{"id":10021548,"date":"2026-11-09T09:54:49","date_gmt":"2026-11-09T01:54:49","guid":{"rendered":"https:\/\/itmax.cn\/?p=10021548"},"modified":"2026-11-09T09:54:59","modified_gmt":"2026-11-09T01:54:59","slug":"distributed-llama-20261109","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10021548","title":{"rendered":"\u4e24\u53f0\u8bbe\u5907\u4e00\u8d77\u7b97\uff0c\u7b54\u6848\u4ece Paris \u53d8\u6210\u4e86 Rome\uff5cdistributed-llama \u6211\u7f16\u8bd1\u771f\u8dd1\u4e86\u4e00\u904d"},"content":{"rendered":"<h1 id=\"paris-romedistributed-llama\">\u4e24\u53f0\u8bbe\u5907\u4e00\u8d77\u7b97\uff0c\u7b54\u6848\u4ece Paris \u53d8\u6210\u4e86 Rome\uff5cdistributed-llama \u6211\u7f16\u8bd1\u771f\u8dd1\u4e86\u4e00\u904d<\/h1>\n<h2 id=\"_1\">\u4e00\u53f0\u8bbe\u5907\u7684\u5185\u5b58\u5899\uff0c\u6211\u60f3\u7ed5\u8fc7\u53bb<\/h2>\n<p>\u4e0a\u4e2a\u6708\u6709\u4e2a\u670b\u53cb\u627e\u6211\u51fa\u4e3b\u610f\u3002\u4ed6\u624b\u91cc\u6512\u4e86\u51e0\u53f0\u5403\u7070\u7684\u65e7\u8bbe\u5907\uff1a\u4e00\u53f0 M1 Mac mini\u3001\u4e24\u53f0\u5c0f\u4e3b\u673a\u3001\u8fd8\u6709\u4e00\u5757\u6811\u8393\u6d3e\uff0c\u60f3\u628a\u4e00\u4e2a 70B \u7ea7\u522b\u7684\u6a21\u578b\u642c\u56de\u5bb6\u81ea\u5df1\u8dd1\uff0c\u4e0d\u60f3\u6309\u6708\u7ed9\u4e91\u7aef\u4ea4\u94b1\u3002\u6211\u7ed9\u4ed6\u7b97\u4e86\u7b97\u8d26\uff0c\u5361\u70b9\u7279\u522b\u6e05\u695a\uff1a70B \u6a21\u578b\u5c31\u7b97\u538b\u5230 Q4 \u91cf\u5316\uff0c\u6743\u91cd\u4e5f\u5f97 40GB \u4e0a\u4e0b\uff0c\u4ed6\u4efb\u4f55\u4e00\u53f0\u8bbe\u5907\u7684\u5185\u5b58\u90fd\u88c5\u4e0d\u4e0b\u3002\u4ed6\u628a\u4e09\u53f0\u8bbe\u5907\u7684\u5185\u5b58\u52a0\u8d77\u6765\u95ee\u6211\u300c\u603b\u548c\u591f\u4e86\u5440\uff0c\u80fd\u4e0d\u80fd\u62fc\u4e00\u62fc\u300d\uff0c\u8fd9\u53e5\u8bdd\u6b63\u597d\u95ee\u5728\u4e86\u70b9\u4e0a\u2014\u2014\u5185\u5b58\u62fc\u5f97\u8d77\u6765\uff0c\u7b97\u529b\u62fc\u4e0d\u8d77\u6765\uff0c\u8fd9\u6b63\u662f\u5206\u5e03\u5f0f\u63a8\u7406\u6700\u96be\u5543\u7684\u90a3\u5757\u9aa8\u5934\u3002<\/p>\n<p>\u8001\u529e\u6cd5\u5c31\u4e24\u6761\u8def\u3002\u7b2c\u4e00\u6761\uff0c\u7838\u94b1\u6362\u4e00\u53f0\u5927\u5185\u5b58\u673a\u5668\uff0c64GB \u7684 Mac \u81f3\u5c11\u5c0f\u4e24\u4e07\uff0c\u4e3a\u4e86\u8dd1\u4e2a\u6a21\u578b\u592a\u5962\u4f88\u3002\u7b2c\u4e8c\u6761\uff0c\u7528 llama.cpp \u7684 RPC \u6a21\u5f0f\u628a\u6a21\u578b\u6309\u5c42\u62c6\u5230\u591a\u53f0\u8bbe\u5907\u4e0a\uff0c\u5404\u8282\u70b9\u8f6e\u6d41\u5e72\u6d3b\u2014\u2014\u8fd9\u4e2a\u601d\u8def\u80fd\u8dd1\uff0c\u4f46\u6309\u5c42\u63a5\u529b\u610f\u5473\u7740\u7b2c\u4e00\u4e2a\u8282\u70b9\u7b97\u5b8c\u7b2c\u4e00\u5c42\u624d\u8f6e\u5230\u7b2c\u4e8c\u4e2a\u8282\u70b9\uff0c\u7f51\u7edc\u6765\u56de\u6b21\u6570\u8ddf\u7740\u5c42\u6570\u8d70\uff0c\u901f\u5ea6\u611f\u4eba\u3002\u540e\u6765 Jeff Geerling \u5728\u4ed6\u90a3\u4e2a\u6811\u8393\u6d3e\u96c6\u7fa4\u9879\u76ee\u91cc\u5b9e\u6d4b\uff0cllama.cpp RPC \u8dd1 70B \u6bcf\u79d2\u53ea\u51fa 0.28 \u4e2a token\uff0c\u6162\u5230\u57fa\u672c\u6ca1\u6cd5\u770b\uff08\u51fa\u5904\u89c1\u540e\u6587\uff09\u3002<\/p>\n<p>\u6211\u5f53\u65f6\u8fd8\u6709\u4e2a\u60f3\u6cd5\uff1a\u7528\u63a8\u7406\u670d\u52a1\u5668\u628a\u51e0\u53f0\u8bbe\u5907\u5f53\u4e00\u4e2a\u6c60\u5b50\u8c03\u5ea6\uff0c\u8bf7\u6c42\u8f6e\u6d41\u5206\u53d1\u5230\u4e0d\u540c\u673a\u5668\u3002\u8fd9\u6761\u8def\u66f4\u4e0d\u884c\u2014\u2014\u5b83\u89e3\u51b3\u7684\u662f\u300c\u5f88\u591a\u4eba\u540c\u65f6\u7528\u300d\u7684\u95ee\u9898\uff0c\u89e3\u51b3\u4e0d\u4e86\u300c\u4e00\u4e2a\u6a21\u578b\u88c5\u4e0d\u4e0b\u300d\u7684\u95ee\u9898\uff0c\u5355\u4e2a\u8bf7\u6c42\u8fd8\u662f\u5f97\u5b8c\u6574\u585e\u8fdb\u4e00\u53f0\u8bbe\u5907\u7684\u5185\u5b58\u91cc\u3002<\/p>\n<p>\u5c31\u5728\u6211\u529d\u4ed6\u6b7b\u5fc3\u7684\u65f6\u5019\uff0c\u6211\u628a distributed-llama \u8fd9\u4e2a\u4ed3\u5e93\u7ffb\u4e86\u51fa\u6765\u3002\u5b83\u8ddf llama.cpp RPC \u8d70\u7684\u662f\u5b8c\u5168\u4e0d\u540c\u7684\u5207\u6cd5\uff1a\u4e0d\u662f\u6309\u5c42\u63a5\u529b\uff0c\u800c\u662f\u5f20\u91cf\u5e76\u884c\u2014\u2014\u6240\u6709\u8282\u70b9\u540c\u65f6\u53c2\u4e0e\u6bcf\u4e00\u6b65\u8ba1\u7b97\uff0c\u9760\u9ad8\u901f\u540c\u6b65\u6362\u901f\u5ea6\u3002\u4ed3\u5e93\u9996\u9875\u4e00\u53e5\u8bdd\u5f88\u6233\u4eba\uff1a\u300c\u628a\u5bb6\u91cc\u7684\u8bbe\u5907\u8fde\u6210\u96c6\u7fa4\u6765\u52a0\u901f\u63a8\u7406\uff0c\u8bbe\u5907\u8d8a\u591a\u8d8a\u5feb\u300d\u3002\u8fd9\u8bdd\u5230\u5e95\u6709\u51e0\u6210\u662f\u771f\u7684\uff0c\u6211\u51b3\u5b9a\u4eb2\u624b\u7f16\u4e00\u904d\u3001\u8dd1\u4e00\u904d\u518d\u8bf4\u3002<\/p>\n<h2 id=\"_2\">\u5b83\u600e\u4e48\u628a\u4e00\u4e2a\u6a21\u578b\u5207\u5230\u51e0\u53f0\u673a\u5668\u4e0a<\/h2>\n<p>\u5148\u628a\u67b6\u6784\u8bf4\u900f\uff0c\u56e0\u4e3a\u5b83\u7684\u4ee3\u4ef7\u5168\u85cf\u5728\u8fd9\u91cc\u3002<\/p>\n<p>distributed-llama \u628a\u7f51\u7edc\u91cc\u7684\u8bbe\u5907\u5206\u6210\u4e24\u79cd\u89d2\u8272\u3002\u6839\u8282\u70b9\uff08root\uff09\u8d1f\u8d23\u52a0\u8f7d\u6a21\u578b\u548c\u6743\u91cd\uff0c\u628a\u6743\u91cd\u5207\u7247\u5206\u53d1\u7ed9\u5404\u4e2a\u5de5\u4eba\u8282\u70b9\uff08worker\uff09\uff0c\u540c\u65f6\u7ef4\u62a4\u5168\u7f51\u72b6\u6001\u540c\u6b65\uff1b\u6bcf\u4e2a\u5de5\u4eba\u8282\u70b9\u2014\u2014\u5305\u62ec\u6839\u8282\u70b9\u81ea\u5df1\u2014\u2014\u90fd\u63e1\u6709\u795e\u7ecf\u7f51\u7edc\u7684\u4e00\u5757\u5207\u7247\uff0c\u6bcf\u4e00\u6b65\u63a8\u7406\u5927\u5bb6\u5404\u7b97\u5404\u7684\u90a3\u4e00\u4efd\uff0c\u7b97\u5b8c\u518d\u540c\u6b65\u3002\u6839\u8282\u70b9\u7684\u5185\u5b58\u5360\u7528\u4f1a\u6bd4\u5de5\u4eba\u591a\u4e00\u70b9\uff0c\u5de5\u4eba\u8282\u70b9\u4e0d\u9700\u8981\u4efb\u4f55\u6a21\u578b\u914d\u7f6e\uff0c\u62ff\u5230 IP \u5c31\u80fd\u4e0a\u5c97\u3002<\/p>\n<p>\u8fd9\u4e2a\u8bbe\u8ba1\u7684\u6536\u76ca\u548c\u4ee3\u4ef7\u662f\u7ed1\u6b7b\u7684\u3002\u6536\u76ca\uff1a\u6a21\u578b\u5185\u5b58\u6309\u8282\u70b9\u6570\u644a\u8584\uff0c\u4e00\u53f0\u88c5\u4e0d\u4e0b\u7684\u6a21\u578b\uff0c\u56db\u53f0\u51d1\u4e00\u51d1\u5c31\u88c5\u4e0b\u4e86\u3002\u4ee3\u4ef7\uff1a<strong>\u6bcf\u751f\u6210\u4e00\u4e2a token\uff0c\u6240\u6709\u8282\u70b9\u90fd\u8981\u540c\u6b65\u4e00\u8f6e<\/strong>\u3002\u6211\u5728\u4e24\u8282\u70b9\u5b9e\u6d4b\u91cc\u76ef\u7740\u65e5\u5fd7\u7b97\u8fc7\uff0c\u6bcf\u4e2a token \u8981\u53d1 63kB\u3001\u6536 356kB \u6570\u636e\uff0c\u540c\u6b65\u8017\u65f6 2 \u5230 8 \u6beb\u79d2\u2014\u2014\u8fd9\u8fd8\u662f\u672c\u673a loopback \u7684\u6210\u7ee9\uff0c\u6362\u6210\u5343\u5146\u7f51\u7ebf\u53ea\u4f1a\u66f4\u7cdf\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5b83\u7684\u5929\u82b1\u677f\u5927\u6982\u7387\u4e0d\u662f CPU \u7b97\u529b\uff0c\u800c\u662f\u7f51\u7ebf\u3002<\/p>\n<p>\u8ddf llama.cpp RPC \u6446\u5728\u4e00\u8d77\u770b\u4f1a\u66f4\u6e05\u695a\u3002RPC \u662f\u6d41\u6c34\u7ebf\uff1a\u6bcf\u4e2a\u8282\u70b9\u625b\u82e5\u5e72\u5c42\uff0c\u6570\u636e\u50cf\u63a5\u529b\u68d2\u4e00\u6837\u4e00\u68d2\u4e00\u68d2\u4f20\uff0c\u7f51\u7edc\u6d41\u91cf\u8ddf\u6a21\u578b\u5c42\u6570\u6210\u6b63\u6bd4\uff0c\u67d0\u4e2a\u6162\u8282\u70b9\u4f1a\u6210\u4e3a\u5168\u94fe\u8def\u7684\u8282\u62cd\u5668\u3002distributed-llama \u662f\u5408\u5531\uff1a\u6bcf\u5c42\u5185\u90e8\u5207\u5f00\u5927\u5bb6\u4e00\u8d77\u7b97\uff0c\u7b97\u5b8c\u4e00\u6b21\u5168\u5f52\u7ea6\uff0c\u540c\u6b65\u6b21\u6570\u8ddf\u5c42\u6570\u65e0\u5173\u3001\u53ea\u8ddf\u751f\u6210\u6b65\u6570\u6210\u6b63\u6bd4\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u5b83\u5728\u540c\u6784\u3001\u4f4e\u5ef6\u8fdf\u3001\u6709\u7ebf\u4e92\u8054\u7684\u573a\u666f\u4e0b\u80fd\u8dd1\u51fa RPC \u62ff\u4e0d\u5230\u7684\u901f\u5ea6\uff0c\u4e5f\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u5230\u4e86 Wi-Fi \u6216\u8005\u914d\u7f6e\u4e0d\u9f50\u7684\u8bbe\u5907\u4e0a\u5b83\u5c31\u6c34\u571f\u4e0d\u670d\u2014\u2014\u5408\u5531\u961f\u91cc\u53ea\u8981\u4e00\u4e2a\u4eba\u6162\u534a\u62cd\uff0c\u6240\u6709\u4eba\u90fd\u5f97\u7b49\u4ed6\u3002<\/p>\n<p>\u8fd8\u6709\u4e09\u6761\u786c\u9650\u5236\uff0c\u52a8\u5de5\u4e4b\u524d\u5fc5\u987b\u5148\u80cc\u4e0b\u6765\uff1a<\/p>\n<ol>\n<li><strong>\u8282\u70b9\u6570\u53ea\u80fd\u662f 2 \u7684\u5e42<\/strong>\uff1a1\u30012\u30014\u30018\u3002\u4f60\u6709\u4e09\u53f0\u8bbe\u5907\uff0c\u5c31\u53ea\u80fd\u7528\u4e24\u53f0\uff0c\u5269\u4e0b\u4e00\u53f0\u5e72\u770b\u7740\u3002\u8fd9\u4e2a\u7ea6\u675f\u6765\u81ea\u5f20\u91cf\u5e76\u884c\u7684\u5207\u5206\u7b97\u6cd5\uff0c\u4e0d\u662f\u62cd\u8111\u888b\u5b9a\u7684\uff0c\u4f46\u5bf9\u4f60\u5bb6\u8bbe\u5907\u6e05\u5355\u662f\u5b9e\u6253\u5b9e\u7684\u7ea6\u675f\u3002<\/li>\n<li><strong>\u8282\u70b9\u6570\u4e0a\u9650\u7b49\u4e8e\u6a21\u578b\u7684 KV head \u6570\u91cf<\/strong>\uff08issue <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/70\">#70<\/a>\uff09\u3002\u5927\u591a\u6570\u7a20\u5bc6\u6a21\u578b KV head \u591f\u591a\uff0c\u95ee\u9898\u4e0d\u5927\uff1b\u4e2a\u522b MoE \u6a21\u578b\u8fd9\u4e2a\u6570\u4f1a\u5f88\u4f4e\uff0c\u5929\u82b1\u677f\u4e00\u4e0b\u5c31\u538b\u4e0b\u6765\u4e86\u3002<\/li>\n<li><strong>\u91cf\u5316\u53ea\u652f\u6301\u4e24\u79cd\u7ec4\u5408<\/strong>\uff08issue <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/183\">#183<\/a>\uff09\uff1aq40 \u6743\u91cd\u914d q80 \u540c\u6b65\u7f13\u51b2\uff0c\u6216\u8005\u5168\u7a0b f32\u3002GGUF \u90a3\u4e00\u5957\u5404\u79cd\u91cf\u5316\u5728\u8fd9\u513f\u901a\u901a\u4e0d\u8ba4\u3002q80 \u7f13\u51b2\u662f\u5b83\u7684\u6298\u4e2d\u672f\uff1a\u6743\u91cd\u538b\u5230 4 \u6bd4\u7279\u7701\u5185\u5b58\uff0c\u540c\u6b65\u65f6\u5347\u5230 8 \u6bd4\u7279\u4fdd\u7cbe\u5ea6\uff0c\u4e24\u8fb9\u5404\u8ba9\u4e00\u6b65\u3002<\/li>\n<\/ol>\n<p>\u53e6\u5916\u6709\u4e2a\u5b9e\u7528\u53c2\u6570\u503c\u5f97\u8bb0\u4f4f\uff1a<code>--max-seq-len<\/code>\u3002\u628a\u6700\u5927\u5e8f\u5217\u957f\u5ea6\u4ece\u9ed8\u8ba4\u503c\u538b\u5230\u4f60\u771f\u6b63\u9700\u8981\u7684\u6570\uff08\u6bd4\u5982 2048\uff09\uff0cKV cache \u7684\u5185\u5b58\u8ddf\u7740\u7f29\uff0c\u5bf9\u5185\u5b58\u521a\u597d\u5361\u7ebf\u7684\u8282\u70b9\u662f\u6551\u547d\u7a3b\u8349\u3002<\/p>\n<p>\u6700\u540e\u4e00\u6761\u5bb9\u6613\u88ab\u5ffd\u7565\uff1a\u5b83<strong>\u4e0d\u80fd\u76f4\u63a5\u5403 GGUF<\/strong>\u3002\u6a21\u578b\u5f97\u8f6c\u6362\u6210\u81ea\u5bb6 <code>.m<\/code> \u683c\u5f0f\uff0c\u4f5c\u8005\u5728 Hugging Face \u4e0a\u653e\u4e86\u9884\u8f6c\u6362\u597d\u7684\u6a21\u578b\u4ed3\u5e93\uff0c\u6216\u8005\u81ea\u5df1\u7528\u4ed3\u5e93\u91cc\u7684 <code>convert-hf.py<\/code> \u8f6c\u2014\u2014\u5927\u6a21\u578b\u8f6c\u6362\u8981\u51e0\u5c0f\u65f6\u5916\u52a0\u53cc\u4efd\u78c1\u76d8\u7a7a\u95f4\uff0c405B \u6a21\u578b\u5149\u4e0b\u8f7d\u5c31\u8981 238GB\uff0c\u8f6c\u6362\u671f\u95f4\u4f60\u5f97\u5907\u51fa\u7ea6 1.1TB \u7684\u7a7a\u5730\u3002<\/p>\n<h2 id=\"_3\">\u54ea\u4e9b\u662f\u771f\u672c\u4e8b\uff0c\u54ea\u4e9b\u8fd8\u6b20\u7740\u706b\u5019<\/h2>\n<p>\u6211\u628a\u5b83\u7684\u80fd\u529b\u6309\u300c\u4fe1\u4e0d\u4fe1\u5f97\u8fc7\u300d\u5206\u4e86\u4e09\u6863\u3002<\/p>\n<p><strong>\u4fe1\u5f97\u8fc7\u7684<\/strong>\uff1aCPU \u8def\u5f84\u3002ARM \u548c x86 AVX2 \u90fd\u505a\u4e86\u4f18\u5316\uff0cLinux\u3001macOS\u3001Windows \u5168\u652f\u6301\uff0c\u7f16\u8bd1\u5c31\u4e00\u4e2a <code>make<\/code>\u3002Windows \u8d70 Chocolatey \u88c5 mingw \u5c31\u80fd\u7f16\uff0c\u6811\u8393\u6d3e\u6587\u6863\u4e13\u95e8\u5199\u4e86\u8981\u88c5 64 \u4f4d\u7684 Lite \u7cfb\u7edf\uff08\u4e0d\u5e26\u684c\u9762\u7684\u90a3\u4e2a\uff09\u3002\u6a21\u578b\u4e0b\u8f7d\u6709 <code>launch.py<\/code> \u4e00\u6761\u547d\u4ee4\uff0c\u9884\u8f6c\u6362\u6a21\u578b\u5217\u8868\u8986\u76d6 Llama 3.1\/3.2\/3.3\uff08\u6700\u5927\u5230 405B\uff09\u3001DeepSeek R1 \u84b8\u998f\u7248\u3001Qwen3 \u5168\u7cfb\uff080.6B \u5230 30B MoE\uff09\uff0c\u4e24\u5929\u524d\u521a\u53d1\u7684 v0.17.0 \u53c8\u52a0\u4e86 Qwen3.5 9B\u2014\u2014\u7f16\u8bd1\u65f6\u6211\u4eb2\u773c\u770b\u5230 <code>qwen35-graph.cpp<\/code> \u51fa\u73b0\u5728\u7f16\u8bd1\u8f93\u51fa\u91cc\uff0c\u65b0\u529f\u80fd\u662f\u771f\u7684\u8fdb\u4e86\u4e3b\u4ee3\u7801\u3002API \u670d\u52a1 <code>dllama-api<\/code> \u63d0\u4f9b OpenAI \u517c\u5bb9\u7aef\u70b9\uff0c<code>\/v1\/models<\/code> \u548c <code>\/v1\/chat\/completions<\/code> \u6211\u90fd\u62ff curl \u6253\u8fc7\uff0c\u80fd\u6b63\u5e38\u8fd4\u56de\u3002\u4ed3\u5e93\u7684 discussions \u533a\u8fd8\u662f\u4e2a\u5b9d\u85cf\uff0c\u4f5c\u8005\u628a\u5404\u5730\u7528\u6237\u63d0\u4ea4\u7684\u5b9e\u6d4b\u914d\u7f6e\u90fd\u6c47\u603b\u5728\u90a3\u513f\uff0c\u6211\u540e\u9762\u5f15\u7528\u7684 Mac Mini \u6570\u636e\u5c31\u6765\u81ea\u5b98\u65b9\u7f6e\u9876\u7684\u5e16\u5b50\u3002<\/p>\n<p><strong>\u8981\u6253\u6298\u6263\u7684<\/strong>\uff1aGPU \u8def\u5f84\u3002Vulkan \u540e\u7aef\u4ece 2025 \u5e74 3 \u6708\u7684 v0.13.0 \u8fdb\u6765\uff0c\u81f3\u4eca\u6807\u7740\u5b9e\u9a8c\u6027\uff0cissue <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/50\">#50<\/a>\u300cVulkan Acceleration\u300d\u6302\u7740 35 \u6761\u8bc4\u8bba\u5feb\u4e24\u5e74\u6ca1\u6536\u53e3\u3002\u66f4\u624e\u5fc3\u7684\u662f issue <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/265\">#265<\/a>\uff0c\u6709\u4eba\u5b9e\u6d4b Qwen 8B \u4e0a GPU \u53cd\u800c\u6bd4 CPU \u6162\uff0c\u6807\u9898\u5c31\u53eb\u300cUnexpected Performance\u300d\uff1bissue <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/272\">#272<\/a> \u91cc AMD GPU \u8dd1 70B \u5e72\u8106\u6ca1\u8f93\u51fa\u3002\u6211\u7684\u5224\u65ad\uff1aGPU \u52a0\u901f\u8fd9\u6761\u8def\u73b0\u5728\u53ea\u80fd\u7b97\u300c\u5728\u94fa\u300d\uff0c\u60f3\u7528 GPU \u8282\u70b9\u7ec4\u96c6\u7fa4\u7684\uff0c\u5148\u628a\u8fd9\u4e24\u4e2a issue \u770b\u5b8c\u518d\u51b3\u5b9a\u3002<\/p>\n<p><strong>\u8fd8\u6ca1\u6709\u7684<\/strong>\uff1aembedding \u63a5\u53e3\uff08feature \u8bf7\u6c42 <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/96\">#96<\/a> \u6302\u7740\uff09\u3001\u624b\u673a\u8282\u70b9\uff08<a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/89\">#89<\/a> \u6709\u4eba\u5728\u95ee\uff0c\u6ca1\u4e0b\u6587\uff09\u3002\u60f3\u8981\u7684\u8bdd\u73b0\u5728\u53ea\u80fd\u7b49\u3002<\/p>\n<h2 id=\"_4\">\u7f16\u8bd1\u5b89\u88c5\u6bd4\u6211\u60f3\u7684\u987a\u5229\uff0c\u5751\u5168\u5728\u6a21\u578b\u90a3\u8fb9<\/h2>\n<p>\u6211\u5728 M2 \u7684 MacBook \u4e0a\u5b8c\u6574\u8d70\u4e86\u4e00\u904d\uff0c\u5168\u7a0b\u987a\u5229\u5f97\u6709\u70b9\u610f\u5916\u3002<\/p>\n<pre><code class=\"language-bash\">git clone https:\/\/github.com\/b4rtaz\/distributed-llama.git\ncd distributed-llama\nmake dllama\nmake dllama-api\n<\/code><\/pre>\n<p><code>make<\/code> \u7528 g++ <code>-O3<\/code> \u7f16\u8bd1\uff0c\u6211\u8fd9\u53f0\u673a\u5668 30 \u79d2\u5185\u7f16\u5b8c\uff0c\u4ea7\u7269\u662f\u4e24\u4e2a\u4e8c\u8fdb\u5236\uff1a<code>dllama<\/code> \u548c <code>dllama-api<\/code>\uff0c\u52a0\u8d77\u6765\u4e0d\u5230 1MB\u3002\u6ca1\u6709 Python \u4f9d\u8d56\uff0c\u6ca1\u6709\u5de8\u578b\u8fd0\u884c\u65f6\uff0c\u8fd9\u4e00\u70b9\u6bd4\u5f88\u591a\u540c\u7c7b\u9879\u76ee\u4f53\u9762\u5f97\u591a\u3002<\/p>\n<p>\u6a21\u578b\u4e0b\u8f7d\u7528\u4ed3\u5e93\u81ea\u5e26\u7684 <code>launch.py<\/code>\uff08\u9700\u8981 Python 3\uff09\uff1a<\/p>\n<pre><code class=\"language-bash\">python3 launch.py                # \u6253\u5370\u53ef\u7528\u6a21\u578b\u5217\u8868\npython3 launch.py qwen3_0.6b_q40 # \u4e0b\u8f7d 0.9GB \u7684 Qwen3 0.6B\n<\/code><\/pre>\n<p>\u4e0b\u8f7d\u5b8c\u5b83\u4f1a\u987a\u624b\u751f\u6210\u4e00\u4e2a <code>run_qwen3_0.6b_q40.sh<\/code>\uff0c\u547d\u4ee4\u53c2\u6570\u90fd\u66ff\u4f60\u62fc\u597d\u4e86\u3002\u591a\u8282\u70b9\u90e8\u7f72\u65f6\uff0c\u6a21\u578b\u53ea\u9700\u8981\u4e0b\u8f7d\u5230\u6839\u8282\u70b9\uff0c\u5de5\u4eba\u8282\u70b9\u7a7a\u624b\u8d77\u6b65\u2014\u2014\u8fd9\u4e2a\u7ec6\u8282\u5bf9\u6811\u8393\u6d3e\u7528\u6237\u5f88\u53cb\u597d\uff0c4GB \u7684\u5361\u4e0d\u7528\u817e\u5730\u65b9\u653e\u6743\u91cd\u3002<\/p>\n<p>\u591a\u8282\u70b9\u7684\u7ec4\u7f51\u6d41\u7a0b\u5b98\u65b9\u6587\u6863\u5199\u5f97\u7ec6\uff0c\u6811\u8393\u6d3e\u90a3\u7bc7\u7684\u5173\u952e\u51e0\u6b65\u662f\uff1a\u6bcf\u53f0\u8bbe\u5907\u88c5 64 \u4f4d\u7cfb\u7edf\u3001\u63a5\u7f51\u7ebf\u5230\u540c\u4e00\u4e2a\u4ea4\u6362\u673a\uff08\u4e24\u53f0\u8bbe\u5907\u53ef\u4ee5\u76f4\u8fde\u7701\u6389\u4ea4\u6362\u673a\uff09\u3001\u7ed9\u6bcf\u53f0\u8bbe\u5907\u914d\u9759\u6001 IP\uff08<code>sudo ip addr add 10.0.0.2\/24 dev eth0<\/code> \u8fd9\u79cd\u4e00\u884c\u547d\u4ee4\uff09\u3001\u6240\u6709\u8bbe\u5907\u7f16\u8bd1\u540c\u4e00\u4efd\u4ee3\u7801\uff0c\u7136\u540e\u5de5\u4eba\u5728\u5404\u81ea\u673a\u5668\u4e0a\u8d77\u670d\u52a1\uff0c\u6839\u8282\u70b9\u542f\u52a8\u65f6\u628a\u5de5\u4eba\u5730\u5740\u7528\u7a7a\u683c\u4e32\u5728 <code>--workers<\/code> \u53c2\u6570\u540e\u9762\u3002\u5de5\u4eba\u8282\u70b9\u90a3\u6761\u547d\u4ee4\u5b98\u65b9\u5199\u7684\u662f <code>sudo nice -n -20<\/code>\uff0c\u610f\u601d\u662f\u628a\u8c03\u5ea6\u4f18\u5148\u7ea7\u62c9\u6ee1\u2014\u2014\u540c\u6b65\u662f\u5b9e\u65f6\u6027\u654f\u611f\u64cd\u4f5c\uff0c\u8fd9\u6761\u540e\u9762\u8fd8\u4f1a\u518d\u63d0\u3002<\/p>\n<p>\u5b98\u65b9\u6587\u6863\u6ca1\u5199\u6574\u673a\u6700\u4f4e\u914d\u7f6e\u8868\uff0c\u8fd9\u8bdd\u6211\u5148\u8bf4\u5728\u524d\u5934\u3002\u6211\u6309\u4e24\u5904\u5b98\u65b9\u53e3\u5f84\u63a8\u4e86\u4e24\u6863\uff1a\u4e00\u662f README \u7684\u6a21\u578b\u4f53\u79ef\u8868\uff08\u5185\u5b58\u5360\u7528\u6309\u8282\u70b9\u6570\u644a\u8584\uff0c\u6839\u8282\u70b9\u7565\u591a\uff09\uff0c\u4e8c\u662f\u6811\u8393\u6d3e\u6587\u6863\u660e\u786e\u8981\u6c42 64 \u4f4d\u7cfb\u7edf\u3002<\/p>\n<ul>\n<li><strong>\u6700\u4f4e\u53ef\u7528\u6863<\/strong>\uff1a2 \u53f0 8GB \u5185\u5b58\u7684 64 \u4f4d\u8bbe\u5907\uff0c\u5343\u5146\u7f51\u7ebf\u76f4\u8fde\uff0c\u8dd1 0.6B \u5230 3B \u9a8c\u8bc1\u94fe\u8def\u6ca1\u95ee\u9898\u3002\u8fd9\u662f\u6309\u6a21\u578b\u4f53\u79ef\u8868\u63a8\u7684\uff0c\u4e0d\u662f\u5b98\u65b9\u6570\u5b57\u3002<\/li>\n<li><strong>\u5efa\u8bae\u6863<\/strong>\uff1a4 \u53f0 24GB \u5185\u5b58\u3001\u4e07\u5146\u6216\u96f7\u7535\u4e92\u8054\u3002\u8fd9\u4e0d\u662f\u6211\u62cd\u8111\u888b\u2014\u2014\u5b98\u65b9 README \u5f15\u7528\u7684\u5b9e\u6d4b\u5e16\uff08discussion <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/discussions\/147\">#147<\/a>\uff09\u5c31\u662f 4 \u53f0 Mac Mini M4 Pro 24GB \u96f7\u7535 5 \u4e92\u8054\u8dd1 Llama 3.3 70B\uff0c\u6bcf\u79d2 6.51 \u4e2a token\uff0c2 \u53f0\u65f6\u662f 3.31\uff0c\u540c\u6837\u914d\u7f6e\u8d70 10G \u7f51\u7ebf\u662f 6.04\u3002\u78c1\u76d8\u65b9\u9762\uff0c\u4ece 0.9GB\uff080.6B\uff09\u5230 238GB\uff08405B\uff09\u6309\u6a21\u578b\u8868\u5bf9\u53f7\u5165\u5ea7\u3002<\/li>\n<\/ul>\n<h2 id=\"_5\">\u6211\u8dd1\u51fa\u6765\u7684\u56db\u4e2a\u6570\u5b57<\/h2>\n<p>\u88c5\u5b8c\u6211\u5c31\u60f3\u9a8c\u8bc1\u4e24\u4ef6\u4e8b\uff1a\u5355\u673a\u6027\u80fd\u5230\u5e95\u4ec0\u4e48\u6c34\u5e73\uff0c\u62c6\u6210\u4e24\u53f0\u5230\u5e95\u662f\u52a0\u901f\u8fd8\u662f\u62d6\u7d2f\u3002<\/p>\n<p>\u7b2c\u4e00\u4e2a\u6570\u5b57\uff0c\u5355\u8282\u70b9 Qwen3 0.6B Q40\uff0c8 \u7ebf\u7a0b\uff1a<\/p>\n<pre><code class=\"language-text\">Evaluation    tokens\/s: 122.10 (8.19 ms\/tok)\nPrediction    tokens\/s: 77.05 (12.98 ms\/tok)\n<\/code><\/pre>\n<p>\u8f93\u51fa\u300cThe capital of France is Paris.\u300d\uff0c\u7b54\u6848\u6b63\u786e\u3002\u987a\u4fbf\u649e\u51fa\u4e00\u4e2a\u53cd\u5e38\u8bc6\u7684\u53d1\u73b0\uff1a\u6211\u628a\u7ebf\u7a0b\u6570\u964d\u5230 4\uff0c\u9884\u6d4b\u901f\u5ea6\u53cd\u800c\u51b2\u5230 <strong>165.31 tok\/s<\/strong>\uff0c\u6bd4 8 \u7ebf\u7a0b\u5feb\u4e86\u4e00\u500d\u8fd8\u591a\u3002\u6211\u8fd9\u9897 M2 \u662f 4 \u6027\u80fd\u6838 + 4 \u80fd\u6548\u6838\u7684\u5f02\u6784\u5e03\u5c40\uff0c8 \u7ebf\u7a0b\u628a\u4efb\u52a1\u644a\u5230\u80fd\u6548\u6838\u4e0a\u53cd\u800c\u62d6\u540e\u817f\u3002\u5728 Mac \u4e0a\u8dd1\u7684\u670b\u53cb\uff0c<code>--nthreads<\/code> \u5148\u586b\u7269\u7406\u6027\u80fd\u6838\u6570\u8bd5\u8bd5\u3002<\/p>\n<p>\u7b2c\u4e8c\u4e2a\u6570\u5b57\uff0c\u672c\u673a\u518d\u8d77\u4e00\u4e2a worker \u7ec4\u6210\u4e24\u8282\u70b9\uff1a<\/p>\n<pre><code class=\"language-text\">Prediction tokens\/s: 44.26\nPred 8 ms Sync 6 ms | Sent 63 kB Recv 356 kB\n<\/code><\/pre>\n<p>\u540c\u4e00\u53f0\u673a\u5668\u4e0a\u62c6\u4e24\u4e2a\u8282\u70b9\uff0c\u901f\u5ea6\u4ece 165 \u6389\u5230 44\u2014\u2014\u672c\u673a\u6d4b\u8bd5\u53ea\u80fd\u8bc1\u660e\u529f\u80fd\u901a\u4e86\uff0c\u8bc1\u660e\u4e0d\u4e86\u52a0\u901f\u3002\u771f\u6b63\u7684\u6536\u76ca\u8981\u770b discussion #147 \u90a3\u79cd\u771f\u5b9e\u96c6\u7fa4\uff1a70B \u4ece\u5355\u673a\u88c5\u4e0d\u4e0b\uff0c\u5230 2 \u53f0 3.31\u30014 \u53f0 6.51\uff0c\u7ffb\u500d\u5173\u7cfb\u57fa\u672c\u6210\u7acb\u3002<\/p>\n<p>\u7b2c\u4e09\u4e2a\u6570\u5b57\u6700\u6709\u610f\u601d\u3002\u5355\u8282\u70b9\u8f93\u51fa\u300cParis\u300d\uff0c\u4e24\u8282\u70b9\u8dd1\u540c\u4e00\u4e2a prompt\u3001\u540c\u6837\u7684\u8d2a\u5fc3\u89e3\u7801\uff0c\u8f93\u51fa\u53d8\u6210\u4e86\u300c<strong>Rome<\/strong>\u300d\u3002\u6743\u91cd\u5207\u7247\u91cd\u6392\u4e4b\u540e\u6d6e\u70b9\u7d2f\u52a0\u987a\u5e8f\u53d8\u4e86\uff0c\u6570\u503c\u6f02\u79fb\u628a argmax \u9876\u7ffb\u4e86\u4e00\u4e2a\u4f4d\u7f6e\u3002\u8fd9\u4e0d\u662f\u6211\u624b\u6296\uff0c\u540e\u9762\u7ffb\u793e\u533a\u5e16\u5b50\u65f6\u53d1\u73b0 Jeff Geerling \u5728\u5927\u89c4\u6a21\u96c6\u7fa4\u4e0a\u62a5\u544a\u8fc7\u300c\u751f\u6210\u7684 token \u6709\u65f6\u662f\u4e71\u7801\u300d\uff0c\u8ddf\u6211\u5728 loopback \u4e0a\u649e\u5230\u7684\u662f\u540c\u4e00\u7c7b\u95ee\u9898\u7684\u8f7b\u91cd\u4e24\u6863\u3002<strong>\u5206\u5e03\u5f0f\u5207\u5206\u4f1a\u6539\u53d8\u8f93\u51fa\u8d28\u91cf<\/strong>\uff0c\u8fd9\u4ef6\u4e8b README \u91cc\u4e00\u4e2a\u5b57\u6ca1\u63d0\uff0c\u6211\u5efa\u8bae\u5b98\u65b9\u5199\u8fdb\u6587\u6863\u3002<\/p>\n<p>\u7b2c\u56db\u4e2a\uff0cAPI \u670d\u52a1\u9a8c\u8bc1\uff1a<\/p>\n<pre><code class=\"language-bash\">curl http:\/\/127.0.0.1:9999\/v1\/models\n# {&quot;data&quot;:[{&quot;id&quot;:&quot;dllama_model_qwen3_0.6b_q40.m&quot;,&quot;object&quot;:&quot;model&quot;,...}]}\ncurl http:\/\/127.0.0.1:9999\/v1\/chat\/completions \\\n  -H &quot;Content-Type: application\/json&quot; \\\n  -d '{&quot;messages&quot;:[{&quot;role&quot;:&quot;user&quot;,&quot;content&quot;:&quot;\u7528\u4e00\u53e5\u8bdd\u4ecb\u7ecd\u676d\u5dde&quot;}]}'\n# \u6b63\u5e38\u8fd4\u56de\uff0cQwen3 \u7684 &lt;think&gt; \u63a8\u7406\u6d41\u4e5f\u80fd\u51fa\n<\/code><\/pre>\n<p>\u63a5 OpenAI SDK \u5199\u7684\u4e0b\u6e38\u5e94\u7528\u57fa\u672c\u53ef\u4ee5\u65e0\u75db\u6362\u57fa\u5ea7\u3002\u4e0d\u8fc7 <code>finish_reason<\/code> \u5b57\u6bb5\u8fd4\u56de\u7684\u662f\u7a7a\u4e32\uff0c\u505a\u4e25\u683c\u7684\u6d41\u5f0f\u6536\u5c3e\u5224\u65ad\u65f6\u8981\u7559\u4e2a\u5fc3\u773c\u3002\u53e6\u5916\u63d0\u9192\u4e00\u53e5\uff1a\u804a\u5929\u6a21\u5f0f <code>dllama chat<\/code> \u662f\u7ed9\u547d\u4ee4\u884c\u51c6\u5907\u7684\uff0ctoken \u4e00\u4e2a\u4e2a\u5f80\u5916\u8e66\u7684\u4f53\u9a8c\u5f88\u76f4\u89c2\uff0c\u9002\u5408\u5148\u62ff\u5b83\u786e\u8ba4\u6a21\u578b\u884c\u4e3a\uff0c\u518d\u5207\u5230 API \u6a21\u5f0f\u63a5\u5de5\u5177\u94fe\u3002<\/p>\n<p>\u56db\u4e2a\u6570\u5b57\u653e\u5728\u4e00\u8d77\uff0c\u7ed3\u8bba\u5176\u5b9e\u5df2\u7ecf\u6d6e\u51fa\u6765\u4e86\uff1a\u5355\u673a\u6027\u80fd\u662f\u5408\u683c\u7684\uff0c\u591a\u8282\u70b9\u5185\u5b58\u644a\u8584\u662f\u771f\u7684\uff0c\u4f46\u300c\u8bbe\u5907\u8d8a\u591a\u8d8a\u5feb\u300d\u8fd9\u53e5\u5ba3\u4f20\u8bed\u5f97\u6253\u4e2a\u8865\u4e01\u2014\u2014\u540c\u6784\u8bbe\u5907\u3001\u6709\u7ebf\u4e92\u8054\u3001\u6a21\u578b\u5927\u5230\u5355\u673a\u88c5\u4e0d\u4e0b\uff0c\u4e09\u4e2a\u6761\u4ef6\u540c\u65f6\u6ee1\u8db3\u65f6\u5b83\u624d\u5151\u73b0\u627f\u8bfa\uff0c\u7f3a\u4e00\u6761\u90fd\u53ef\u80fd\u8d8a\u62c6\u8d8a\u6162\uff0c\u751a\u81f3\u628a\u7b54\u6848\u7b97\u6b6a\u3002<\/p>\n<h2 id=\"issue\">\u7ffb\u5b8c issue \u548c\u793e\u533a\u5e16\u5b50\uff0c\u6211\u5f52\u7eb3\u51fa\u4e94\u6761\u5751<\/h2>\n<p>\u4ed3\u5e93\u5f00\u653e issue \u53ea\u6709 48 \u4e2a\uff0c\u4e0d\u7b97\u70ed\u95f9\uff0c\u4f46\u8bc4\u8bba\u533a\u8d28\u91cf\u610f\u5916\u5730\u9ad8\uff0c\u52a0\u4e0a\u7b2c\u4e09\u65b9\u5b9e\u6d4b\u5e16\uff0c\u80fd\u62fc\u51fa\u4e00\u5f20\u5b8c\u6574\u7684\u907f\u5751\u56fe\uff1a<\/p>\n<ol>\n<li><strong>\u52a0\u8282\u70b9\u4e0d\u4e00\u5b9a\u7ebf\u6027\u52a0\u901f\uff0c\u751a\u81f3\u53ef\u80fd\u5012\u9000<\/strong>\u3002issue <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/20\">#20<\/a> \u91cc\u6709\u4eba\u4ece 4 \u53f0\u6811\u8393\u6d3e\u6269\u5230 8 \u53f0\uff0c\u540c\u6b65\u65f6\u95f4\u7a81\u7136\u66b4\u6da8\uff0c\u9ad8\u8d5e\u8ba8\u8bba\u4e86\u5f88\u4e45\uff1bissue <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/179\">#179<\/a> \u5728 12 \u4ee3 NUC \u4e0a\u8dd1 8B \u770b\u4e0d\u5230\u6269\u5c55\u6536\u76ca\u3002\u7b2c\u4e09\u65b9\u6c47\u603b\uff08<a href=\"https:\/\/specpicks.com\/reviews\/raspberry-pi-4-local-llm-cluster-realistic-2026\">SpecPicks 2026 \u5e74\u7684\u5206\u6790\u6587<\/a>\uff09\u7ed9\u7684\u89c4\u5f8b\u662f\uff1a4 \u8282\u70b9\u7ea6\u6478 2 \u500d\u901f\u5ea6\u30018 \u8282\u70b9\u7ea6\u6478 3 \u500d\uff0c\u5343\u5146\u4ee5\u592a\u7f51\u5728 8 \u8282\u70b9\u4ee5\u4e0a\u5c31\u6210\u5899\u4e86\u3002<\/li>\n<li><strong>\u522b\u62ff\u5b83\u8ddf\u6b63\u7ecf\u786c\u4ef6\u62fc\u6027\u4ef7\u6bd4<\/strong>\u3002Geerling \u5728<a href=\"https:\/\/github.com\/geerlingguy\/beowulf-ai-cluster\/issues\/6\">10 \u8282\u70b9\u6811\u8393\u6d3e 5 \u96c6\u7fa4\u7684\u57fa\u51c6\u8bb0\u5f55<\/a>\u91cc\u5199\u5f97\u5f88\u76f4\u767d\uff1a70B \u6a21\u578b 8 \u8282\u70b9\u8dd1\u5230 0.85 tok\/s\uff0c\u6bd4\u7528 Vulkan \u52a0\u901f\u7684 x86 \u96c6\u7fa4\u6162 5 \u500d\uff0c\u8fd8\u5076\u5c14\u5410\u4e71\u7801\u3002\u4ed6\u7684\u7ed3\u8bba\u662f\u4e09\u5343\u7f8e\u5143\u7684\u6811\u8393\u6d3e\u96c6\u7fa4\u4e0d\u5212\u7b97\u2014\u2014\u8fd9\u7b14\u8d26\u5bf9\u60f3\u65b0\u4e70\u8bbe\u5907\u7ec4\u96c6\u7fa4\u7684\u4eba\u540c\u6837\u6210\u7acb\u3002<\/li>\n<li><strong>\u6a21\u578b\u8f6c\u6362\u662f\u9690\u5f62\u5de5\u7a0b<\/strong>\u3002issue <a href=\"https:\/\/github.com\/b4rtaz\/distributed-llama\/issues\/185\">#185<\/a>\uff1aMistral Large \u7684 tokenizer \u8f6c\u4e0d\u8fc7\u53bb\uff0c\u62a5\u300cCannot resolve bosId or eosIds\u300d\u3002\u5b98\u65b9\u9884\u8f6c\u6362\u4ed3\u5e93\u6ca1\u8986\u76d6\u7684\u6a21\u578b\uff0c\u4f60\u5f97\u81ea\u5df1\u8fc7 <code>convert-hf.py<\/code> \u8fd9\u9053\u574e\uff0c\u62a5\u9519\u4e86\u6ca1\u4eba\u66ff\u4f60\u515c\u3002<\/li>\n<li><strong>\u5de5\u4eba\u8282\u70b9\u8981\u63d0\u6743\u8c03\u5ea6<\/strong>\u3002\u6811\u8393\u6d3e\u6587\u6863\u91cc\u5199\u7684\u662f <code>sudo nice -n -20 .\/dllama worker<\/code>\u2014\u2014\u540c\u6b65\u662f\u5b9e\u65f6\u6027\u654f\u611f\u64cd\u4f5c\uff0c\u666e\u901a\u4f18\u5148\u7ea7\u4e0b\u522b\u7684\u8fdb\u7a0b\u4e00\u62a2\u65f6\u95f4\u7247\uff0c\u6574\u6761\u6d41\u6c34\u7ebf\u8ddf\u7740\u6296\u3002\u8fd9\u6761\u5f88\u591a\u4eba\u4f1a\u6f0f\u3002<\/li>\n<li><strong>\u4e5f\u6709\u987a\u5229\u7684\u6b63\u4f8b<\/strong>\u3002Seeed Studio \u5b98\u65b9 wiki \u6709\u7bc7\u90e8\u7f72\u8bb0\u5f55\uff1a1 \u53f0 8GB \u6811\u8393\u6d3e\u5f53\u6839\u8282\u70b9\u5e26 3 \u53f0 4GB \u5de5\u4eba\u8282\u70b9\u8dd1 DeepSeek 8B\uff0c\u5b9e\u6d4b 6.06 tok\/s\u30024GB \u7684\u5361\u80fd\u5f53\u5de5\u4eba\u7528\uff0c\u8bf4\u660e\u5185\u5b58\u644a\u8584\u7684\u673a\u5236\u662f\u771f\u7684\u5728\u5de5\u4f5c\u3002<\/li>\n<\/ol>\n<h2 id=\"_6\">\u6211\u6700\u540e\u7684\u5224\u65ad<\/h2>\n<p>\u503c\u4e0d\u503c\u5f97\u4e0a\u624b\uff0c\u6211\u6309\u4eba\u5206\u3002<\/p>\n<p><strong>\u8be5\u4e0a\u624b\u7684\u4eba<\/strong>\uff1a\u624b\u91cc\u5df2\u7ecf\u8eba\u7740 2 \u53f0\u6216 4 \u53f0\u540c\u6784\u8bbe\u5907\u3001\u5185\u5b58\u52a0\u8d77\u6765\u88c5\u5f97\u4e0b\u4f60\u60f3\u8dd1\u7684\u6a21\u578b\u3001\u613f\u610f\u62ff\u4e00\u4e0b\u5348\u6362\u4e00\u4e2a\u300c\u5927\u6a21\u578b\u5728\u6211\u81ea\u5df1\u8bbe\u5907\u4e0a\u8dd1\u8d77\u6765\u300d\u7684\u4f53\u9a8c\u2014\u2014\u8fd9\u4e2a\u9879\u76ee\u7684\u5b8c\u6210\u5ea6\u5bf9\u5f97\u8d77\u4f60\u7684\u534a\u5929\u65f6\u95f4\uff0cCPU \u8def\u5f84\u5f00\u7bb1\u5373\u7528\uff0c\u7f16\u8bd1 30 \u79d2\uff0cAPI \u8fd8\u80fd\u76f4\u63a5\u63a5\u73b0\u6709\u5de5\u5177\u94fe\u3002<\/p>\n<p><strong>\u522b\u78b0\u7684\u4eba<\/strong>\uff1a\u60f3\u4e3a\u65b0\u786c\u4ef6\u7ec4\u96c6\u7fa4\u7684\u2014\u2014\u540c\u6837\u7684\u94b1\u4e70\u4e00\u53f0 32GB \u4e8c\u624b mini \u4e3b\u673a\uff0c\u901f\u5ea6\u5927\u6982\u7387\u66f4\u5feb\u3001\u8fd8\u4e0d\u6298\u817e\uff0cSpecPicks \u7684\u5bf9\u6bd4\u8868\u7b97\u5f97\u5f88\u6e05\u695a\uff1b\u8981\u7ed9\u751f\u4ea7\u73af\u5883\u4f9b API \u7684\u2014\u2014<code>finish_reason<\/code> \u7a7a\u4e32\u3001\u4e24\u8282\u70b9\u7684\u7b54\u6848\u6f02\u79fb\u3001GPU \u8def\u5f84\u672a\u6210\u719f\uff0c\u8fd9\u4e09\u6761\u4efb\u4f55\u4e00\u6761\u90fd\u591f\u4f60\u5934\u75bc\uff1b\u60f3\u6df7\u642d\u5f02\u6784\u8bbe\u5907\u7684\u2014\u2014\u5f20\u91cf\u5e76\u884c\u5bf9\u8282\u70b9\u95f4\u5e26\u5bbd\u592a\u654f\u611f\uff0c\u7b14\u8bb0\u672c Wi-Fi \u5c31\u522b\u60f3\u4e86\u3002<\/p>\n<p>\u66ff\u4ee3\u65b9\u6848\u6211\u70b9\u540d\u4e24\u4e2a\u3002<strong>llama.cpp RPC<\/strong>\uff1a\u6309\u5c42\u5207\u5206\uff0c\u76f4\u63a5\u5403 GGUF\uff0c\u652f\u6301\u5f02\u6784\u8282\u70b9\uff0c\u751f\u6001\u6210\u719f\u5f97\u591a\uff1b\u4ee3\u4ef7\u662f Geerling \u5b9e\u6d4b\u7684\u6781\u7aef\u6162\u901f\uff080.28 tok\/s \u5bf9 0.85\uff09\u548c\u66f4\u9ad8\u7684\u7f51\u7edc\u6765\u56de\u6b21\u6570\u3002<strong>exo<\/strong>\uff1a\u91ce\u5fc3\u66f4\u5927\u3001\u63a5\u53e3\u66f4\u82b1\u54e8\uff0c\u4f46 Geerling \u5728 2 \u5230 3 \u8282\u70b9\u7684\u5c0f\u96c6\u7fa4\u4e0a\u8fde 3B \u6a21\u578b\u90fd\u6ca1\u8dd1\u8d77\u6765\uff0c\u6211\u73b0\u5728\u4e0d\u6562\u628a\u5b83\u5217\u6210\u6b63\u7ecf\u9009\u9879\u3002\u4e24\u4e2a\u7ef4\u5ea6\u6446\u5728\u8fd9\uff1a\u8981\u7075\u6d3b\u548c\u751f\u6001\u9009 llama.cpp RPC\uff0c\u8981\u901f\u5ea6\u548c\u5185\u5b58\u644a\u8584\u9009 distributed-llama\uff0c\u4e0d\u5b58\u5728\u4e24\u5934\u90fd\u5360\u7684\u3002<\/p>\n<p>\u4e00\u53e5\u8bdd\u6536\u5c3e\uff1a\u603b\u800c\u8a00\u4e4b\uff0cdistributed-llama \u7528\u6700\u6734\u7d20\u7684 C++ \u548c TCP \u628a\u300c\u5bb6\u7528\u8bbe\u5907\u7ec4\u63a8\u7406\u96c6\u7fa4\u300d\u8fd9\u4ef6\u4e8b\u505a\u5230\u4e86\u80fd\u7528\u7684\u7a0b\u5ea6\uff0c\u6211\u987a\u624b\u62c9\u4e86\u4e0b\uff0c\u73b0\u5728\u662f 3,132 \u4e2a star\uff0cv0.17.0 \u4e24\u5929\u524d\u521a\u52a0\u4e0a Qwen3.5 \u652f\u6301\uff0c\u6700\u8fd1\u56db\u5468 11 \u6b21\u63d0\u4ea4\u2014\u2014\u8282\u594f\u4e0d\u5feb\uff0c\u4f46\u6bcf\u4e2a\u53d1\u7248\u90fd\u6709\u771f\u4e1c\u897f\u3002\u6700\u540e\u8bf4\u53e5\u5b9e\u5728\u7684\uff0c\u5b83\u6551\u4e0d\u4e86\u60f3\u767d\u5ad6 70B \u7684\u4eba\uff0c\u4f46\u80fd\u8ba9\u4f60\u5bb6\u5403\u7070\u7684\u65e7\u8bbe\u5907\u7b2c\u4e00\u6b21\u4e3a\u8dd1\u5927\u6a21\u578b\u51fa\u4e0a\u529b\uff1b\u81f3\u4e8e\u90a3\u4e2a Paris \u53d8 Rome \u7684\u5c0f\u60ca\u559c\uff0c\u5c31\u5f53\u662f\u5206\u5e03\u5f0f\u7cfb\u7edf\u9001\u7ed9\u8f83\u771f\u7528\u6237\u7684\u89c1\u9762\u793c\u5427\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u628a\u5bb6\u91cc\u95f2\u7f6e\u8bbe\u5907\u4e32\u6210\u63a8\u7406\u96c6\u7fa4\u8fd9\u4e2a\u5ff5\u5934\uff0c\u6211\u62ff distributed-llama \u771f\u7f16\u771f\u8dd1\u4e86\u4e00\u904d\uff1a\u5355\u8282\u70b9 Qwen3 0.6B \u6bcf\u79d2 165 \u4e2a token\uff0c\u53ef\u62c6\u6210\u4e24\u8282\u70b9\u540e\u7b54\u6848\u7adf\u4ece Paris \u6f02\u6210\u4e86 Rome\u3002\u54ea\u4e09\u6761\u9650\u5236\u8981\u5148\u80cc\u4e0b\u6765\uff0c\u4e94\u6761\u5751\u662f\u4ec0\u4e48\uff0c\u9002\u5408\u8c01\u3001\u8c01\u522b\u78b0\uff0c\u4e00\u6b21\u8bf4\u6e05\u3002<\/p>\n","protected":false},"author":1,"featured_media":10021547,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10021548","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"https:\/\/itmax.cn\/wp-content\/uploads\/2026\/11\/distributed-llama-cover.png","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021548","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10021548"}],"version-history":[{"count":1,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021548\/revisions"}],"predecessor-version":[{"id":10021549,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021548\/revisions\/10021549"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/media\/10021547"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10021548"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10021548"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10021548"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}