{"id":10021524,"date":"2026-10-30T09:38:30","date_gmt":"2026-10-30T01:38:30","guid":{"rendered":"https:\/\/itmax.cn\/?p=10021524"},"modified":"2026-10-30T09:38:42","modified_gmt":"2026-10-30T01:38:42","slug":"ms-swift-20261030","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10021524","title":{"rendered":"\u8ba9\u6a21\u578b\u5b66\u4f1a\u8bf4\u4f60\u7684\u884c\u8bdd\uff5cms-swift \u4e0a\u624b\u524d\uff0c\u6211\u628a\u5b83 600 \u6761 issue \u7ffb\u4e86\u4e00\u904d"},"content":{"rendered":"<figure style=\"margin:0 0 32px;text-align:center;\"><img decoding=\"async\" src=\"https:\/\/itmax.cn\/wp-content\/uploads\/2026\/10\/ms-swift-cover.png\" alt=\"ms-swift \u9879\u76ee\u5c01\u9762\uff1a\u9b54\u642d\u51fa\u54c1\u7684\u5927\u6a21\u578b\u5fae\u8c03\u5168\u5bb6\u6876\" style=\"width:100%;height:auto;max-width:960px;border-radius:8px;display:block;margin:0 auto;\" \/><figcaption style=\"margin-top:12px;font-size:14px;color:#868EAA;line-height:1.6;\">ms-swift \u5728 GitHub \u4e0a\u7684\u5f53\u524d\u72b6\u6001\uff1a15,986 star\u3001\u652f\u6301 600+ \u6a21\u578b\u3001\u8fd1\u56db\u5468 75 \u6b21\u63d0\u4ea4<\/figcaption><\/figure>\n<h2 id=\"_1\">\u6211\u4e3a\u4ec0\u4e48\u76ef\u4e0a\u5b83<\/h2>\n<p>\u53bb\u5e74\u5e2e\u670b\u53cb\u516c\u53f8\u8c03\u4e00\u4e2a\u552e\u540e\u95ee\u7b54\u673a\u5668\u4eba\uff0c\u6298\u817e\u4e86\u5feb\u4e00\u4e2a\u6708\uff0c\u6700\u540e\u5361\u5728\u4e00\u4e2a\u5f88\u5c34\u5c2c\u7684\u5730\u65b9\uff1a\u5ba2\u6237\u95ee\u300c\u8fd9\u6279\u8f74\u627f\u7684\u6e38\u9699\u80fd\u8c03\u5230 C3 \u7ec4\u5417\u300d\uff0c\u673a\u5668\u4eba\u7b54\u5f97\u50cf\u6559\u79d1\u4e66\uff0c\u504f\u504f\u7b54\u4e0d\u5230\u70b9\u5b50\u4e0a\u3002\u884c\u8bdd\u5168\u5728 prompt \u91cc\u585e\u7740\uff0c\u4e00\u585e\u5c31\u662f\u4e09\u5343\u591a token\uff0c\u8d35\u4e0d\u8bf4\uff0c\u6a21\u578b\u8fd8\u662f\u6293\u4e0d\u4f4f\u91cd\u70b9\u3002\u6362 RAG \u4e5f\u8bd5\u4e86\uff0c\u68c0\u7d22\u56de\u6765\u7684\u6bb5\u843d\u5357\u8f95\u5317\u8f99\uff0c\u6a21\u578b\u62ff\u7740\u9519\u8bef\u7684\u4e0a\u4e0b\u6587\u4e00\u672c\u6b63\u7ecf\u5730\u80e1\u8bf4\u3002<\/p>\n<p>\u90a3\u65f6\u5019\u6211\u4eec\u5750\u4e0b\u6765\u7b97\u4e86\u4e00\u7b14\u8d26\uff1a\u672f\u8bed\u5c31\u90a3\u4e48\u591a\uff0c\u95ee\u7b54\u5bf9\u6512\u4e86\u51e0\u767e\u6761\uff0c\u4e3a\u4ec0\u4e48\u4e0d\u76f4\u63a5\u628a\u8fd9\u70b9\u4e1c\u897f\u8bad\u8fdb\u6a21\u578b\u91cc\uff1f\u5ff5\u5934\u6709\u4e86\uff0c\u7136\u540e\u6211\u6253\u5f00\u4e86 transformers \u7684 Trainer\u2014\u2014\u6570\u636e\u5904\u7406\u8981\u81ea\u5df1\u5199 collator\uff0c\u5bf9\u8bdd\u6a21\u677f\u8981\u81ea\u5df1\u5bf9\u9f50\uff0c\u663e\u5b58\u4e0d\u591f\u8981\u81ea\u5df1\u7814\u7a76 gradient checkpointing \u548c ZeRO \u7684\u7ec4\u5408\uff0c\u65ad\u70b9\u7eed\u8bad\u8981\u81ea\u5df1\u7ba1\u3002\u6211\u627f\u8ba4\u8fd9\u4e9b\u6211\u90fd\u4f1a\u5199\uff0c\u4f46\u6bcf\u6362\u4e00\u4e2a\u57fa\u5ea7\u6a21\u578b\u5c31\u5f97\u91cd\u5bf9\u4e00\u904d\u6a21\u677f\uff0c\u8fd9\u79cd\u4f53\u529b\u6d3b\u5e72\u5230\u7b2c\u4e09\u5929\u6211\u5c31\u5f00\u59cb\u6000\u7591\u4eba\u751f\u3002<\/p>\n<p>\u540e\u6765\u4e00\u4e2a\u5728\u9b54\u642d\u793e\u533a\u6df7\u7684\u670b\u53cb\u8ddf\u6211\u8bf4\uff1a\u4f60\u8fd9\u4e9b\u6d3b\uff0cms-swift \u4e00\u6761\u547d\u4ee4\u5c31\u5305\u4e86\u3002\u6211\u5f53\u65f6\u662f\u4e0d\u4fe1\u7684\uff0c\u300c\u4e00\u6761\u547d\u4ee4\u5fae\u8c03\u300d\u8fd9\u79cd\u8bdd\u6211\u542c\u8fc7\u592a\u591a\u6b21\uff0c\u591a\u534a\u662f demo \u80fd\u8dd1\u3001\u751f\u4ea7\u62c9\u80ef\u3002\u4f46\u8fd9\u6b21\u6211\u628a\u5b83\u7684\u6587\u6863\u548c issue \u533a\u7ffb\u4e86\u4e00\u904d\uff0c\u8d8a\u7ffb\u8d8a\u89c9\u5f97\u8fd9\u4e2a\u9879\u76ee\u88ab\u4e25\u91cd\u4f4e\u4f30\u4e86\u2014\u2014star \u624d\u4e00\u4e07\u4e94\u51fa\u5934\uff0c\u8ddf\u5b83\u7684\u5b9e\u9645\u5b8c\u6210\u5ea6\u5b8c\u5168\u4e0d\u5339\u914d\u3002<\/p>\n<h2 id=\"readme\">\u80fd\u5e72\u7684\u4e8b\u6bd4 README \u4e0a\u770b\u7740\u591a\uff0c\u4f46\u5206\u771f\u6d3b\u548c\u6446\u8bbe<\/h2>\n<p>ms-swift \u662f ModelScope \u5b98\u65b9\u51fa\u7684\u8bad\u7ec3\u6846\u67b6\uff0c\u540d\u5b57\u5c55\u5f00\u662f Scalable lightWeight Infrastructure for Fine-Tuning\u3002\u5b83\u8986\u76d6\u7684\u4e1c\u897f\u53ef\u4ee5\u5206\u4e94\u62e8\u770b\u3002<\/p>\n<p><strong>\u7b2c\u4e00\u62e8\uff1a\u5168\u94fe\u8def\u547d\u4ee4\u884c\u3002<\/strong> <code>swift sft<\/code> \u8bad\u7ec3\u3001<code>swift infer<\/code> \u63a8\u7406\u3001<code>swift export<\/code> \u5408\u5e76\u5bfc\u51fa\u3001<code>swift deploy<\/code> \u8d77\u670d\u52a1\u3001<code>swift eval<\/code> \u8dd1\u57fa\u51c6\uff0c\u4e00\u6761\u547d\u4ee4\u5bf9\u5e94\u4e00\u4e2a\u73af\u8282\uff0c\u6a21\u578b ID \u76f4\u63a5\u586b\u9b54\u642d\u6216 HuggingFace \u7684\u4ed3\u5e93\u8def\u5f84\uff0c\u6743\u91cd\u81ea\u52a8\u4e0b\u8f7d\u3002\u5b98\u65b9\u5ba3\u79f0\u652f\u6301 600+ \u6587\u672c\u6a21\u578b\u548c 400+ \u591a\u6a21\u6001\u6a21\u578b\uff0c\u5185\u7f6e 150+ \u6570\u636e\u96c6\u3002\u6211\u7279\u610f\u53bb\u7ffb\u4e86\u5b83\u7684\u63d0\u4ea4\u8bb0\u5f55\uff1a10 \u6708 29 \u65e5\u8fd8\u5728\u7ed9 NVIDIA \u65b0\u51fa\u7684\u8f6c\u5199\u6a21\u578b\u505a Day-0 \u652f\u6301\uff0c\u8ddf\u8fdb\u901f\u5ea6\u662f\u771f\u5feb\uff0c\u4e0d\u662f README \u91cc\u7684\u6b7b\u6570\u5b57\u3002<\/p>\n<p><strong>\u7b2c\u4e8c\u62e8\uff1a\u8f7b\u91cf\u8bad\u7ec3\u65b9\u6cd5\u3002<\/strong> LoRA\u3001QLoRA\u3001DoRA\u3001LoRA+\u3001LongLoRA\u3001LISA \u8fd9\u4e9b\u57fa\u672c\u76d8\u90fd\u9f50\uff0c\u8fd8\u80fd\u76f4\u63a5\u5728 BNB\u3001AWQ\u3001GPTQ \u91cf\u5316\u8fc7\u7684\u6a21\u578b\u4e0a\u505a\u8bad\u7ec3\u3002README \u91cc\u6709\u4e2a\u6570\u5b57\u6211\u4e13\u95e8\u6838\u4e86\u51fa\u5904\uff1a7B \u6a21\u578b\u7528\u91cf\u5316\u8bad\u7ec3\u53ea\u8981 9GB \u663e\u5b58\uff0c\u8fd9\u4e2a\u5199\u5728\u5b98\u65b9\u7279\u6027\u5217\u8868\u91cc\uff0c\u4e0d\u662f\u6211\u8f6c\u8ff0\u7684\u4e8c\u624b\u4f20\u95fb\u3002<\/p>\n<p><strong>\u7b2c\u4e09\u62e8\uff1a\u5927\u89c4\u6a21\u8bad\u7ec3\u3002<\/strong> Megatron \u540e\u7aef\u7684 TP\/PP\/EP \u5e76\u884c\uff0cDeepSpeed ZeRO2\/ZeRO3\uff0c\u8fd8\u6709\u4e00\u6574\u5957 GRPO \u5bb6\u65cf\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u2014\u2014DAPO\u3001GSPO\u3001CISPO\u3001CHORD\u3001RLOO\u3001Reinforce++\uff0crollout \u9636\u6bb5\u53ef\u4ee5\u6302 vLLM \u63d0\u901f\u3002\u8fd9\u4e00\u62e8\u770b\u7740\u552c\u4eba\uff0c\u4f46\u8bf4\u5b9e\u8bdd\uff0cMegatron \u90a3\u5957\u5bf9\u6ca1\u6709\u51e0\u5341\u5f20\u5361\u7684\u4eba\u5c31\u662f\u6446\u8bbe\uff0cGRPO \u5355\u5361\u8dd1\u4e2a\u5c0f\u6a21\u578b\u73a9\u73a9\u53ef\u4ee5\uff0c\u8ba4\u771f\u505a RL \u90a3\u662f\u53e6\u4e00\u4e2a\u8d5b\u9053\u7684\u4e8b\u3002<\/p>\n<p><strong>\u7b2c\u56db\u62e8\uff1a\u6570\u636e\u4e0e\u5bfc\u51fa\u3002<\/strong> \u8fd9\u5757\u662f\u5bb9\u6613\u88ab\u7565\u8fc7\u3001\u4f46\u771f\u7528\u8d77\u6765\u6700\u7701\u5fc3\u7684\u90e8\u5206\u3002\u81ea\u5b9a\u4e49\u6570\u636e\u96c6\u53ea\u8981\u6574\u7406\u6210\u6807\u51c6\u7684 messages \u683c\u5f0f\u5b58\u6210 jsonl\uff0c\u672c\u5730\u8def\u5f84\u548c\u9b54\u642d\u7684\u4ed3\u5e93 ID \u53ef\u4ee5\u6df7\u7740\u4f20\u7ed9 <code>--dataset<\/code>\uff0c<code>#500<\/code> \u8fd9\u79cd\u540e\u7f00\u76f4\u63a5\u63a7\u5236\u91c7\u6837\u6761\u6570\uff0c\u5199\u5c0f\u6837\u9a8c\u8bc1\u811a\u672c\u7684\u65f6\u5019\u7279\u522b\u987a\u624b\u3002\u8bad\u5b8c\u7684\u5bfc\u51fa\u4fa7\u4e5f\u9f50\uff1aGPTQ\u3001AWQ\u3001BNB\u3001FP8 \u56db\u79cd\u91cf\u5316\u5bfc\u51fa\uff0c\u5bfc\u51fa\u7684\u6a21\u578b\u76f4\u63a5\u63a5 vLLM\u3001SGLang\u3001LMDeploy \u505a\u63a8\u7406\u52a0\u901f\u3002\u53e6\u5916\u5b83\u8fd8\u6709\u4e2a <code>swift sample<\/code> \u547d\u4ee4\uff0c\u53ef\u4ee5\u4ece\u5927\u6a21\u578b\u7684\u63a5\u53e3\u505a\u84b8\u998f\u91c7\u6837\u6765\u9020\u8bad\u7ec3\u6570\u636e\u2014\u2014\u628a\u300c\u751f\u6210\u6570\u636e\u300d\u8fd9\u4e00\u6b65\u4e5f\u6536\u8fdb\u4e86\u540c\u4e00\u5957\u547d\u4ee4\u884c\uff0c\u8fd9\u4e2a\u8bbe\u8ba1\u6211\u7ed9\u597d\u8bc4\u3002<\/p>\n<p><strong>\u7b2c\u4e94\u62e8\uff1aWeb-UI\u3002<\/strong> <code>swift web-ui<\/code> \u8d77\u4e00\u4e2a Gradio \u754c\u9762\uff0c\u70b9\u70b9\u70b9\u5c31\u80fd\u914d\u8bad\u7ec3\u3002\u6211\u7684\u5224\u65ad\u662f\uff1a\u62ff\u5b83\u7ed9\u56e2\u961f\u91cc\u4e0d\u505a\u8bad\u7ec3\u7684\u540c\u5b66\u6f14\u793a\u6d41\u7a0b\u591f\u7528\uff0c\u771f\u8dd1\u4efb\u52a1\u8fd8\u662f\u56de\u5230\u547d\u4ee4\u884c\u2014\u2014\u53c2\u6570\u591a\u5230 Web \u754c\u9762\u91cc\u7ffb\u627e\u4e0d\u5982\u5199\u811a\u672c\u53ef\u590d\u73b0\u3002<\/p>\n<p>\u591a\u6a21\u6001\u8fd9\u5757\u4e5f\u503c\u5f97\u4e00\u8bf4\uff1a\u56fe\u7247\u3001\u89c6\u9891\u3001\u97f3\u9891\u6df7\u7740\u8bad\u662f\u5b83\u539f\u751f\u652f\u6301\u7684\uff0c\u89c6\u89c9\u5854\u548c\u8bed\u8a00\u5c42\u7684\u5b66\u4e60\u7387\u8fd8\u80fd\u5206\u5f00\u63a7\u5236\uff0c packed \u8bad\u7ec3\u628a\u591a\u6a21\u6001\u6837\u672c\u6253\u5305\u63d0\u901f\u3002\u6211\u7ffb\u4e86\u4e00\u5708 issue \u533a\u7684\u591a\u6a21\u6001\u5e16\u5b50\uff0c\u8dd1\u901a\u7684\u591a\uff0c\u4f46\u56fe\u7247\u5206\u8fa8\u7387\u3001\u89c6\u9891\u5e27\u6570\u8fd9\u7c7b\u9884\u5904\u7406\u53c2\u6570\u9ed8\u8ba4\u503c\u504f\u4fdd\u5b88\uff0c\u6570\u636e\u4fa7\u8981\u81ea\u5df1\u538b\u4e00\u538b\u3002<\/p>\n<p>\u6240\u4ee5\u6211\u7684\u7ed3\u8bba\u662f\uff1a\u7b2c\u4e00\u62e8\u3001\u7b2c\u4e8c\u62e8\u548c\u7b2c\u56db\u62e8\u662f\u771f\u80fd\u7528\u7684\u6d3b\uff0c\u7b2c\u4e09\u62e8\u770b\u4f60\u7684\u5361\uff0c\u7b2c\u4e94\u62e8\u5f53\u4e2a\u95e8\u9762\u3002\u8fd9\u4e2a\u533a\u5206\u5f88\u91cd\u8981\uff0c\u540e\u9762\u8bb2\u5224\u65ad\u7684\u65f6\u5019\u8fd8\u4f1a\u7528\u5230\u3002<\/p>\n<h2 id=\"_2\">\u4e00\u5c42\u58f3\u7684\u806a\u660e\u4e4b\u5904\uff0c\u548c\u5b83\u6b20\u4e0b\u7684\u503a<\/h2>\n<p>\u770b\u5b83\u7684\u67b6\u6784\uff0c\u4f60\u4f1a\u53d1\u73b0 ms-swift \u81ea\u5df1\u6ca1\u6709\u53d1\u660e\u8f6e\u5b50\uff1a\u5e95\u4e0b\u662f transformers Trainer \u52a0 peft \u52a0 deepspeed\uff0c\u5b83\u505a\u7684\u662f\u4e0a\u9762\u90a3\u4e00\u5c42\u547d\u4ee4\u884c\u6ce8\u518c\u8868\u3001\u5bf9\u8bdd\u6a21\u677f\u7cfb\u7edf\u548c\u6570\u636e\u683c\u5f0f\u9002\u914d\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u7684\u597d\u5904\u663e\u800c\u6613\u89c1\u2014\u2014\u4e0a\u6e38\u66f4\u65b0\u5b83\u8ddf\u5f97\u5feb\uff0ctransformers \u4ece 4.33 \u4e00\u8def\u652f\u6301\u5230 5.x\uff0c\u65b0\u6a21\u578b\u53d1\u5e03\u57fa\u672c\u662f\u6539\u4e2a\u6a21\u677f\u5c31\u80fd\u8bad\u3002<\/p>\n<p>\u4f46\u503a\u4e5f\u662f\u8fd9\u4e48\u6b20\u4e0b\u7684\u3002\u7b2c\u4e00\u7b14\u662f\u4f9d\u8d56\u9489\u5f97\u6b7b\uff1a<code>datasets<\/code> \u8981 <code>&gt;=3.0,&lt;4.8.5<\/code>\uff0c<code>peft<\/code> \u8981 <code>&lt;0.21<\/code>\uff0c<code>deepspeed<\/code> \u8981 <code>&gt;=0.19.3,&lt;0.20<\/code>\u3002\u8fd9\u4e0d\u662f\u4ed6\u4eec\u6311\u5254\uff0c\u662f\u4e0a\u6e38\u63a5\u53e3\u771f\u5728\u53d8\uff0c\u4f46\u5bf9\u4f60\u6765\u8bf4\u610f\u5473\u7740\u4e00\u4e2a\u5df2\u7ecf\u88c5\u4e86\u522b\u7684\u8bad\u7ec3\u6846\u67b6\u7684 Python \u73af\u5883\u91cc\uff0c\u5341\u6709\u516b\u4e5d\u8981\u4e3a\u5b83\u5355\u72ec\u5f00\u4e00\u4e2a venv\u3002\u6211\u5efa\u8bae\u4ece\u4e00\u5f00\u59cb\u5c31\u4e60\u60ef\u8fd9\u4ef6\u4e8b\uff0c\u522b\u786c\u585e\u8fdb\u4e3b\u73af\u5883\u3002<\/p>\n<p>\u7b2c\u4e8c\u7b14\u662f\u62bd\u8c61\u7a0e\u3002\u8bad\u7ec3\u5b8c\u7684 checkpoint \u76ee\u5f55\u91cc\u6709 <code>args.json<\/code>\uff0c\u4e0b\u6b21\u63a8\u7406\u5b83\u4f1a\u81ea\u52a8\u56de\u8bfb\u6240\u6709\u8bad\u7ec3\u53c2\u6570\uff0c\u4e0d\u7528\u91cd\u590d\u6307\u5b9a <code>--model<\/code>\u2014\u2014\u65b9\u4fbf\u662f\u771f\u65b9\u4fbf\u3002\u53ef\u4ee3\u4ef7\u662f\u6392\u67e5\u95ee\u9898\u65f6\u94fe\u8def\u53d8\u957f\uff1a\u4f60\u4ee5\u4e3a\u4f20\u4e86\u67d0\u4e2a\u53c2\u6570\uff0c\u5b9e\u9645\u5b83\u4ece\u54ea\u4e00\u5c42\u8bfb\u8fdb\u53bb\u7684\uff0c\u5f97\u7ffb\u534a\u5929\u3002\u9047\u5230\u8981\u9b54\u6539\u8bad\u7ec3\u5faa\u73af\u3001\u81ea\u5b9a\u4e49 loss \u7684\u7814\u7a76\u573a\u666f\uff0c\u8fd9\u5c42\u58f3\u7684\u539a\u5ea6\u5c31\u4f1a\u4ece\u4fbf\u5229\u53d8\u6210\u963b\u788d\uff0c\u8fd9\u79cd\u60c5\u51b5\u4e0b\u6211\u4f1a\u76f4\u63a5\u529d\u4f60\u7ed5\u8fc7\u5b83\u53bb\u5199 Trainer\u3002<\/p>\n<p>\u8fd9\u91cc\u6211\u6709\u4e2a\u4e0d\u592a\u4e00\u6837\u7684\u770b\u6cd5\uff1a\u793e\u533a\u91cc\u603b\u6709\u4eba\u8bf4\u5b83\u300c\u5927\u800c\u5168\u7b49\u4e8e\u80f6\u6c34\u6846\u67b6\u300d\uff0c\u6211\u4e0d\u8fd9\u4e48\u8ba4\u4e3a\u3002\u5bf9\u5de5\u7a0b\u56e2\u961f\u6765\u8bf4\uff0c\u8bad\u7ec3\u3001\u5bfc\u51fa\u3001\u90e8\u7f72\u3001\u91cf\u5316\u7528\u540c\u4e00\u5957\u53c2\u6570\u4f53\u7cfb\u4e32\u8d77\u6765\uff0c\u8fd9\u4e2a\u4e00\u81f4\u6027\u7701\u4e0b\u7684\u4eba\u529b\u8fdc\u8d85\u62bd\u8c61\u7a0e\u2014\u2014\u4f60\u4e0d\u7528\u5728\u56db\u4e2a\u5de5\u5177\u4e4b\u95f4\u5bf9\u9f50\u6a21\u578b\u683c\u5f0f\u3002\u771f\u6b63\u88ab\u8fd9\u5c42\u58f3\u4f24\u5230\u7684\u53ea\u6709\u4e00\u7c7b\u4eba\uff1a\u9700\u8981\u770b\u6e05\u8bad\u7ec3\u5faa\u73af\u91cc\u6bcf\u4e00\u6b65\u7684\u7814\u7a76\u8005\u3002\u5de5\u5177\u6ca1\u6709\u9519\uff0c\u9519\u7684\u662f\u62ff\u5b83\u7684\u4eba\u4e0d\u5bf9\u3002<\/p>\n<h2 id=\"_3\">\u88c5\u8d77\u6765\u4e0d\u96be\uff0c\u96be\u7684\u662f\u522b\u8e29\u7248\u672c<\/h2>\n<p>\u5148\u8001\u5b9e\u4ea4\u4ee3\uff1a\u6211\u8fd9\u53f0\u673a\u5668\u662f MacBook\uff0c\u6ca1\u6709 N \u5361\uff0c\u8bad\u7ec3\u8fd9\u6b65\u6211\u6ca1\u6cd5\u5728\u672c\u673a\u771f\u8dd1\u3002\u6240\u4ee5\u4e0b\u9762\u8fd9\u4e9b\u6b65\u9aa4\u548c\u7248\u672c\u7ea6\u675f\uff0c\u662f\u6211\u5bf9\u7740\u5b98\u65b9 README \u548c\u4e2d\u6587\u6587\u6863\u9010\u6761\u6838\u8fc7\u7684\uff0c\u786c\u4ef6\u6570\u5b57\u5168\u90e8\u6ce8\u660e\u51fa\u5904\uff0c\u6ca1\u6709\u4e00\u4e2a\u662f\u4f30\u7684\u3002<\/p>\n<p>\u5b89\u88c5\u672c\u4f53\u5c31\u4e00\u884c\uff1a<\/p>\n<pre><code class=\"language-shell\">pip install ms-swift -U\n\n# \u6216\u8005\u7528 uv\uff0ctorch \u540e\u7aef\u81ea\u52a8\u5339\u914d\npip install uv\nuv pip install ms-swift -U --torch-backend=auto\n<\/code><\/pre>\n<p>\u8fd0\u884c\u73af\u5883\u5b98\u65b9\u7ed9\u4e86\u660e\u786e\u8868\u683c\uff1aPython <code>&gt;=3.10<\/code>\uff08\u63a8\u8350 3.12\uff09\uff0ctorch <code>&gt;=2.0<\/code>\uff08\u63a8\u8350 2.8.0\uff09\uff0cCUDA \u63a8\u8350 12.8 \u6216 13.0\uff0ctransformers <code>&gt;=4.33<\/code>\u3002\u8fd9\u5f20\u8868\u662f\u6211\u89c1\u8fc7\u5199\u5f97\u6700\u5168\u7684\u56fd\u4ea7\u8bad\u7ec3\u6846\u67b6\u6587\u6863\u4e4b\u4e00\uff0c\u8fde flash-attn \u548c vllm \u7684\u63a8\u8350\u7248\u672c\u90fd\u6807\u4e86\u3002<\/p>\n<p>\u786c\u4ef6\u95e8\u69db\uff0c\u5b98\u65b9\u53e3\u5f84\u5206\u4e09\u6863\uff1a<\/p>\n<ul>\n<li><strong>\u6700\u4f4e\u53ef\u7528<\/strong>\uff1a\u5355\u5361 24GB \u6d88\u8d39\u5361\uff083090\/4090 \u7ea7\u522b\uff09\u3002\u5b98\u65b9 Quick Start \u539f\u6587\u5199\u660e\uff1aQwen3-4B \u505a LoRA \u5fae\u8c03\u5360 13GB \u663e\u5b58\uff0c\u5355\u5f20 3090 \u5341\u5206\u949f\u8dd1\u5b8c\u2014\u2014\u8fd9\u662f\u5b98\u65b9\u539f\u8bdd\uff0c\u4e0d\u662f\u6211\u7684\u4f30\u7b97\u3002<\/li>\n<li><strong>\u91cf\u5316\u8bad\u7ec3<\/strong>\uff1a7B \u6a21\u578b QLoRA \u5b98\u65b9\u6807 9GB\uff0c\u4e00\u5f20 3060 12G \u6216\u8005 4070 \u90fd\u80fd\u585e\u4e0b\u3002<\/li>\n<li><strong>\u5efa\u8bae\u914d\u7f6e<\/strong>\uff1a\u8ba4\u771f\u505a\u5bf9\u9f50\u8bad\u7ec3\uff08DPO\/GRPO\uff09\u60f3\u8981\u4e0d\u618b\u5c48\uff0c\u53cc\u5361\u8d77\u6b65\uff1bREADME \u7684 GRPO \u793a\u4f8b\u91cc\u6709 4 \u5f20 80G A100 \u8dd1 72B \u6df7\u5408\u6a21\u5f0f\u7684\u5148\u4f8b\uff0c\u91cf\u529b\u800c\u884c\u3002<\/li>\n<\/ul>\n<p>Mac \u7528\u6237\u522b\u7070\u5fc3\uff0c\u5b83\u652f\u6301 Apple MPS \u548c\u7eaf CPU\u2014\u2014\u4f46\u90a3\u662f\u7ed9\u4f60\u8dd1\u63a8\u7406 demo \u7528\u7684\uff0c\u8bad\u7ec3\u522b\u60f3\u3002\u53e6\u5916\u63d0\u9192\u4e00\u53e5\u78c1\u76d8\uff1atorch \u4e00\u4e2a\u5305\u5c31\u4e24\u4e2a\u591a GB\uff0c\u52a0\u4e0a Qwen3-4B \u8fd9\u79cd 4B \u6a21\u578b\u7684\u6743\u91cd\u672c\u8eab\u5c31\u6709 8GB \u4e0a\u4e0b\uff0c\u7f13\u5b58\u76ee\u5f55\u4f1a\u8ddf\u7740\u4e0b\u8f7d\u91cf\u81a8\u80c0\uff0c\u78c1\u76d8\u9884\u7559 30GB \u8d77\u6b65\u6bd4\u8f83\u7a33\uff0c\u7f13\u5b58\u4f4d\u7f6e\u63d0\u524d\u7528\u73af\u5883\u53d8\u91cf\u6307\u5230\u5927\u76d8\u4e0a\uff0c\u522b\u5806\u5728\u7cfb\u7edf\u76d8\u3002\u5b98\u65b9\u6ca1\u7ed9\u6574\u673a\u914d\u7f6e\u603b\u8868\uff0c\u54ea\u4e9b\u5361\u786e\u5b9a\u80fd\u7528\uff1aNVIDIA \u5168\u7cfb\u8d70 CUDA\uff0cAMD MI300 \u7cfb\u8d70 ROCm\uff0c\u6607\u817e NPU\u3001\u6469\u5c14\u7ebf\u7a0b\u3001\u6d77\u5149\u8fd9\u4e9b\u56fd\u4ea7\u786c\u4ef6\u4e5f\u5217\u5728\u652f\u6301\u5217\u8868\u91cc\uff0c\u8fd9\u4e2a\u8986\u76d6\u9762\u5728\u5f00\u6e90\u8bad\u7ec3\u6846\u67b6\u91cc\u76f8\u5f53\u5c11\u89c1\u3002<\/p>\n<h2 id=\"_4\">\u7167\u7740\u5b98\u65b9\u90a3\u6761\u8def\u8d70\u4e00\u904d<\/h2>\n<p>\u5b98\u65b9 Quick Start \u7684\u4f8b\u5b50\u9009\u5f97\u5f88\u806a\u660e\uff1a\u62ff Qwen3-4B-Instruct-2507 \u505a\u300c\u81ea\u6211\u8ba4\u77e5\u300d\u5fae\u8c03\uff0c\u5341\u6765\u6761\u6570\u636e\u5c31\u80fd\u8ba9\u6a21\u578b\u6539\u53e3\u8bf4\u81ea\u5df1\u662f\u8c01\uff0c\u53cd\u9988\u76f4\u89c2\uff0c\u6700\u9002\u5408\u7b2c\u4e00\u6b21\u4e0a\u624b\u3002\u4e0b\u9762\u8fd9\u5957\u6d41\u7a0b\u6309\u5b98\u65b9\u6587\u6863\u539f\u6837\u8d70\uff0c\u6211\u5728\u6ca1\u6709\u5361\u7684\u673a\u5668\u4e0a\u628a\u6bcf\u6761\u547d\u4ee4\u7684\u53c2\u6570\u90fd\u5bf9\u7740\u6587\u6863\u6838\u4e86\u4e00\u904d\uff0c\u9884\u671f\u8f93\u51fa\u4ee5\u6587\u6863\u63cf\u8ff0\u4e3a\u51c6\u3002<\/p>\n<p>\u5148\u51c6\u5907\u4e00\u4e2a\u81ea\u6211\u8ba4\u77e5\u6570\u636e\u96c6\uff0c\u683c\u5f0f\u5c31\u662f\u6807\u51c6 messages \u5bf9\u8bdd\uff0c\u5b58\u6210 <code>my.jsonl<\/code>\uff1a<\/p>\n<pre><code class=\"language-json\">{&quot;messages&quot;: [{&quot;role&quot;: &quot;system&quot;, &quot;content&quot;: &quot;\u4f60\u662f\u670b\u53cb\u516c\u53f8\u7684\u552e\u540e\u52a9\u624b&quot;}, {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;\u4f60\u662f\u8c01\uff1f&quot;}, {&quot;role&quot;: &quot;assistant&quot;, &quot;content&quot;: &quot;\u6211\u662f\u552e\u540e\u5c0f\u52a9\uff0c\u4e13\u7ba1\u8f74\u627f\u548c\u6da6\u6ed1\u7684\u884c\u8bdd\u95ee\u9898\u3002&quot;}]}\n<\/code><\/pre>\n<p>\u7136\u540e\u4e00\u6761\u547d\u4ee4\u5f00\u8bad\uff1a<\/p>\n<pre><code class=\"language-shell\">CUDA_VISIBLE_DEVICES=0 \\\nswift sft \\\n    --model Qwen\/Qwen3-4B-Instruct-2507 \\\n    --tuner_type lora \\\n    --dataset \/path\/to\/my.jsonl 'swift\/self-cognition#500' \\\n    --torch_dtype bfloat16 \\\n    --num_train_epochs 1 \\\n    --per_device_train_batch_size 1 \\\n    --learning_rate 1e-4 \\\n    --lora_rank 8 \\\n    --lora_alpha 32 \\\n    --target_modules all-linear \\\n    --gradient_accumulation_steps 16 \\\n    --max_length 2048 \\\n    --model_author swift \\\n    --model_name swift-robot \\\n    --output_dir output\n<\/code><\/pre>\n<p>\u51e0\u4e2a\u53c2\u6570\u8bf4\u4e00\u4e0b\u6211\u7684\u7406\u89e3\uff1a<code>swift\/self-cognition#500<\/code> \u662f\u5b98\u65b9\u81ea\u5e26\u7684\u81ea\u6211\u8ba4\u77e5\u6570\u636e\u96c6\uff0c<code>#500<\/code> \u8868\u793a\u91c7\u6837 500 \u6761\uff1b<code>--model_author<\/code> \u548c <code>--model_name<\/code> \u53ea\u5728\u8fd9\u4e2a\u6570\u636e\u96c6\u5b58\u5728\u65f6\u751f\u6548\uff0c\u5206\u522b\u51b3\u5b9a\u6a21\u578b\u300c\u62a5\u6237\u53e3\u300d\u65f6\u7684\u4f5c\u8005\u540d\u548c\u81ea\u5df1\u7684\u540d\u5b57\u3002\u663e\u5b58\u5360\u7528\u5b98\u65b9\u6807\u6ce8 13GB\u3002<\/p>\n<p>\u8bad\u5b8c\u63a8\u7406\uff0c<code>--adapters<\/code> \u586b\u8bad\u7ec3\u8f93\u51fa\u76ee\u5f55\u91cc\u6700\u540e\u7684\u90a3\u4e2a checkpoint \u6587\u4ef6\u5939\uff0c\u4e0d\u7528\u518d\u4f20 <code>--model<\/code>\uff0c\u5b83\u4f1a\u4ece checkpoint \u91cc\u7684 <code>args.json<\/code> \u81ea\u52a8\u8bfb\uff1a<\/p>\n<pre><code class=\"language-shell\">CUDA_VISIBLE_DEVICES=0 \\\nswift infer \\\n    --adapters output\/vx-xxx\/checkpoint-xxx \\\n    --stream true \\\n    --temperature 0\n<\/code><\/pre>\n<p>\u6309\u5b98\u65b9\u6587\u6863\u7684\u8bf4\u6cd5\uff0c\u8fd9\u65f6\u5019\u95ee\u5b83\u300c\u4f60\u662f\u8c01\u300d\uff0c\u5b83\u4f1a\u7528\u4f60\u6570\u636e\u96c6\u91cc\u5199\u7684\u8eab\u4efd\u56de\u7b54\uff0c\u800c\u4e0d\u662f Qwen \u7684\u81ea\u6211\u4ecb\u7ecd\u3002\u60f3\u8ba9\u5b83\u53d8\u6210\u4e00\u4e2a\u72ec\u7acb\u6a21\u578b\u800c\u4e0d\u662f\u6302\u5728\u57fa\u5ea7\u4e0a\u7684 LoRA\uff0c\u52a0 <code>--merge_lora true<\/code> \u5408\u5e76\u6743\u91cd\uff0c\u518d\u8d70 <code>swift export --push_to_hub true<\/code> \u5c31\u80fd\u63a8\u56de\u9b54\u642d\u3002\u6574\u6761\u8def\u6ca1\u6709\u4e00\u6b65\u9700\u8981\u5199 Python\uff0c\u8fd9\u662f\u6211\u770b\u5b8c\u6700\u76f4\u89c2\u7684\u611f\u53d7\u3002<\/p>\n<h2 id=\"_5\">\u5751\u90fd\u5728\u6df1\u6c34\u533a<\/h2>\n<p>\u5355\u5361 LoRA \u8fd9\u6761\u4e3b\u5e72\u9053\u94fa\u5f97\u5f88\u5e73\uff0c\u5751\u96c6\u4e2d\u5728\u4e24\u8fb9\uff1aQwen3 \u7684\u601d\u8003\u6807\u7b7e\uff0c\u548c\u5206\u5e03\u5f0f\u5f3a\u5316\u5b66\u4e60\u3002<\/p>\n<p><strong>\u7b2c\u4e00\u5751\uff1aQwen3 \u7684 think \u6807\u7b7e\u3002<\/strong> issue <a href=\"https:\/\/github.com\/modelscope\/ms-swift\/issues\/5507\">#5507<\/a> \u6709 33 \u6761\u8ba8\u8bba\uff0c\u63d0\u95ee\u8005\u95ee\u5f97\u5f88\u5178\u578b\uff1aSFT \u6570\u636e\u96c6\u91cc\u6ca1\u5199 <code>&lt;think&gt;<\/code> \u6807\u7b7e\uff0c\u8bad\u7ec3\u540e\u6a21\u578b\u7684\u884c\u4e3a\u662f\u4ec0\u4e48\uff1f\u8fd9\u80cc\u540e\u662f\u4e2a\u771f\u5b9e\u75db\u70b9\u2014\u2014Qwen3 \u8fd9\u7c7b\u6df7\u5408\u63a8\u7406\u6a21\u578b\uff0c\u8bad\u7ec3\u6570\u636e\u91cc\u601d\u8003\u6bb5\u7684\u6709\u65e0\u4f1a\u76f4\u63a5\u5f71\u54cd\u8f93\u51fa\u98ce\u683c\uff0c\u6570\u636e\u4e0d\u5e72\u51c0\u7684\u8bdd\u8bad\u51fa\u6765\u7684\u6a21\u578b\u4e00\u4f1a\u513f\u601d\u8003\u4e00\u4f1a\u513f\u4e0d\u601d\u8003\u3002\u8fd9\u4e2a issue \u4e0b\u9762\u793e\u533a\u7ed9\u4e86\u4e0d\u5c11\u5b9e\u8df5\u5efa\u8bae\uff0c\u8bad\u7ec3\u524d\u52a1\u5fc5\u628a\u6570\u636e\u91cc\u601d\u8003\u6807\u7b7e\u7684\u5206\u5e03\u634b\u6e05\u695a\u3002<\/p>\n<p><strong>\u7b2c\u4e8c\u5751\uff1aGRPO \u8bad\u7ec3\u8d85\u65f6\u3002<\/strong> issue <a href=\"https:\/\/github.com\/modelscope\/ms-swift\/issues\/8053\">#8053<\/a>\uff08bug \u6807\u7b7e\uff09\u63cf\u8ff0\u7684\u573a\u666f\u662f rollout server \u660e\u660e\u62c9\u8d77\u6765\u4e86\uff0c\u8bad\u7ec3\u7aef\u5374\u62a5\u8fde\u4e0d\u4e0a vLLM \u7684 51216 \u7aef\u53e3\u7136\u540e\u8d85\u65f6\uff0c21 \u6761\u8ba8\u8bba\u91cc\u5168\u662f\u540c\u7c7b\u53d7\u5bb3\u8005\u3002\u591a\u673a RL \u8bad\u7ec3\u7684\u7f51\u7edc\u914d\u7f6e\u6bd4\u60f3\u8c61\u7684\u8106\uff0c\u9632\u706b\u5899\u3001\u7f51\u5361\u7ed1\u5b9a\u3001\u7aef\u53e3\u66b4\u9732\u4efb\u4f55\u4e00\u4e2a\u73af\u8282\u90fd\u5f97\u81ea\u5df1\u68c0\u67e5\u3002<\/p>\n<p><strong>\u7b2c\u4e09\u5751\uff1a\u5927\u6a21\u578b\u6d41\u6c34\u7ebf\u5e76\u884c\u5361\u6b7b\u3002<\/strong> issue <a href=\"https:\/\/github.com\/modelscope\/ms-swift\/issues\/9451\">#9451<\/a> \u66f4\u786c\u6838\uff1aGLM 5.1 \u7684 MoE \u6a21\u578b\u4e0a 128 \u5361\u505a\u6d41\u6c34\u7ebf\u5e76\u884c LoRA\uff0c\u8bad\u7ec3\u5361\u6b7b\u5728\u7b2c\u4e00\u6b65\uff0c\u6839\u56e0\u6307\u5411 Megatron \u7684 <code>batch_p2p_comm<\/code> \u914d\u7f6e\u4e0e NCCL \u61d2\u521d\u59cb\u5316\u7684\u76f8\u4e92\u4f5c\u7528\u3002\u8fd9\u4e2a issue \u5230\u6211\u67e5\u7684\u65f6\u5019\u8fd8\u6302\u7740 bug \u6807\u7b7e\u2014\u2014\u8bf4\u660e Megatron \u8fd9\u6761\u8def\u867d\u7136\u901a\u7740\uff0c\u4f46\u8fb9\u7f18\u7ec4\u5408\u8fd8\u662f\u4f1a\u6709\u6ca1\u586b\u5e73\u7684\u6c9f\u3002<\/p>\n<p><strong>\u7b2c\u56db\u5751\uff1a\u540c\u4e00\u6a21\u578b\u4e0d\u540c\u540e\u7aef\u5411\u91cf\u4e0d\u4e00\u81f4\u3002<\/strong> issue <a href=\"https:\/\/github.com\/modelscope\/ms-swift\/issues\/6364\">#6364<\/a> \u6709 21 \u6761\u8ba8\u8bba\uff1a\u540c\u4e00\u4e2a embedding \u6a21\u578b\uff0cvLLM \u90e8\u7f72\u548c swift \u81ea\u5e26\u90e8\u7f72\u8fd4\u56de\u7684\u5411\u91cf\u5bf9\u4e0d\u4e0a\u3002\u505a\u68c0\u7d22\u7684\u540c\u5b66\u8981\u7279\u522b\u5f53\u5fc3\uff0c\u8bad\u7ec3\u548c\u90e8\u7f72\u6362\u540e\u7aef\u4e4b\u524d\u5148\u5bf9\u4e00\u904d\u5411\u91cf\u3002<\/p>\n<p>issue \u533a\u6574\u4f53\u5f00\u7740 600 \u4e2a open issue\uff0c\u770b\u7740\u5413\u4eba\uff0c\u4f46\u6211\u7ffb\u4e86\u7ffb\uff0c\u54cd\u5e94\u901f\u5ea6\u4e0d\u6162\uff0c\u4e2d\u6587\u63d0\u95ee\u4e5f\u63a5\u5f85\u3002\u9ad8\u9891\u8ba8\u8bba\u91cc\u8fd8\u6709\u4e00\u4e2a\u503c\u5f97\u5355\u72ec\u8bf4\u7684\uff1aissue <a href=\"https:\/\/github.com\/modelscope\/ms-swift\/issues\/4030\">#4030<\/a>\uff0c\u6807\u9898\u5c31\u662f\u8bad\u7ec3 Qwen3 \u7cfb\u5217\u7684\u6700\u4f73\u5b9e\u8df5\uff0c\u6302\u4e86 146 \u6761\u8ba8\u8bba\uff0c\u4ece MoE \u7ed3\u6784\u7684\u5b66\u4e60\u7387\u8bbe\u5230 MoE \u8bad\u7ec3\u7684\u6570\u636e\u914d\u6bd4\uff0c\u582a\u79f0\u534a\u4efd\u516c\u5f00\u7684\u8c03\u53c2\u7b14\u8bb0\u3002\u8fd9\u79cd\u300c\u5b98\u65b9\u51fa\u9898\u3001\u793e\u533a\u8865\u7b54\u6848\u300d\u7684\u957f\u8d34\u662f ms-swift issue \u533a\u7684\u7279\u8272\uff0c\u52a8\u624b\u524d\u5148\u641c\u4e00\u641c\u540c\u7c7b\u6a21\u578b\u6709\u6ca1\u6709\u73b0\u6210\u8d34\uff0c\u80fd\u7701\u6389\u4e0d\u5c11\u81ea\u5df1\u8bd5\u9519\u7684\u65f6\u95f4\u3002<\/p>\n<p>\u793e\u533a\u4e4b\u5916\u7684\u5b9e\u6218\u6587\u7ae0\u91cc\uff0cCSDN \u4e0a\u90a3\u7bc7<a href=\"https:\/\/blog.csdn.net\/weixin_33481022\/article\/details\/158751654\">\u4ece\u5b89\u88c5\u5230\u63a8\u9001\u7684\u5b8c\u6574\u6d41\u7a0b\u5b9e\u6d4b<\/a>\u603b\u7ed3\u7684\u51e0\u6761\u8ddf issue \u533a\u80fd\u5bf9\u4e0a\uff1aOOM \u4e86\u6309\u300c\u5f00 gradient checkpointing \u2192 \u8f6c bf16 \u2192 \u6362 QLoRA \u2192 batch \u964d\u5230 1\u300d\u7684\u987a\u5e8f\u5904\u7406\uff1bloss \u4e0d\u964d\u5148\u6000\u7591\u6570\u636e\u683c\u5f0f\u548c\u6a21\u677f\u5339\u4e0d\u5339\u914d\uff0c\u800c\u4e0d\u662f\u731b\u8c03\u5b66\u4e60\u7387\uff1b\u5fae\u8c03\u540e\u6a21\u578b\u53d8\u7b28\u662f\u707e\u96be\u6027\u9057\u5fd8\uff0c\u4e0d\u662f\u4f60\u8bad\u574f\u4e86\u3002\u53e6\u4e00\u7bc7\u6846\u67b6\u6a2a\u8bc4\u7c7b\u7684<a href=\"https:\/\/devpress.csdn.net\/awstech\/6a72e63610ee7a33f2965e3c.html\">\u516d\u6b3e\u5fae\u8c03\u5de5\u5177\u5bf9\u6bd4<\/a>\u91cc\u7ed9 ms-swift \u7684\u5b9a\u4f4d\u4e5f\u8ddf\u6211\u7684\u5224\u65ad\u4e00\u81f4\uff1a\u80fd\u529b\u8986\u76d6\u6700\u5168\u3001\u504f\u547d\u4ee4\u884c\u3001\u7ed1\u5b9a\u9b54\u642d\u751f\u6001\u3002\u516c\u5f00\u7684\u82f1\u6587\u793e\u533a\u8ba8\u8bba\u4e0d\u7b97\u591a\uff0cReddit \u548c Hacker News \u4e0a\u5c11\u6709\u6210\u89c4\u6a21\u7684\u8ba8\u8bba\u5e16\uff0c\u8fd9\u8ddf\u5b83\u7528\u6237\u96c6\u4e2d\u5728\u4e2d\u6587\u5708\u6709\u5173\u3002<\/p>\n<h2 id=\"_6\">\u6211\u6700\u540e\u7684\u5224\u65ad<\/h2>\n<p>\u503c\u4e0d\u503c\u5f97\u4e0a\u624b\uff0c\u6211\u7ed9\u660e\u786e\u7b54\u6848\uff1a\u5982\u679c\u4f60\u7684\u8bad\u7ec3\u5bf9\u8c61\u662f Qwen\u3001GLM\u3001InternLM \u8fd9\u6279\u56fd\u4ea7\u6a21\u578b\uff0c\u6216\u8005\u4efb\u52a1\u6d89\u53ca\u591a\u6a21\u6001\uff0cms-swift \u662f\u5f53\u4e0b\u8ddf\u8fdb\u6700\u5feb\u3001\u94fe\u8def\u6700\u5168\u7684\u9009\u62e9\uff0c\u6ca1\u6709\u4e4b\u4e00\u3002\u65b0\u6a21\u578b Day-0 \u652f\u6301\u8fd9\u4ef6\u4e8b\uff0c\u522b\u4eba\u662f\u771f\u505a\u4e0d\u5230\u8fd9\u4e2a\u901f\u5ea6\u2014\u2014\u5b83\u662f\u9b54\u642d\u5b98\u65b9\u7684\u57fa\u5efa\uff0c\u6a21\u578b\u4e00\u53d1\u5e03\u5b83\u5c31\u6709\u52a8\u529b\u5f53\u5929\u9002\u914d\u3002\u547d\u4ee4\u884c\u5168\u94fe\u8def\u7684\u5f62\u6001\u4e5f\u9002\u5408\u8981\u5728\u8bad\u7ec3\u548c\u90e8\u7f72\u4e4b\u95f4\u6765\u56de\u642c\u8fd0\u6a21\u578b\u7684\u5de5\u7a0b\u56e2\u961f\u3002<\/p>\n<p>\u5b83\u4e0d\u9002\u5408\u8c01\uff1f\u4e24\u79cd\u4eba\u3002\u7b2c\u4e00\u79cd\u662f\u8981\u96f6\u4ee3\u7801\u4f53\u9a8c\u7684\u65b0\u624b\uff0c\u76f4\u63a5\u53bb LLaMA-Factory\uff0c\u5b83\u7684 Web-UI \u751f\u6001\u548c\u793e\u533a\u6559\u7a0b\u539a\u5ea6\u662f\u540c\u8d5b\u9053\u6700\u539a\u7684\uff0cAWS SageMaker \u8fd9\u4e9b\u5e73\u53f0\u7684\u96c6\u6210\u4e5f\u66f4\u6210\u719f\u2014\u2014\u4e24\u5bb6\u90fd\u662f Apache-2.0\uff0c\u6362\u7740\u7528\u6ca1\u6709\u6c89\u6ca1\u6210\u672c\uff0c\u6211\u7684\u5efa\u8bae\u662f\u65b0\u624b\u4ece LLaMA-Factory \u5165\u95e8\u3001\u4efb\u52a1\u590d\u6742\u4e86\u518d\u8fc1\u5230 ms-swift\uff0c\u6a21\u677f\u548c\u6570\u636e\u683c\u5f0f\u7684\u601d\u8def\u662f\u76f8\u901a\u7684\u3002\u7b2c\u4e8c\u79cd\u662f\u5355\u5361\u62a0\u663e\u5b58\u5230\u6781\u81f4\u7684\u73a9\u5bb6\uff0cUnsloth \u624b\u5199\u7684 Triton kernel \u5728\u7701\u663e\u5b58\u548c\u63d0\u901f\u4e0a\u4ecd\u662f\u72ec\u4e00\u6863\uff0c7B \u5361\u5728 12G \u91cc\u7684\u573a\u666f\u5b83\u66f4\u5bf9\u8def\uff1b\u4ee3\u4ef7\u662f\u6a21\u578b\u8986\u76d6\u7a84\u3001\u591a\u5361\u80fd\u529b\u5f31\u3002\u5224\u65ad\u7ef4\u5ea6\u5c31\u4e09\u6761\uff1a\u6a21\u578b\u5728\u4e0d\u5728\u5b83\u7684\u652f\u6301\u5217\u8868\u3001\u4f60\u6709\u51e0\u5f20\u5361\u3001\u8981\u4e0d\u8981\u591a\u6a21\u6001\u2014\u2014\u4e09\u6761\u91cc\u547d\u4e2d\u524d\u4e24\u6761\u91cc\u7684\u540e\u8005\uff0c\u5c31\u9009 ms-swift\u3002<\/p>\n<p>\u603b\u800c\u8a00\u4e4b\uff0c\u8fd9\u4e2a\u9879\u76ee\u6211\u5df2\u7ecf\u52a0\u8fdb\u4e86\u81ea\u5df1\u7684\u5e38\u7528\u6e05\u5355\uff0c\u4e0b\u4e00\u5757\u6361\u5230\u7684\u5361\u5230\u624b\uff0c\u7b2c\u4e00\u4e2a\u8981\u8dd1\u7684\u5c31\u662f\u4e0a\u9762\u90a3\u4e2a\u81ea\u6211\u8ba4\u77e5\u7684\u4f8b\u5b50\u3002\u987a\u624b\u62c9\u4e86\u4e0b\u5b83\u73b0\u5728\u7684\u6570\u636e\uff1a15,986 \u4e2a star\uff0c\u6700\u65b0\u7248\u672c v4.6.0 \u662f 10 \u6708 13 \u53f7\u53d1\u7684\uff0c\u6700\u8fd1\u56db\u5468\u8fd8\u6709 75 \u6b21\u63d0\u4ea4\uff0c\u5341\u6708\u5e95\u8fd8\u5728\u7ed9\u65b0\u6a21\u578b\u505a\u5f53\u5929\u9002\u914d\u2014\u2014\u4e00\u4e2a 2023 \u5e74 8 \u6708\u7acb\u9879\u7684\u9879\u76ee\u8fd8\u80fd\u4fdd\u6301\u8fd9\u4e2a\u8282\u594f\uff0c\u540e\u52b2\u662f\u771f\u7684\u8db3\u3002\u6700\u540e\u8bf4\u53e5\u5b9e\u5728\u7684\uff0c\u5fae\u8c03\u8fd9\u4ef6\u4e8b\u7684\u95e8\u69db\u8fd9\u51e0\u5e74\u662f\u88ab\u8fd9\u7c7b\u6846\u67b6\u4e00\u5bf8\u5bf8\u538b\u4e0b\u6765\u7684\uff0c\u4e09\u5e74\u524d\u6211\u5199 Trainer \u811a\u672c\u8c03\u6a21\u677f\u7684\u90a3\u4e00\u5468\uff0c\u653e\u5230\u4eca\u5929\u5c31\u662f\u4e00\u6761\u547d\u4ee4\u52a0\u5341\u5206\u949f\u7b49\u6743\u91cd\u4e0b\u8f7d\u3002\u4f60\u8981\u662f\u4e5f\u6709\u51e0\u767e\u6761\u6512\u7740\u6ca1\u7528\u7684\u5bf9\u8bdd\u6570\u636e\uff0c\u771f\u7684\u53ef\u4ee5\u8bd5\u8bd5\u8ba9\u5b83\u8fdb\u5230\u6a21\u578b\u91cc\uff0c\u800c\u4e0d\u662f\u6c38\u8fdc\u8eba\u5728 prompt \u91cc\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>ms-swift \u662f\u9b54\u642d\u5b98\u65b9\u7684\u5fae\u8c03\u6846\u67b6\uff0c\u6211\u5bf9\u7740\u6587\u6863\u548c 600 \u6761 issue \u628a\u5b83\u7ffb\u4e86\u4e00\u904d\uff1a600 \u591a\u4e2a\u6a21\u578b\u4e00\u6761\u547d\u4ee4\u5fae\u8c03\uff0c\u5355\u5361 13GB \u5c31\u80fd\u8dd1 Qwen3\uff0c\u65b0\u6a21\u578b\u5f53\u5929\u9002\u914d\u3002\u5751\u96c6\u4e2d\u5728\u601d\u8003\u6807\u7b7e\u548c\u5206\u5e03\u5f0f RL\uff0c\u7ed3\u8bba\u662f\u56fd\u4ea7\u6a21\u578b\u5fae\u8c03\u9996\u9009\u3002<\/p>\n","protected":false},"author":1,"featured_media":10021523,"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-10021524","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\/10\/ms-swift-cover.png","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021524","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=10021524"}],"version-history":[{"count":1,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021524\/revisions"}],"predecessor-version":[{"id":10021525,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021524\/revisions\/10021525"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/media\/10021523"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10021524"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10021524"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10021524"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}