7684\u3002\u5728\u64b0\u5199\u672c\u6587\u65f6\uff0c\u53ef\u4ee5\u201c\u5f00\u7bb1\u5373\u7528\u201d\u5730\u7528\u4e8e\u8bb8\u591a\u4efb\u52a1\u7684\u201c\u6700\u4f73\u201d\u5f00\u6e90 LLM \u662f\u6307\u4ee4\u5fae\u8c03 LLMs\u3002\u8457\u540d\u7684\u6a21\u578b\u6709: BLOOMZ<\/a> Flan-T5<\/a>\u3001Flan-UL2<\/a> \u548c OPT-IML<\/a>\u3002\u8fd9\u4e9b\u6a21\u578b\u7684\u7f3a\u70b9\u662f\u5b83\u4eec\u7684\u5c3a\u5bf8\u3002\u8981\u83b7\u5f97\u4e00\u4e2a\u50cf\u6837\u7684\u6a21\u578b\uff0c\u4f60\u81f3\u5c11\u9700\u8981\u73a9 10B+ \u7ea7\u522b\u7684\u6a21\u578b\uff0c\u5728\u5168\u7cbe\u5ea6\u60c5\u51b5\u4e0b\u8fd9\u5c06\u9700\u8981\u9ad8\u8fbe 40GB GPU \u5185\u5b58\uff0c\u53ea\u662f\u4e3a\u4e86\u5c06\u6a21\u578b\u88c5\u5728\u5355\u4e2a GPU \u8bbe\u5907\u4e0a\u800c\u4e0d\u8fdb\u884c\u4efb\u4f55\u8bad\u7ec3\uff01<\/p>\n\n\n\n

    <\/a>\u4ec0\u4e48\u662f TRL\uff1f<\/h3>\n\n\n\n

    trl<\/code> \u5e93\u7684\u76ee\u7684\u662f\u4f7f RL \u7684\u6b65\u9aa4\u66f4\u5bb9\u6613\u548c\u7075\u6d3b\uff0c\u8ba9\u6bcf\u4e2a\u4eba\u53ef\u4ee5\u5728\u4ed6\u4eec\u81ea\u5df1\u7684\u6570\u636e\u96c6\u548c\u8bad\u7ec3\u8bbe\u7f6e\u4e0a\u7528 RL \u5fae\u8c03 LM\u3002\u5728\u8bb8\u591a\u5176\u4ed6\u5e94\u7528\u7a0b\u5e8f\u4e2d\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528\u6b64\u7b97\u6cd5\u5fae\u8c03\u6a21\u578b\u4ee5\u751f\u6210 \u6b63\u9762\u7535\u5f71\u8bc4\u8bba<\/a>\u3001\u8fdb\u884c \u53d7\u63a7\u751f\u6210<\/a> \u6216 \u964d\u4f4e\u6a21\u578b\u7684\u6bd2\u6027<\/a>\u3002<\/p>\n\n\n\n

    \u4f7f\u7528 trl<\/code> \u4f60\u53ef\u4ee5\u5728\u5206\u5e03\u5f0f\u7ba1\u7406\u5668\u6216\u8005\u5355\u4e2a\u8bbe\u5907\u4e0a\u8fd0\u884c\u6700\u53d7\u6b22\u8fce\u7684\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u4e4b\u4e00: PPO<\/a>\u3002\u6211\u4eec\u5229\u7528 Hugging Face \u751f\u6001\u7cfb\u7edf\u4e2d\u7684 accelerate<\/code> \u6765\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u8fd9\u6837\u4efb\u4f55\u7528\u6237\u90fd\u53ef\u4ee5\u5c06\u5b9e\u9a8c\u6269\u5927\u5230\u4e00\u4e2a\u6709\u8da3\u7684\u89c4\u6a21\u3002<\/p>\n\n\n\n

    \u4f7f\u7528 RL \u5fae\u8c03\u8bed\u8a00\u6a21\u578b\u5927\u81f4\u9075\u5faa\u4e0b\u9762\u8be6\u8ff0\u7684\u534f\u8bae\u3002\u8fd9\u9700\u8981\u6709 2 \u4e2a\u539f\u59cb\u6a21\u578b\u7684\u526f\u672c; \u4e3a\u907f\u514d\u6d3b\u8dc3\u6a21\u578b\u4e0e\u5176\u539f\u59cb\u884c\u4e3a\/\u5206\u5e03\u504f\u79bb\u592a\u591a\uff0c\u4f60\u9700\u8981\u5728\u6bcf\u4e2a\u4f18\u5316\u6b65\u9aa4\u4e2d\u8ba1\u7b97\u53c2\u8003\u6a21\u578b\u7684 logits \u3002\u8fd9\u5bf9\u4f18\u5316\u8fc7\u7a0b\u589e\u52a0\u4e86\u786c\u7ea6\u675f\uff0c\u56e0\u4e3a\u4f60\u59cb\u7ec8\u9700\u8981\u6bcf\u4e2a GPU \u8bbe\u5907\u81f3\u5c11\u6709\u4e24\u4e2a\u6a21\u578b\u526f\u672c\u3002\u5982\u679c\u6a21\u578b\u7684\u5c3a\u5bf8\u53d8\u5927\uff0c\u5728\u5355\u4e2a GPU \u4e0a\u5b89\u88c5\u8bbe\u7f6e\u4f1a\u53d8\u5f97\u8d8a\u6765\u8d8a\u68d8\u624b\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    TRL \u4e2d PPO \u8bad\u7ec3\u8bbe\u7f6e\u6982\u8ff0\u3002<\/strong><\/p>\n\n\n\n

    \u5728 trl<\/code> \u4e2d\uff0c\u4f60\u8fd8\u53ef\u4ee5\u5728\u53c2\u8003\u6a21\u578b\u548c\u6d3b\u8dc3\u6a21\u578b\u4e4b\u95f4\u4f7f\u7528\u5171\u4eab\u5c42\u4ee5\u907f\u514d\u6574\u4e2a\u526f\u672c\u3002 \u6a21\u578b\u89e3\u6bd2\u793a\u4f8b\u4e2d\u5c55\u793a\u4e86\u6b64\u529f\u80fd\u7684\u5177\u4f53\u793a\u4f8b\u3002<\/p>\n\n\n\n

    <\/a>\u5927\u89c4\u6a21\u8bad\u7ec3<\/h3>\n\n\n\n

    \u5927\u89c4\u6a21\u8bad\u7ec3\u662f\u5177\u6709\u6311\u6218\u6027\u7684\u3002\u7b2c\u4e00\u4e2a\u6311\u6218\u662f\u5728\u53ef\u7528\u7684 GPU \u8bbe\u5907\u4e0a\u62df\u5408\u6a21\u578b\uff0c\u53ca\u5176\u4f18\u5316\u5668\u72b6\u6001\u3002 \u5355\u4e2a\u53c2\u6570\u5360\u7528\u7684 GPU \u5185\u5b58\u91cf\u53d6\u51b3\u4e8e\u5176\u201c\u7cbe\u5ea6\u201d(\u6216\u66f4\u5177\u4f53\u5730\u8bf4\u662f dtype<\/code>)\u3002 \u6700\u5e38\u89c1\u7684 dtype<\/code> \u662f float32<\/code> (32 \u4f4d) \u3001 float16<\/code> \u548c bfloat16<\/code> (16 \u4f4d)\u3002 \u6700\u8fd1\uff0c\u201c\u5947\u5f02\u7684\u201d\u7cbe\u5ea6\u652f\u6301\u5f00\u7bb1\u5373\u7528\u7684\u8bad\u7ec3\u548c\u63a8\u7406 (\u5177\u6709\u7279\u5b9a\u6761\u4ef6\u548c\u7ea6\u675f)\uff0c\u4f8b\u5982 int8<\/code> (8 \u4f4d)\u3002 \u7b80\u800c\u8a00\u4e4b\uff0c\u8981\u5728 GPU \u8bbe\u5907\u4e0a\u52a0\u8f7d\u4e00\u4e2a\u6a21\u578b\uff0c\u6bcf\u5341\u4ebf\u4e2a\u53c2\u6570\u5728 float32 \u7cbe\u5ea6\u4e0a\u9700\u8981 4GB\uff0c\u5728 float16 \u4e0a\u9700\u8981 2GB\uff0c\u5728 int8 \u4e0a\u9700\u8981 1GB\u3002 \u5982\u679c\u4f60\u60f3\u4e86\u89e3\u5173\u4e8e\u8fd9\u4e2a\u8bdd\u9898\u7684\u66f4\u591a\u4fe1\u606f\uff0c\u8bf7\u67e5\u770b\u8fd9\u7bc7\u7814\u7a76\u6df1\u5165\u7684 \u6587\u7ae0<\/a>\u3002<\/p>\n\n\n\n

    \u5982\u679c\u60a8\u4f7f\u7528 AdamW \u4f18\u5316\u5668\uff0c\u6bcf\u4e2a\u53c2\u6570\u9700\u8981 8 \u4e2a\u5b57\u8282 (\u4f8b\u5982\uff0c\u5982\u679c\u60a8\u7684\u6a21\u578b\u6709 1B \u4e2a\u53c2\u6570\uff0c\u5219\u6a21\u578b\u7684\u5b8c\u6574 AdamW \u4f18\u5316\u5668\u5c06\u9700\u8981 8GB GPU \u5185\u5b58 \u6765\u6e90<\/a>)\u3002<\/p>\n\n\n\n

    \u8bb8\u591a\u6280\u672f\u5df2\u7ecf\u88ab\u91c7\u7528\u4ee5\u5e94\u5bf9\u5927\u89c4\u6a21\u8bad\u7ec3\u4e0a\u7684\u6311\u6218\u3002\u6700\u719f\u6089\u7684\u8303\u5f0f\u662f\u7ba1\u9053\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u548c\u6570\u636e\u5e76\u884c\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u56fe\u7247\u6765\u81ea\u00a0\u8fd9\u7bc7\u535a\u6587<\/a><\/strong><\/p>\n\n\n\n

    \u901a\u8fc7\u6570\u636e\u5e76\u884c\u6027\uff0c\u540c\u4e00\u6a21\u578b\u5e76\u884c\u6258\u7ba1\u5728\u591a\u53f0\u673a\u5668\u4e0a\uff0c\u5e76\u4e14\u6bcf\u4e2a\u5b9e\u4f8b\u90fd\u88ab\u63d0\u4f9b\u4e0d\u540c\u7684\u6570\u636e\u6279\u6b21\u3002 \u8fd9\u662f\u6700\u76f4\u63a5\u7684\u5e76\u884c\u7b56\u7565\uff0c\u672c\u8d28\u4e0a\u662f\u590d\u5236\u5355 GPU \u7684\u60c5\u51b5\uff0c\u5e76\u4e14\u5df2\u7ecf\u88ab trl<\/code> \u652f\u6301\u3002 \u4f7f\u7528\u7ba1\u9053\u5e76\u884c\u548c\u5f20\u91cf\u5e76\u884c\uff0c\u6a21\u578b\u672c\u8eab\u5206\u5e03\u5728\u673a\u5668\u4e0a: \u5728\u7ba1\u9053\u5e76\u884c\u4e2d\uff0c\u6a21\u578b\u6309\u5c42\u62c6\u5206\uff0c\u800c\u5f20\u91cf\u5e76\u884c\u5219\u8de8 GPU \u62c6\u5206\u5f20\u91cf\u64cd\u4f5c (\u4f8b\u5982\u77e9\u9635\u4e58\u6cd5)\u3002\u4f7f\u7528\u8fd9\u4e9b\u6a21\u578b\u5e76\u884c\u7b56\u7565\uff0c\u4f60\u9700\u8981\u5c06\u6a21\u578b\u6743\u91cd\u5206\u7247\u5230\u8bb8\u591a\u8bbe\u5907\u4e0a\uff0c\u8fd9\u9700\u8981\u4f60\u5b9a\u4e49\u8de8\u8fdb\u7a0b\u7684\u6fc0\u6d3b\u548c\u68af\u5ea6\u7684\u901a\u4fe1\u534f\u8bae\u3002 \u8fd9\u5b9e\u73b0\u8d77\u6765\u5e76\u4e0d\u7b80\u5355\uff0c\u53ef\u80fd\u9700\u8981\u91c7\u7528\u4e00\u4e9b\u6846\u67b6\uff0c\u4f8b\u5982 Megatron-DeepSpeed<\/code><\/a> \u6216 Nemo<\/code><\/a>\u3002\u5176\u4ed6\u5bf9\u6269\u5c55\u8bad\u7ec3\u81f3\u5173\u91cd\u8981\u7684\u5de5\u5177\u4e5f\u9700\u8981\u88ab\u5f3a\u8c03\uff0c\u4f8b\u5982\u81ea\u9002\u5e94\u6fc0\u6d3b\u68c0\u67e5\u70b9\u548c\u878d\u5408\u5185\u6838\u3002 \u53ef\u4ee5\u5728 \u6269\u5c55\u9605\u8bfb<\/a> \u627e\u5230\u6709\u5173\u5e76\u884c\u8303\u5f0f\u7684\u8fdb\u4e00\u6b65\u9605\u8bfb\u3002<\/p>\n\n\n\n

    \u56e0\u6b64\uff0c\u6211\u4eec\u95ee\u81ea\u5df1\u4e0b\u9762\u4e00\u4e2a\u95ee\u9898: \u4ec5\u7528\u6570\u636e\u5e76\u884c\u6211\u4eec\u53ef\u4ee5\u8d70\u591a\u8fdc\uff1f\u6211\u4eec\u80fd\u5426\u4f7f\u7528\u73b0\u6709\u7684\u5de5\u5177\u5728\u5355\u4e2a\u8bbe\u5907\u4e2d\u9002\u5e94\u8d85\u5927\u578b\u8bad\u7ec3\u8fc7\u7a0b (\u5305\u62ec\u6d3b\u8dc3\u6a21\u578b\u3001\u53c2\u8003\u6a21\u578b\u548c\u4f18\u5316\u5668\u72b6\u6001)\uff1f \u7b54\u6848\u4f3c\u4e4e\u662f\u80af\u5b9a\u7684\u3002 \u4e3b\u8981\u56e0\u7d20\u662f: \u9002\u914d\u5668\u548c 8 \u4f4d\u77e9\u9635\u4e58\u6cd5\uff01 \u8ba9\u6211\u4eec\u5728\u4ee5\u4e0b\u90e8\u5206\u4e2d\u4ecb\u7ecd\u8fd9\u4e9b\u4e3b\u9898:<\/p>\n\n\n\n

    <\/a>8 \u4f4d\u77e9\u9635\u4e58\u6cd5<\/h3>\n\n\n\n

    \u9ad8\u6548\u7684 8 \u4f4d\u77e9\u9635\u4e58\u6cd5\u662f\u8bba\u6587 LLM.int8() \u4e2d\u9996\u6b21\u5f15\u5165\u7684\u4e00\u79cd\u65b9\u6cd5\uff0c\u65e8\u5728\u89e3\u51b3\u91cf\u5316\u5927\u89c4\u6a21\u6a21\u578b\u65f6\u7684\u6027\u80fd\u4e0b\u964d\u95ee\u9898\u3002 \u6240\u63d0\u51fa\u7684\u65b9\u6cd5\u5c06\u5728\u7ebf\u6027\u5c42\u4e2d\u5e94\u7528\u7684\u77e9\u9635\u4e58\u6cd5\u5206\u89e3\u4e3a\u4e24\u4e2a\u9636\u6bb5: \u5728 float16 \u4e2d\u5c06\u88ab\u6267\u884c\u7684\u5f02\u5e38\u503c\u9690\u85cf\u72b6\u6001\u90e8\u5206\u548c\u5728 int8 \u4e2d\u88ab\u6267\u884c\u7684\u201c\u975e\u5f02\u5e38\u503c\u201d\u90e8\u5206\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u9ad8\u6548\u7684 8 \u4f4d\u77e9\u9635\u4e58\u6cd5\u662f\u8bba\u6587\u00a0LLM.int8()<\/a>\u00a0\u4e2d\u9996\u6b21\u5f15\u5165\u7684\u4e00\u79cd\u65b9\u6cd5\uff0c\u65e8\u5728\u89e3\u51b3\u91cf\u5316\u5927\u89c4\u6a21\u6a21\u578b\u65f6\u7684\u6027\u80fd\u4e0b\u964d\u95ee\u9898\u3002 \u6240\u63d0\u51fa\u7684\u65b9\u6cd5\u5c06\u5728\u7ebf\u6027\u5c42\u4e2d\u5e94\u7528\u7684\u77e9\u9635\u4e58\u6cd5\u5206\u89e3\u4e3a\u4e24\u4e2a\u9636\u6bb5: \u5728 float16 \u4e2d\u88ab\u6267\u884c\u7684\u5f02\u5e38\u503c\u9690\u85cf\u72b6\u6001\u90e8\u5206\u548c\u5728 int8 \u4e2d\u88ab\u6267\u884c\u7684\u201c\u975e\u5f02\u5e38\u503c\u201d\u90e8\u5206\u3002<\/strong><\/p>\n\n\n\n

    \u7b80\u800c\u8a00\u4e4b\uff0c\u5982\u679c\u4f7f\u7528 8 \u4f4d\u77e9\u9635\u4e58\u6cd5\uff0c\u5219\u53ef\u4ee5\u5c06\u5168\u7cbe\u5ea6\u6a21\u578b\u7684\u5927\u5c0f\u51cf\u5c0f\u5230 4 \u5206\u4e4b\u4e00 (\u56e0\u6b64\uff0c\u5bf9\u4e8e\u534a\u7cbe\u5ea6\u6a21\u578b\uff0c\u53ef\u4ee5\u51cf\u5c0f 2 \u5206\u4e4b\u4e00)\u3002<\/p>\n\n\n\n

    <\/a>\u4f4e\u79e9\u9002\u914d\u548c PEFT<\/h3>\n\n\n\n

    \u5728 2021 \u5e74\uff0c\u4e00\u7bc7\u53eb LoRA: Low-Rank Adaption of Large Language Models<\/strong> \u7684\u8bba\u6587\u8868\u660e\uff0c\u53ef\u4ee5\u901a\u8fc7\u51bb\u7ed3\u9884\u8bad\u7ec3\u6743\u91cd\uff0c\u5e76\u521b\u5efa\u67e5\u8be2\u548c\u503c\u5c42\u7684\u6ce8\u610f\u529b\u77e9\u9635\u7684\u4f4e\u79e9\u7248\u672c\u6765\u5bf9\u5927\u578b\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u4e9b\u4f4e\u79e9\u77e9\u9635\u7684\u53c2\u6570\u8fdc\u5c11\u4e8e\u539f\u59cb\u6a21\u578b\uff0c\u56e0\u6b64\u53ef\u4ee5\u4f7f\u7528\u66f4\u5c11\u7684 GPU \u5185\u5b58\u8fdb\u884c\u5fae\u8c03\u3002 \u4f5c\u8005\u8bc1\u660e\uff0c\u4f4e\u9636\u9002\u914d\u5668\u7684\u5fae\u8c03\u53d6\u5f97\u4e86\u4e0e\u5fae\u8c03\u5b8c\u6574\u9884\u8bad\u7ec3\u6a21\u578b\u76f8\u5f53\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u539f\u59cb (\u51bb\u7ed3\u7684) \u9884\u8bad\u7ec3\u6743\u91cd (\u5de6) \u7684\u8f93\u51fa\u6fc0\u6d3b\u7531\u4e00\u4e2a\u7531\u6743\u91cd\u77e9\u9635 A \u548c B \u7ec4\u6210\u7684\u4f4e\u79e9\u9002\u914d\u5668 (\u53f3) \u589e\u5f3a\u3002<\/strong><\/p>\n\n\n\n

    \u8fd9\u79cd\u6280\u672f\u5141\u8bb8\u4f7f\u7528\u4e00\u5c0f\u90e8\u5206\u5185\u5b58\u6765\u5fae\u8c03 LLM\u3002 \u7136\u800c\uff0c\u4e5f\u6709\u4e00\u4e9b\u7f3a\u70b9\u3002\u7531\u4e8e\u9002\u914d\u5668\u5c42\u4e2d\u7684\u989d\u5916\u77e9\u9635\u4e58\u6cd5\uff0c\u524d\u5411\u548c\u53cd\u5411\u4f20\u9012\u7684\u901f\u5ea6\u5927\u7ea6\u662f\u539f\u6765\u7684\u4e24\u500d\u3002<\/p>\n\n\n\n

    <\/a>\u4ec0\u4e48\u662f PEFT\uff1f<\/h3>\n\n\n\n

    Parameter-Efficient Fine-Tuning (PEFT)<\/a> \u662f\u4e00\u4e2a Hugging Face \u7684\u5e93\uff0c\u5b83\u88ab\u521b\u9020\u51fa\u6765\u4ee5\u652f\u6301\u5728 LLM \u4e0a\u521b\u5efa\u548c\u5fae\u8c03\u9002\u914d\u5668\u5c42\u3002 peft<\/code> \u4e0e \ud83e\udd17 Accelerate \u65e0\u7f1d\u96c6\u6210\uff0c\u7528\u4e8e\u5229\u7528\u4e86 DeepSpeed \u548c Big Model Inference \u7684\u5927\u89c4\u6a21\u6a21\u578b\u3002<\/p>\n\n\n\n

    \u6b64\u5e93\u652f\u6301\u5f88\u591a\u5148\u8fdb\u7684\u6a21\u578b\uff0c\u5e76\u4e14\u6709\u5927\u91cf\u7684\u4f8b\u5b50\uff0c\u5305\u62ec:<\/p>\n\n\n\n