{"id":116225,"date":"2026-10-20T23:08:50","date_gmt":"2026-10-20T15:08:50","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/116225.html"},"modified":"2026-10-20T23:08:50","modified_gmt":"2026-10-20T15:08:50","slug":"%e5%88%86%e5%b8%83%e5%bc%8f%e8%ae%ad%e7%bb%83%e6%a1%86%e6%9e%b6_deepspeed_zero%e8%af%a6%e8%a7%a3","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/116225.html","title":{"rendered":"\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6_DeepSpeed_ZeRO\u8be6\u89e3"},"content":{"rendered":"<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261020150846-6ad783fe4e804.png\" alt=\"image.png\" \/><\/p>\n<\/p>\n<h4>\u6587\u7ae0\u76ee\u5f55<\/h4>\n<ul>\n<li>\n<ul>\n<li>1. \u5148\u628a\u95ee\u9898\u8bb2\u6e05\u695a&#xff1a;\u4e3a\u4ec0\u4e48\u201c\u80fd\u63a8\u7406\u201d\u4e0d\u7b49\u4e8e\u201c\u80fd\u8bad\u7ec3\u201d<\/li>\n<li>2. \u8bad\u7ec3\u663e\u5b58\u8d26\u672c&#xff1a;\u5230\u5e95\u54ea\u4e9b\u4e1c\u897f\u5360\u663e\u5b58<\/li>\n<li>3. \u666e\u901a\u6570\u636e\u5e76\u884c\u4e3a\u4ec0\u4e48\u6d6a\u8d39<\/li>\n<li>4. ZeRO-1&#xff1a;\u5148\u5207\u4f18\u5316\u5668\u72b6\u6001<\/li>\n<li>5. ZeRO-2&#xff1a;\u4f18\u5316\u5668\u72b6\u6001\u548c\u68af\u5ea6\u90fd\u5207<\/li>\n<li>6. ZeRO-3&#xff1a;\u8fde\u53c2\u6570\u4e5f\u5207<\/li>\n<li>7. \u7528\u4ee3\u7801\u4f30\u7b97 ZeRO \u9636\u6bb5\u7684\u53c2\u6570\u72b6\u6001\u8282\u7701<\/li>\n<li>8. DeepSpeed \u914d\u7f6e\u600e\u4e48\u8bfb<\/li>\n<li>9. ZeRO-3 \u914d\u7f6e\u548c\u51e0\u4e2a\u5173\u952e\u5f00\u5173<\/li>\n<li>10. Offload&#xff1a;\u628a\u72b6\u6001\u653e\u5230 CPU \u6216 NVMe<\/li>\n<li>11. \u5982\u4f55\u9009\u62e9 ZeRO \u9636\u6bb5<\/li>\n<li>12. \u5e38\u89c1\u95ee\u9898\u6392\u67e5<\/li>\n<li>13. \u5e38\u89c1\u8bef\u533a<\/li>\n<li>14. \u5c0f\u7ed3<\/li>\n<li>\u5927\u6a21\u578b\u89c6\u89d2<\/li>\n<li>\u4e0b\u4e00\u7bc7<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>\u6458\u8981&#xff1a;\u5927\u6a21\u578b\u8bad\u7ec3\u663e\u5b58\u8d35&#xff0c;\u4e0d\u53ea\u662f\u56e0\u4e3a\u6a21\u578b\u53c2\u6570\u591a&#xff0c;\u8fd8\u56e0\u4e3a\u8bad\u7ec3\u65f6\u8981\u4fdd\u5b58\u68af\u5ea6\u3001\u4f18\u5316\u5668\u72b6\u6001\u3001\u6fc0\u6d3b\u503c\u548c\u901a\u4fe1\u7f13\u51b2\u533a\u3002\u666e\u901a\u6570\u636e\u5e76\u884c\u4f1a\u8ba9\u6bcf\u5f20 GPU \u90fd\u4fdd\u5b58\u4e00\u4efd\u5b8c\u6574\u8bad\u7ec3\u72b6\u6001&#xff0c;\u663e\u5b58\u6d6a\u8d39\u4e25\u91cd\u3002DeepSpeed ZeRO \u7684\u6838\u5fc3\u601d\u60f3\u662f\u628a\u8fd9\u4e9b\u8bad\u7ec3\u72b6\u6001\u5206\u7247\u5230\u591a\u5f20 GPU \u4e0a&#xff1a;ZeRO-1 \u5207\u4f18\u5316\u5668\u72b6\u6001&#xff0c;ZeRO-2 \u518d\u5207\u68af\u5ea6&#xff0c;ZeRO-3 \u8fde\u53c2\u6570\u4e5f\u5207\u3002\u672c\u6587\u4f1a\u4ece\u663e\u5b58\u8d26\u672c\u5f00\u59cb\u8bb2\u6e05 ZeRO \u4e3a\u4ec0\u4e48\u6709\u6548\u3001\u6bcf\u4e2a\u9636\u6bb5\u7701\u4ec0\u4e48\u3001\u4ee3\u4ef7\u662f\u4ec0\u4e48&#xff0c;\u5e76\u7ed9\u51fa\u53ef\u8bfb\u7684\u914d\u7f6e\u793a\u4f8b\u548c\u6392\u67e5\u6e05\u5355\u3002<\/p>\n<ul>\n<li>\u524d\u7f6e\u77e5\u8bc6&#xff1a;\u4e13\u680f\u4e00\u5206\u5e03\u5f0f\u8bad\u7ec3\u3001 AdamW\u3001\u6df7\u5408\u7cbe\u5ea6&#xff1b;\u4e13\u680f\u4e8c\u9884\u8bad\u7ec3\u6d41\u7a0b<\/li>\n<li>\u9605\u8bfb\u65f6\u95f4&#xff1a;50-65 \u5206\u949f<\/li>\n<li>\u4ee3\u7801\u73af\u5883&#xff1a;Python 3.10&#043;&#xff0c;\u793a\u4f8b\u53ea\u4f9d\u8d56\u6807\u51c6\u5e93&#xff1b;\u914d\u7f6e\u793a\u4f8b\u4f7f\u7528 JSON<\/li>\n<li>\u8bfb\u5b8c\u76ee\u6807&#xff1a;\u80fd\u4f30\u7b97\u8bad\u7ec3\u663e\u5b58\u7684\u4e3b\u8981\u7ec4\u6210&#xff0c;\u80fd\u89e3\u91ca ZeRO-1\/2\/3 \u7684\u5dee\u522b&#xff0c;\u80fd\u6839\u636e\u6a21\u578b\u89c4\u6a21\u548c\u786c\u4ef6\u9009\u62e9 ZeRO \u9636\u6bb5<\/li>\n<\/ul>\n<hr \/>\n<h3>1. \u5148\u628a\u95ee\u9898\u8bb2\u6e05\u695a&#xff1a;\u4e3a\u4ec0\u4e48\u201c\u80fd\u63a8\u7406\u201d\u4e0d\u7b49\u4e8e\u201c\u80fd\u8bad\u7ec3\u201d<\/h3>\n<p>\u5f88\u591a\u5f00\u53d1\u8005\u7b2c\u4e00\u6b21\u63a5\u89e6\u5927\u6a21\u578b\u8bad\u7ec3\u65f6&#xff0c;\u4f1a\u8fd9\u6837\u4f30\u7b97&#xff1a;7B \u6a21\u578b\u7528 BF16 \u6743\u91cd\u5b58\u50a8&#xff0c;\u6bcf\u4e2a\u53c2\u6570 2 \u5b57\u8282&#xff0c;\u6240\u4ee5\u6743\u91cd\u5927\u7ea6 14GB\u3002\u90a3\u4e00\u5f20 24GB \u6216 48GB \u7684\u5361\u662f\u4e0d\u662f\u5c31\u80fd\u8bad\u7ec3&#xff1f; <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261020150848-6ad7840038f0d.png\" alt=\"image.png\" \/><\/p>\n<p>\u901a\u5e38\u4e0d\u80fd\u3002\u56e0\u4e3a\u63a8\u7406\u4e3b\u8981\u4fdd\u5b58\u6743\u91cd\u548c KV Cache&#xff0c;\u800c\u8bad\u7ec3\u8fd8\u8981\u4fdd\u5b58\u68af\u5ea6\u3001\u4f18\u5316\u5668\u72b6\u6001\u548c\u524d\u5411\u4f20\u64ad\u4e2d\u95f4\u6fc0\u6d3b\u3002\u5c24\u5176\u662f AdamW \u8fd9\u7c7b\u4f18\u5316\u5668&#xff0c;\u4f1a\u4e3a\u6bcf\u4e2a\u53c2\u6570\u4fdd\u5b58\u989d\u5916\u72b6\u6001\u3002\u8bad\u7ec3\u663e\u5b58\u5e38\u5e38\u8fdc\u5927\u4e8e\u6a21\u578b\u6743\u91cd\u672c\u8eab\u3002<\/p>\n<p>\u666e\u901a\u6570\u636e\u5e76\u884c&#xff08;Data Parallel, DP&#xff09;\u8fd8\u6709\u4e00\u4e2a\u95ee\u9898&#xff1a;\u6bcf\u5f20 GPU \u90fd\u4fdd\u5b58\u5b8c\u6574\u6a21\u578b\u3001\u5b8c\u6574\u68af\u5ea6\u548c\u5b8c\u6574\u4f18\u5316\u5668\u72b6\u6001\u3002\u589e\u52a0 GPU \u6570\u91cf\u53ef\u4ee5\u63d0\u9ad8\u541e\u5410&#xff0c;\u4f46\u4e0d\u4e00\u5b9a\u964d\u4f4e\u5355\u5361\u663e\u5b58\u3002<\/p>\n<p>ZeRO&#xff08;Zero Redundancy Optimizer&#xff09;\u8981\u89e3\u51b3\u7684\u5c31\u662f\u8fd9\u4e2a\u201c\u91cd\u590d\u5b58\u50a8\u201d\u95ee\u9898\u3002\u5b83\u7684\u6838\u5fc3\u601d\u8def\u5f88\u6734\u7d20&#xff1a;\u65e2\u7136\u591a\u5f20 GPU \u4e00\u8d77\u8bad\u7ec3&#xff0c;\u5c31\u4e0d\u8981\u8ba9\u6bcf\u5f20\u5361\u90fd\u4fdd\u5b58\u5b8c\u6574\u8bad\u7ec3\u72b6\u6001&#xff0c;\u628a\u72b6\u6001\u62c6\u5f00&#xff0c;\u6bcf\u5f20\u5361\u53ea\u4fdd\u5b58\u4e00\u90e8\u5206\u3002<\/p>\n<hr \/>\n<h3>2. \u8bad\u7ec3\u663e\u5b58\u8d26\u672c&#xff1a;\u5230\u5e95\u54ea\u4e9b\u4e1c\u897f\u5360\u663e\u5b58<\/h3>\n<p>\u5148\u628a\u8bad\u7ec3\u65f6\u7684\u663e\u5b58\u62c6\u5f00\u3002\u4ee5\u6df7\u5408\u7cbe\u5ea6 AdamW \u4e3a\u4f8b&#xff0c;\u53c2\u6570\u76f8\u5173\u72b6\u6001\u5e38\u89c1\u5305\u62ec&#xff1a;<\/p>\n<table>\n<tr>\u72b6\u6001\u4f5c\u7528\u5e38\u89c1\u5b57\u8282\u6570<\/tr>\n<tbody>\n<tr>\n<td>\u6a21\u578b\u53c2\u6570<\/td>\n<td>\u524d\u5411\u548c\u53cd\u5411\u8ba1\u7b97<\/td>\n<td align=\"right\">BF16\/FP16 \u7ea6 2 bytes<\/td>\n<\/tr>\n<tr>\n<td>\u68af\u5ea6<\/td>\n<td>\u53cd\u5411\u4f20\u64ad\u5f97\u5230\u7684\u66f4\u65b0\u65b9\u5411<\/td>\n<td align=\"right\">BF16\/FP16 \u7ea6 2 bytes<\/td>\n<\/tr>\n<tr>\n<td>FP32 master weights<\/td>\n<td>\u4f18\u5316\u5668\u66f4\u7a33\u5b9a\u5730\u66f4\u65b0\u53c2\u6570&#xff0c;\u89c6\u65b9\u6848\u800c\u5b9a<\/td>\n<td align=\"right\">4 bytes<\/td>\n<\/tr>\n<tr>\n<td>Adam \u4e00\u9636\u52a8\u91cf m<\/td>\n<td>\u68af\u5ea6\u6ed1\u52a8\u5e73\u5747<\/td>\n<td align=\"right\">4 bytes<\/td>\n<\/tr>\n<tr>\n<td>Adam \u4e8c\u9636\u52a8\u91cf v<\/td>\n<td>\u68af\u5ea6\u5e73\u65b9\u6ed1\u52a8\u5e73\u5747<\/td>\n<td align=\"right\">4 bytes<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u7c97\u7565\u76f8\u52a0&#xff0c;\u6bcf\u4e2a\u53c2\u6570\u53ef\u80fd\u9700\u8981 12 \u5230 16 \u5b57\u8282\u7684\u53c2\u6570\u76f8\u5173\u8bad\u7ec3\u72b6\u6001\u3002\u8fd8\u6ca1\u7b97\u6fc0\u6d3b\u503c\u3001\u4e34\u65f6\u5f20\u91cf\u3001\u901a\u4fe1 buffer \u548c\u6846\u67b6\u5f00\u9500\u3002<\/p>\n<p>\u4e0b\u9762\u5199\u4e00\u4e2a\u4f30\u7b97\u5668\u3002\u5b83\u4e0d\u662f\u7cbe\u786e\u5bb9\u91cf\u89c4\u5212\u5de5\u5177&#xff0c;\u53ea\u7528\u4e8e\u5efa\u7acb\u76f4\u89c9&#xff1a;<\/p>\n<p><span class=\"token comment\"># Python 3.10&#043;<\/span><\/p>\n<p><span class=\"token keyword\">def<\/span> <span class=\"token function\">param_state_memory_gb<\/span><span class=\"token punctuation\">(<\/span>params_billion<span class=\"token punctuation\">,<\/span> master_weights<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    params <span class=\"token operator\">&#061;<\/span> params_billion <span class=\"token operator\">*<\/span> <span class=\"token number\">1_000_000_000<\/span><br \/>\n    bytes_per_param <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">2<\/span>  <span class=\"token comment\"># bf16\/fp16 params<\/span><br \/>\n    bytes_per_param <span class=\"token operator\">&#043;&#061;<\/span> <span class=\"token number\">2<\/span>  <span class=\"token comment\"># gradients<\/span><br \/>\n    bytes_per_param <span class=\"token operator\">&#043;&#061;<\/span> <span class=\"token number\">4<\/span>  <span class=\"token comment\"># Adam m<\/span><br \/>\n    bytes_per_param <span class=\"token operator\">&#043;&#061;<\/span> <span class=\"token number\">4<\/span>  <span class=\"token comment\"># Adam v<\/span><br \/>\n    <span class=\"token keyword\">if<\/span> master_weights<span class=\"token punctuation\">:<\/span><br \/>\n        bytes_per_param <span class=\"token operator\">&#043;&#061;<\/span> <span class=\"token number\">4<\/span><br \/>\n    <span class=\"token keyword\">return<\/span> params <span class=\"token operator\">*<\/span> bytes_per_param <span class=\"token operator\">\/<\/span> <span class=\"token number\">1024<\/span><span class=\"token operator\">**<\/span><span class=\"token number\">3<\/span><\/p>\n<p><span class=\"token keyword\">for<\/span> n <span class=\"token keyword\">in<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">7<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">13<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">70<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string-interpolation\"><span class=\"token string\">f&#034;<\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>n<span class=\"token punctuation\">:<\/span><span class=\"token format-spec\">&gt;2<\/span><span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">B params -&gt; param states \u2248 <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>param_state_memory_gb<span class=\"token punctuation\">(<\/span>n<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><span class=\"token format-spec\">.1f<\/span><span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\"> GB&#034;<\/span><\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>\u8fd0\u884c\u540e\u4f60\u4f1a\u53d1\u73b0&#xff0c;7B \u6a21\u578b\u7684\u53c2\u6570\u76f8\u5173\u8bad\u7ec3\u72b6\u6001\u5c31\u53ef\u80fd\u63a5\u8fd1\u767e GB \u7ea7\u522b\u3002\u5355\u5361\u653e\u4e0d\u4e0b&#xff0c;\u5c31\u9700\u8981\u5206\u5e03\u5f0f\u8bad\u7ec3\u548c\u72b6\u6001\u5207\u5206\u3002<\/p>\n<p>\u6fc0\u6d3b\u503c\u4e5f\u5f88\u91cd\u8981\u3002\u5e8f\u5217\u957f\u5ea6\u8d8a\u957f\u3001micro batch \u8d8a\u5927\u3001\u5c42\u6570\u8d8a\u6df1&#xff0c;\u6fc0\u6d3b\u503c\u8d8a\u5927\u3002\u68af\u5ea6\u68c0\u67e5\u70b9&#xff08;activation checkpointing&#xff09;\u53ef\u4ee5\u901a\u8fc7\u91cd\u65b0\u8ba1\u7b97\u90e8\u5206\u524d\u5411\u7ed3\u679c\u6765\u8282\u7701\u6fc0\u6d3b\u663e\u5b58&#xff0c;\u4f46\u4f1a\u589e\u52a0\u8ba1\u7b97\u65f6\u95f4\u3002<\/p>\n<p>\u6240\u4ee5\u8bad\u7ec3\u663e\u5b58\u53ef\u4ee5\u7c97\u7565\u770b\u6210&#xff1a;<\/p>\n<p>\u8bad\u7ec3\u663e\u5b58 &#061; \u53c2\u6570\u76f8\u5173\u72b6\u6001 &#043; \u6fc0\u6d3b\u503c &#043; \u901a\u4fe1 buffer &#043; \u4e34\u65f6\u5f20\u91cf &#043; \u6846\u67b6\u5f00\u9500<\/p>\n<p>ZeRO \u4e3b\u8981\u4f18\u5316\u53c2\u6570\u76f8\u5173\u72b6\u6001\u3002\u6fc0\u6d3b\u503c\u8fd8\u8981\u9760 micro batch\u3001seq_len\u3001activation checkpointing\u3001FlashAttention \u7b49\u624b\u6bb5\u914d\u5408\u3002<\/p>\n<hr \/>\n<h3>3. \u666e\u901a\u6570\u636e\u5e76\u884c\u4e3a\u4ec0\u4e48\u6d6a\u8d39<\/h3>\n<p>\u666e\u901a\u6570\u636e\u5e76\u884c\u7684\u6d41\u7a0b\u662f&#xff1a;\u6bcf\u5f20 GPU \u90fd\u6709\u4e00\u4efd\u5b8c\u6574\u6a21\u578b&#xff0c;\u5904\u7406\u4e0d\u540c batch&#xff0c;\u53cd\u5411\u4f20\u64ad\u540e\u540c\u6b65\u68af\u5ea6&#xff0c;\u518d\u5404\u81ea\u66f4\u65b0\u53c2\u6570\u3002<\/p>\n<p>\u53ef\u4ee5\u60f3\u8c61 4 \u5f20 GPU&#xff1a;<\/p>\n<p>GPU0: \u53c2\u6570 W &#043; \u68af\u5ea6 G &#043; Adam\u72b6\u6001 O<br \/>\nGPU1: \u53c2\u6570 W &#043; \u68af\u5ea6 G &#043; Adam\u72b6\u6001 O<br \/>\nGPU2: \u53c2\u6570 W &#043; \u68af\u5ea6 G &#043; Adam\u72b6\u6001 O<br \/>\nGPU3: \u53c2\u6570 W &#043; \u68af\u5ea6 G &#043; Adam\u72b6\u6001 O<\/p>\n<p>\u8fd9\u91cc\u7684 W\u3001G\u3001O \u5728\u6bcf\u5f20\u5361\u4e0a\u90fd\u662f\u5b8c\u6574\u526f\u672c\u3002\u8ba1\u7b97\u4e0a\u5b83\u4eec\u5904\u7406\u4e0d\u540c\u6570\u636e&#xff0c;\u663e\u5b58\u4e0a\u5374\u91cd\u590d\u4fdd\u5b58\u540c\u6837\u72b6\u6001\u3002<\/p>\n<p>\u666e\u901a\u6570\u636e\u5e76\u884c\u7684\u4f18\u70b9\u662f\u7b80\u5355\u3001\u541e\u5410\u9ad8\u3001\u901a\u4fe1\u6a21\u5f0f\u6210\u719f&#xff1b;\u7f3a\u70b9\u662f\u5355\u5361\u663e\u5b58\u538b\u529b\u4e0d\u4f1a\u56e0\u4e3a GPU \u6570\u91cf\u589e\u52a0\u800c\u660e\u663e\u4e0b\u964d\u3002\u5bf9\u4e8e\u5927\u6a21\u578b\u8bad\u7ec3&#xff0c;\u4f18\u5316\u5668\u72b6\u6001\u548c\u68af\u5ea6\u526f\u672c\u4f1a\u6210\u4e3a\u4e3b\u8981\u74f6\u9888\u3002<\/p>\n<p>ZeRO \u7684\u540d\u5b57\u91cc\u6709 \u201cZero Redundancy\u201d&#xff0c;\u610f\u601d\u5c31\u662f\u5c3d\u91cf\u6d88\u9664\u8fd9\u4e9b\u5197\u4f59\u526f\u672c\u3002<\/p>\n<hr \/>\n<h3>4. ZeRO-1&#xff1a;\u5148\u5207\u4f18\u5316\u5668\u72b6\u6001<\/h3>\n<p>ZeRO-1 \u5207\u5206\u7684\u662f\u4f18\u5316\u5668\u72b6\u6001\u3002\u6bcf\u5f20\u5361\u4e0d\u518d\u4fdd\u5b58\u5b8c\u6574 Adam \u7684 m\u3001v \u548c\u53ef\u80fd\u7684 master weights&#xff0c;\u800c\u53ea\u4fdd\u5b58\u5176\u4e2d\u4e00\u90e8\u5206\u3002<\/p>\n<p>\u7b80\u5316\u540e\u53ef\u4ee5\u7406\u89e3\u4e3a&#xff1a;<\/p>\n<p>GPU0: \u5b8c\u6574\u53c2\u6570 W &#043; \u5b8c\u6574\u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<br \/>\nGPU1: \u5b8c\u6574\u53c2\u6570 W &#043; \u5b8c\u6574\u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<br \/>\nGPU2: \u5b8c\u6574\u53c2\u6570 W &#043; \u5b8c\u6574\u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<br \/>\nGPU3: \u5b8c\u6574\u53c2\u6570 W &#043; \u5b8c\u6574\u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<\/p>\n<p>\u4e3a\u4ec0\u4e48\u5148\u5207\u4f18\u5316\u5668\u72b6\u6001&#xff1f;\u56e0\u4e3a AdamW \u72b6\u6001\u5f88\u5927\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u53c2\u6570&#xff0c;Adam \u81f3\u5c11\u4fdd\u5b58 m \u548c v \u4e24\u4e2a FP32 \u5f20\u91cf&#xff0c;\u663e\u5b58\u5f00\u9500\u5f88\u53ef\u89c2\u3002<\/p>\n<p>ZeRO-1 \u7684\u4f18\u70b9&#xff1a;<\/p>\n<ul>\n<li>\u6539\u52a8\u76f8\u5bf9\u5c0f&#xff1b;<\/li>\n<li>\u901a\u4fe1\u590d\u6742\u5ea6\u8f83\u4f4e&#xff1b;<\/li>\n<li>\u80fd\u8282\u7701\u4e00\u5927\u5757\u4f18\u5316\u5668\u72b6\u6001\u663e\u5b58&#xff1b;<\/li>\n<li>\u5bf9\u8bad\u7ec3\u541e\u5410\u5f71\u54cd\u901a\u5e38\u6bd4\u66f4\u9ad8\u9636\u6bb5\u5c0f\u3002<\/li>\n<\/ul>\n<p>ZeRO-1 \u7684\u9650\u5236&#xff1a;\u53c2\u6570\u548c\u68af\u5ea6\u4ecd\u7136\u5b8c\u6574\u590d\u5236\u5728\u6bcf\u5f20\u5361\u4e0a\u3002\u5982\u679c\u6a21\u578b\u53c2\u6570\u672c\u8eab\u5c31\u653e\u4e0d\u4e0b&#xff0c;ZeRO-1 \u4e0d\u591f\u3002<\/p>\n<hr \/>\n<h3>5. ZeRO-2&#xff1a;\u4f18\u5316\u5668\u72b6\u6001\u548c\u68af\u5ea6\u90fd\u5207<\/h3>\n<p>ZeRO-2 \u5728 ZeRO-1 \u57fa\u7840\u4e0a\u7ee7\u7eed\u5207\u5206\u68af\u5ea6\u3002\u53cd\u5411\u4f20\u64ad\u65f6&#xff0c;\u68af\u5ea6\u4e0d\u518d\u6bcf\u5f20\u5361\u4fdd\u7559\u5b8c\u6574\u526f\u672c&#xff0c;\u800c\u662f\u901a\u8fc7 reduce-scatter \u7b49\u901a\u4fe1\u628a\u68af\u5ea6\u5206\u7247\u5230\u5bf9\u5e94 GPU\u3002<\/p>\n<p>\u7b80\u5316\u56fe&#xff1a;<\/p>\n<p>GPU0: \u5b8c\u6574\u53c2\u6570 W &#043; 1\/4 \u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<br \/>\nGPU1: \u5b8c\u6574\u53c2\u6570 W &#043; 1\/4 \u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<br \/>\nGPU2: \u5b8c\u6574\u53c2\u6570 W &#043; 1\/4 \u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<br \/>\nGPU3: \u5b8c\u6574\u53c2\u6570 W &#043; 1\/4 \u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<\/p>\n<p>\u8fd9\u901a\u5e38\u662f\u5f88\u591a\u5fae\u8c03\u548c\u4e2d\u7b49\u89c4\u6a21\u8bad\u7ec3\u7684\u5b9e\u7528\u8d77\u70b9\u3002\u56e0\u4e3a\u5b83\u5728\u663e\u5b58\u8282\u7701\u548c\u901a\u4fe1\u590d\u6742\u5ea6\u4e4b\u95f4\u6bd4\u8f83\u5e73\u8861\u3002<\/p>\n<p>ZeRO-2 \u9002\u5408&#xff1a;<\/p>\n<p>1. \u6a21\u578b\u53c2\u6570\u80fd\u5b8c\u6574\u653e\u8fdb\u5355\u5361&#xff1b;<br \/>\n2. \u68af\u5ea6\u548c\u4f18\u5316\u5668\u72b6\u6001\u5bfc\u81f4 OOM&#xff1b;<br \/>\n3. \u5e0c\u671b\u6bd4 ZeRO-3 \u66f4\u7b80\u5355\u3001\u66f4\u5feb&#xff1b;<br \/>\n4. LoRA\/QLoRA \u4ee5\u5916\u7684\u8f83\u5927\u89c4\u6a21 SFT \u6216 CPT\u3002<\/p>\n<p>\u4f46\u5982\u679c\u5b8c\u6574\u53c2\u6570 W \u90fd\u653e\u4e0d\u4e0b&#xff0c;ZeRO-2 \u4ecd\u7136\u4e0d\u591f&#xff0c;\u56e0\u4e3a\u6bcf\u5f20\u5361\u8fd8\u4fdd\u7559\u5b8c\u6574\u53c2\u6570\u3002<\/p>\n<hr \/>\n<h3>6. ZeRO-3&#xff1a;\u8fde\u53c2\u6570\u4e5f\u5207<\/h3>\n<p>ZeRO-3 \u8fdb\u4e00\u6b65\u5207\u5206\u6a21\u578b\u53c2\u6570\u3002\u6bcf\u5f20 GPU \u53ea\u4fdd\u5b58\u4e00\u90e8\u5206\u53c2\u6570&#xff0c;\u9700\u8981\u8ba1\u7b97\u67d0\u4e00\u5c42\u65f6&#xff0c;\u518d\u4e34\u65f6 all-gather \u6240\u9700\u53c2\u6570&#xff0c;\u8ba1\u7b97\u5b8c\u540e\u91ca\u653e\u6216\u91cd\u65b0\u5206\u7247\u3002<\/p>\n<p>\u7b80\u5316\u56fe&#xff1a;<\/p>\n<p>GPU0: 1\/4 \u53c2\u6570 W &#043; 1\/4 \u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<br \/>\nGPU1: 1\/4 \u53c2\u6570 W &#043; 1\/4 \u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<br \/>\nGPU2: 1\/4 \u53c2\u6570 W &#043; 1\/4 \u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<br \/>\nGPU3: 1\/4 \u53c2\u6570 W &#043; 1\/4 \u68af\u5ea6 G &#043; 1\/4 \u4f18\u5316\u5668\u72b6\u6001 O<\/p>\n<p>\u8fd9\u5c31\u662f ZeRO-3 \u6700\u7701\u663e\u5b58\u7684\u539f\u56e0\u3002\u5b83\u8ba9\u6a21\u578b\u53c2\u6570\u672c\u8eab\u4e5f\u4e0d\u518d\u5b8c\u6574\u590d\u5236&#xff0c;\u5355\u5361\u663e\u5b58\u538b\u529b\u5927\u5e45\u4e0b\u964d\u3002<\/p>\n<p>\u4ee3\u4ef7\u4e5f\u5f88\u660e\u663e&#xff1a;<\/p>\n<ul>\n<li>\u524d\u5411\u548c\u53cd\u5411\u8fc7\u7a0b\u4e2d\u66f4\u9891\u7e41\u901a\u4fe1&#xff1b;<\/li>\n<li>\u5bf9\u7f51\u7edc\u5e26\u5bbd\u548c\u5ef6\u8fdf\u66f4\u654f\u611f&#xff1b;<\/li>\n<li>checkpoint \u4fdd\u5b58\u548c\u52a0\u8f7d\u66f4\u590d\u6742&#xff1b;<\/li>\n<li>\u5c0f batch \u6216\u4f4e\u5e26\u5bbd\u73af\u5883\u4e0b\u541e\u5410\u53ef\u80fd\u660e\u663e\u4e0b\u964d&#xff1b;<\/li>\n<li>\u6392\u67e5\u95ee\u9898\u65f6\u65e5\u5fd7\u548c\u5206\u7247\u72b6\u6001\u66f4\u96be\u7406\u89e3\u3002<\/li>\n<\/ul>\n<p>\u56e0\u6b64 ZeRO-3 \u4e0d\u662f\u201c\u6c38\u8fdc\u66f4\u597d\u201d\u3002\u5b83\u662f\u5bb9\u91cf\u4f18\u5148\u7684\u65b9\u6848\u3002\u5f53\u663e\u5b58\u591f\u7528\u65f6&#xff0c;ZeRO-2 \u751a\u81f3\u666e\u901a DDP \u53ef\u80fd\u66f4\u5feb\u3002<\/p>\n<hr \/>\n<h3>7. \u7528\u4ee3\u7801\u4f30\u7b97 ZeRO \u9636\u6bb5\u7684\u53c2\u6570\u72b6\u6001\u8282\u7701<\/h3>\n<p>\u4e0b\u9762\u7528\u4e00\u4e2a\u7b80\u5316\u4f30\u7b97\u5668\u611f\u53d7 ZeRO \u5bf9\u53c2\u6570\u76f8\u5173\u72b6\u6001\u7684\u5f71\u54cd\u3002\u6ce8\u610f\u8fd9\u53ea\u4f30\u7b97\u53c2\u6570\u3001\u68af\u5ea6\u3001\u4f18\u5316\u5668\u72b6\u6001&#xff0c;\u4e0d\u5305\u62ec\u6fc0\u6d3b\u503c\u3002<\/p>\n<p><span class=\"token comment\"># Python 3.10&#043;<\/span><\/p>\n<p><span class=\"token keyword\">def<\/span> <span class=\"token function\">zero_memory_estimate<\/span><span class=\"token punctuation\">(<\/span>params_billion<span class=\"token punctuation\">,<\/span> num_gpus<span class=\"token punctuation\">,<\/span> stage<span class=\"token punctuation\">,<\/span> master_weights<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    params <span class=\"token operator\">&#061;<\/span> params_billion <span class=\"token operator\">*<\/span> <span class=\"token number\">1_000_000_000<\/span><\/p>\n<p>    param_bytes <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">2<\/span><br \/>\n    grad_bytes <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">2<\/span><br \/>\n    optim_bytes <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">4<\/span> <span class=\"token operator\">&#043;<\/span> <span class=\"token number\">4<\/span> <span class=\"token operator\">&#043;<\/span> <span class=\"token punctuation\">(<\/span><span class=\"token number\">4<\/span> <span class=\"token keyword\">if<\/span> master_weights <span class=\"token keyword\">else<\/span> <span class=\"token number\">0<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>    <span class=\"token keyword\">if<\/span> stage <span class=\"token operator\">&#061;&#061;<\/span> <span class=\"token number\">0<\/span><span class=\"token punctuation\">:<\/span>  <span class=\"token comment\"># \u666e\u901a\u6570\u636e\u5e76\u884c<\/span><br \/>\n        per_gpu <span class=\"token operator\">&#061;<\/span> params <span class=\"token operator\">*<\/span> <span class=\"token punctuation\">(<\/span>param_bytes <span class=\"token operator\">&#043;<\/span> grad_bytes <span class=\"token operator\">&#043;<\/span> optim_bytes<span class=\"token punctuation\">)<\/span><br \/>\n    <span class=\"token keyword\">elif<\/span> stage <span class=\"token operator\">&#061;&#061;<\/span> <span class=\"token number\">1<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        per_gpu <span class=\"token operator\">&#061;<\/span> params <span class=\"token operator\">*<\/span> <span class=\"token punctuation\">(<\/span>param_bytes <span class=\"token operator\">&#043;<\/span> grad_bytes <span class=\"token operator\">&#043;<\/span> optim_bytes <span class=\"token operator\">\/<\/span> num_gpus<span class=\"token punctuation\">)<\/span><br \/>\n    <span class=\"token keyword\">elif<\/span> stage <span class=\"token operator\">&#061;&#061;<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        per_gpu <span class=\"token operator\">&#061;<\/span> params <span class=\"token operator\">*<\/span> <span class=\"token punctuation\">(<\/span>param_bytes <span class=\"token operator\">&#043;<\/span> grad_bytes <span class=\"token operator\">\/<\/span> num_gpus <span class=\"token operator\">&#043;<\/span> optim_bytes <span class=\"token operator\">\/<\/span> num_gpus<span class=\"token punctuation\">)<\/span><br \/>\n    <span class=\"token keyword\">elif<\/span> stage <span class=\"token operator\">&#061;&#061;<\/span> <span class=\"token number\">3<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        per_gpu <span class=\"token operator\">&#061;<\/span> params <span class=\"token operator\">*<\/span> <span class=\"token punctuation\">(<\/span>param_bytes <span class=\"token operator\">&#043;<\/span> grad_bytes <span class=\"token operator\">&#043;<\/span> optim_bytes<span class=\"token punctuation\">)<\/span> <span class=\"token operator\">\/<\/span> num_gpus<br \/>\n    <span class=\"token keyword\">else<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token keyword\">raise<\/span> ValueError<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;stage must be 0, 1, 2, or 3&#034;<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>    <span class=\"token keyword\">return<\/span> per_gpu <span class=\"token operator\">\/<\/span> <span class=\"token number\">1024<\/span><span class=\"token operator\">**<\/span><span class=\"token number\">3<\/span><\/p>\n<p><span class=\"token keyword\">for<\/span> stage <span class=\"token keyword\">in<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">3<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    mem <span class=\"token operator\">&#061;<\/span> zero_memory_estimate<span class=\"token punctuation\">(<\/span>params_billion<span class=\"token operator\">&#061;<\/span><span class=\"token number\">13<\/span><span class=\"token punctuation\">,<\/span> num_gpus<span class=\"token operator\">&#061;<\/span><span class=\"token number\">8<\/span><span class=\"token punctuation\">,<\/span> stage<span class=\"token operator\">&#061;<\/span>stage<span class=\"token punctuation\">)<\/span><br \/>\n    <span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string-interpolation\"><span class=\"token string\">f&#034;ZeRO stage <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>stage<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">: \u2248 <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>mem<span class=\"token punctuation\">:<\/span><span class=\"token format-spec\">.1f<\/span><span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\"> GB per GPU for param states&#034;<\/span><\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>\u8fd9\u6bb5\u4ee3\u7801\u53ef\u4ee5\u5e2e\u4f60\u7406\u89e3\u6bcf\u4e2a\u9636\u6bb5\u5207\u4e86\u4ec0\u4e48&#xff1a;<\/p>\n<ul>\n<li>stage 0 \u4e0d\u5207&#xff0c;\u6240\u6709\u72b6\u6001\u5b8c\u6574\u590d\u5236&#xff1b;<\/li>\n<li>stage 1 \u53ea\u5207\u4f18\u5316\u5668\u72b6\u6001&#xff1b;<\/li>\n<li>stage 2 \u5207\u4f18\u5316\u5668\u72b6\u6001\u548c\u68af\u5ea6&#xff1b;<\/li>\n<li>stage 3 \u53c2\u6570\u3001\u68af\u5ea6\u3001\u4f18\u5316\u5668\u72b6\u6001\u90fd\u6309 GPU \u6570\u5207\u5206\u3002<\/li>\n<\/ul>\n<p>\u771f\u5b9e\u663e\u5b58\u4f1a\u66f4\u9ad8&#xff0c;\u56e0\u4e3a\u8fd8\u6709\u6fc0\u6d3b\u503c\u548c\u4e34\u65f6\u5f20\u91cf\u3002\u4f30\u7b97\u5668\u7684\u7528\u9014\u4e0d\u662f\u7ed9\u91c7\u8d2d\u62a5\u4ef7&#xff0c;\u800c\u662f\u5e2e\u52a9\u4f60\u5224\u65ad&#xff1a;OOM \u4e3b\u8981\u6765\u81ea\u53c2\u6570\u72b6\u6001&#xff0c;\u8fd8\u662f\u6765\u81ea\u6fc0\u6d3b\u548c batch\u3002<\/p>\n<hr \/>\n<h3>8. DeepSpeed \u914d\u7f6e\u600e\u4e48\u8bfb<\/h3>\n<p>\u4e00\u4e2a ZeRO-2 \u7684\u7b80\u5316\u914d\u7f6e\u53ef\u80fd\u957f\u8fd9\u6837&#xff1a;<\/p>\n<p><span class=\"token punctuation\">{<\/span><br \/>\n  <span class=\"token string-property property\">&#034;train_micro_batch_size_per_gpu&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;gradient_accumulation_steps&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">16<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;bf16&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token punctuation\">{<\/span><br \/>\n    <span class=\"token string-property property\">&#034;enabled&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><br \/>\n  <span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;zero_optimization&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token punctuation\">{<\/span><br \/>\n    <span class=\"token string-property property\">&#034;stage&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;allgather_partitions&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;reduce_scatter&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;overlap_comm&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;contiguous_gradients&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><br \/>\n  <span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;gradient_clipping&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">1.0<\/span><br \/>\n<span class=\"token punctuation\">}<\/span><\/p>\n<p>\u9010\u9879\u89e3\u91ca&#xff1a;<\/p>\n<table>\n<tr>\u5b57\u6bb5\u542b\u4e49<\/tr>\n<tbody>\n<tr>\n<td>train_micro_batch_size_per_gpu<\/td>\n<td>\u5355\u5f20 GPU \u4e00\u6b21\u524d\u5411\u5904\u7406\u591a\u5c11\u6761\u5e8f\u5217<\/td>\n<\/tr>\n<tr>\n<td>gradient_accumulation_steps<\/td>\n<td>\u7d2f\u79ef\u591a\u5c11\u6b21\u68af\u5ea6\u518d\u505a\u4e00\u6b21\u4f18\u5316\u5668\u66f4\u65b0<\/td>\n<\/tr>\n<tr>\n<td>bf16.enabled<\/td>\n<td>\u4f7f\u7528 BF16 \u6df7\u5408\u7cbe\u5ea6&#xff0c;\u901a\u5e38\u6bd4 FP16 \u66f4\u7a33\u5b9a<\/td>\n<\/tr>\n<tr>\n<td>zero_optimization.stage<\/td>\n<td>ZeRO \u9636\u6bb5&#xff0c;1\/2\/3 \u5bf9\u5e94\u4e0d\u540c\u5207\u5206\u8303\u56f4<\/td>\n<\/tr>\n<tr>\n<td>allgather_partitions<\/td>\n<td>\u805a\u5408\u5206\u7247\u53c2\u6570\u6216\u72b6\u6001\u7684\u901a\u4fe1\u7b56\u7565<\/td>\n<\/tr>\n<tr>\n<td>reduce_scatter<\/td>\n<td>\u68af\u5ea6\u5f52\u7ea6\u65f6\u76f4\u63a5\u6563\u5c04\u5206\u7247<\/td>\n<\/tr>\n<tr>\n<td>overlap_comm<\/td>\n<td>\u5c1d\u8bd5\u8ba9\u901a\u4fe1\u548c\u8ba1\u7b97\u91cd\u53e0&#xff0c;\u63d0\u9ad8\u541e\u5410<\/td>\n<\/tr>\n<tr>\n<td>contiguous_gradients<\/td>\n<td>\u8ba9\u68af\u5ea6\u5185\u5b58\u66f4\u8fde\u7eed&#xff0c;\u51cf\u5c11\u788e\u7247<\/td>\n<\/tr>\n<tr>\n<td>gradient_clipping<\/td>\n<td>\u9650\u5236\u68af\u5ea6\u8303\u6570&#xff0c;\u63d0\u5347\u8bad\u7ec3\u7a33\u5b9a\u6027<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u771f\u5b9e\u9879\u76ee\u8fd8\u4f1a\u914d\u7f6e optimizer\u3001scheduler\u3001checkpoint\u3001activation checkpointing\u3001offload\u3001\u65e5\u5fd7\u548c\u8bc4\u4f30\u3002\u4e0d\u8981\u628a\u7f51\u4e0a\u914d\u7f6e\u76f4\u63a5\u590d\u5236\u5230\u751f\u4ea7\u8bad\u7ec3\u3002\u914d\u7f6e\u5fc5\u987b\u7ed3\u5408\u6a21\u578b\u89c4\u6a21\u3001GPU \u663e\u5b58\u3001\u7f51\u7edc\u3001seq_len\u3001batch \u548c\u76ee\u6807\u4efb\u52a1\u8c03\u6574\u3002<\/p>\n<hr \/>\n<h3>9. ZeRO-3 \u914d\u7f6e\u548c\u51e0\u4e2a\u5173\u952e\u5f00\u5173<\/h3>\n<p>ZeRO-3 \u5e38\u89c1\u914d\u7f6e\u4f1a\u66f4\u590d\u6742&#xff0c;\u4f8b\u5982&#xff1a;<\/p>\n<p><span class=\"token punctuation\">{<\/span><br \/>\n  <span class=\"token string-property property\">&#034;train_micro_batch_size_per_gpu&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;gradient_accumulation_steps&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">8<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;bf16&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token punctuation\">{<\/span><br \/>\n    <span class=\"token string-property property\">&#034;enabled&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><br \/>\n  <span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;zero_optimization&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token punctuation\">{<\/span><br \/>\n    <span class=\"token string-property property\">&#034;stage&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">3<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;overlap_comm&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;contiguous_gradients&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;reduce_bucket_size&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">500000000<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;stage3_prefetch_bucket_size&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">50000000<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;stage3_param_persistence_threshold&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">100000<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;stage3_gather_16bit_weights_on_model_save&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><br \/>\n  <span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;gradient_clipping&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">1.0<\/span><br \/>\n<span class=\"token punctuation\">}<\/span><\/p>\n<p>\u51e0\u4e2a\u5b57\u6bb5\u7684\u76f4\u89c9&#xff1a;<\/p>\n<ul>\n<li>reduce_bucket_size&#xff1a;\u68af\u5ea6\u901a\u4fe1\u5206\u6876\u5927\u5c0f&#xff0c;\u5f71\u54cd\u901a\u4fe1\u6548\u7387\u548c\u663e\u5b58\u5cf0\u503c&#xff1b;<\/li>\n<li>stage3_prefetch_bucket_size&#xff1a;\u63d0\u524d\u62c9\u53d6\u53c2\u6570\u5206\u7247\u7684\u5927\u5c0f&#xff0c;\u5f71\u54cd\u901a\u4fe1\u9690\u85cf\u548c\u5185\u5b58&#xff1b;<\/li>\n<li>stage3_param_persistence_threshold&#xff1a;\u5c0f\u53c2\u6570\u662f\u5426\u5e38\u9a7b&#xff0c;\u51cf\u5c11\u9891\u7e41 gather&#xff1b;<\/li>\n<li>stage3_gather_16bit_weights_on_model_save&#xff1a;\u4fdd\u5b58\u6a21\u578b\u65f6\u662f\u5426\u805a\u5408\u5b8c\u6574 16-bit \u6743\u91cd\u3002<\/li>\n<\/ul>\n<p>ZeRO-3 \u7684\u6838\u5fc3\u8c03\u53c2\u601d\u8def\u662f&#xff1a;\u5728\u663e\u5b58\u5cf0\u503c\u3001\u901a\u4fe1\u6548\u7387\u548c\u4fdd\u5b58\u52a0\u8f7d\u4fbf\u5229\u6027\u4e4b\u95f4\u5e73\u8861\u3002\u6876\u592a\u5927\u53ef\u80fd\u663e\u5b58\u5cf0\u503c\u9ad8&#xff0c;\u6876\u592a\u5c0f\u901a\u4fe1\u5f00\u9500\u591a\u3002\u4e0d\u540c\u786c\u4ef6\u548c\u7f51\u7edc\u4e0b&#xff0c;\u6700\u4f73\u503c\u4f1a\u53d8\u5316\u3002<\/p>\n<hr \/>\n<h3>10. Offload&#xff1a;\u628a\u72b6\u6001\u653e\u5230 CPU \u6216 NVMe<\/h3>\n<p>\u5f53 ZeRO \u5207\u5206\u540e GPU \u663e\u5b58\u4ecd\u7136\u4e0d\u591f&#xff0c;\u53ef\u4ee5\u628a\u53c2\u6570\u6216\u4f18\u5316\u5668\u72b6\u6001 offload \u5230 CPU \u5185\u5b58&#xff0c;\u751a\u81f3 NVMe\u3002Offload \u7684\u76f4\u89c9\u662f&#xff1a;GPU \u653e\u4e0d\u4e0b&#xff0c;\u5c31\u628a\u4e00\u90e8\u5206\u72b6\u6001\u653e\u5230\u66f4\u6162\u4f46\u5bb9\u91cf\u66f4\u5927\u7684\u5730\u65b9\u3002<\/p>\n<p>\u4e00\u4e2a\u7b80\u5316\u914d\u7f6e\u793a\u4f8b&#xff1a;<\/p>\n<p><span class=\"token punctuation\">{<\/span><br \/>\n  <span class=\"token string-property property\">&#034;zero_optimization&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token punctuation\">{<\/span><br \/>\n    <span class=\"token string-property property\">&#034;stage&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">3<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;offload_optimizer&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token punctuation\">{<\/span><br \/>\n      <span class=\"token string-property property\">&#034;device&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token string\">&#034;cpu&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n      <span class=\"token string-property property\">&#034;pin_memory&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><br \/>\n    <span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    <span class=\"token string-property property\">&#034;offload_param&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token punctuation\">{<\/span><br \/>\n      <span class=\"token string-property property\">&#034;device&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token string\">&#034;cpu&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n      <span class=\"token string-property property\">&#034;pin_memory&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token boolean\">true<\/span><br \/>\n    <span class=\"token punctuation\">}<\/span><br \/>\n  <span class=\"token punctuation\">}<\/span><br \/>\n<span class=\"token punctuation\">}<\/span><\/p>\n<p>Offload \u7684\u53d6\u820d&#xff1a;<\/p>\n<table>\n<tr>\u65b9\u6848\u4f18\u70b9\u4ee3\u4ef7<\/tr>\n<tbody>\n<tr>\n<td>GPU-only<\/td>\n<td>\u901f\u5ea6\u6700\u5feb&#xff0c;\u901a\u4fe1\u8def\u5f84\u7b80\u5355<\/td>\n<td>\u663e\u5b58\u8981\u6c42\u9ad8<\/td>\n<\/tr>\n<tr>\n<td>CPU offload<\/td>\n<td>\u5927\u5e45\u964d\u4f4e GPU \u663e\u5b58<\/td>\n<td>\u53d7 PCIe\/CPU \u5185\u5b58\u5e26\u5bbd\u5f71\u54cd<\/td>\n<\/tr>\n<tr>\n<td>NVMe offload<\/td>\n<td>\u5bb9\u91cf\u66f4\u5927<\/td>\n<td>\u901f\u5ea6\u66f4\u6162&#xff0c;\u5bf9\u78c1\u76d8\u548c\u914d\u7f6e\u8981\u6c42\u9ad8<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8981\u7279\u522b\u5f3a\u8c03&#xff1a;offload \u901a\u5e38\u4e0d\u662f\u6027\u80fd\u4f18\u5316&#xff0c;\u800c\u662f\u5bb9\u91cf\u515c\u5e95\u3002 \u80fd\u5728 GPU-only \u8dd1\u901a\u4e14\u541e\u5410\u53ef\u63a5\u53d7\u65f6&#xff0c;\u4e0d\u8981\u4e3a\u4e86\u201c\u770b\u8d77\u6765\u66f4\u9ad8\u7ea7\u201d\u5f00\u542f offload\u3002<\/p>\n<hr \/>\n<h3>11. \u5982\u4f55\u9009\u62e9 ZeRO \u9636\u6bb5<\/h3>\n<p>\u53ef\u4ee5\u7528\u4e0b\u9762\u7684\u51b3\u7b56\u987a\u5e8f&#xff1a;<\/p>\n<p>1. \u5148\u4f30\u7b97\u53c2\u6570\u72b6\u6001\u548c\u6fc0\u6d3b\u503c\u663e\u5b58&#xff1b;<br \/>\n2. \u5982\u679c\u666e\u901a DDP \u80fd\u8dd1\u4e14\u541e\u5410\u597d&#xff0c;\u5148\u4e0d\u7528 ZeRO&#xff1b;<br \/>\n3. \u5982\u679c\u4f18\u5316\u5668\u72b6\u6001\u5bfc\u81f4\u663e\u5b58\u7d27\u5f20&#xff0c;\u8003\u8651 ZeRO-1&#xff1b;<br \/>\n4. \u5982\u679c\u68af\u5ea6\u548c\u4f18\u5316\u5668\u72b6\u6001\u5bfc\u81f4 OOM&#xff0c;\u4f18\u5148\u8bd5 ZeRO-2&#xff1b;<br \/>\n5. \u5982\u679c\u5b8c\u6574\u53c2\u6570\u4e5f\u653e\u4e0d\u4e0b&#xff0c;\u4f7f\u7528 ZeRO-3&#xff1b;<br \/>\n6. \u5982\u679c ZeRO-3 \u4ecd\u653e\u4e0d\u4e0b&#xff0c;\u518d\u8003\u8651 offload&#xff1b;<br \/>\n7. \u6bcf\u6b21\u53ea\u6539\u4e00\u4e2a\u4e3b\u8981\u53d8\u91cf&#xff0c;\u8bb0\u5f55\u541e\u5410\u3001\u663e\u5b58\u548c\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u66f4\u5de5\u7a0b\u5316\u7684\u5224\u65ad\u8868&#xff1a;<\/p>\n<table>\n<tr>\u573a\u666f\u63a8\u8350\u8d77\u70b9<\/tr>\n<tbody>\n<tr>\n<td>\u5c0f\u6a21\u578b\u5168\u53c2\u6570\u5fae\u8c03&#xff0c;\u663e\u5b58\u5145\u8db3<\/td>\n<td>DDP \u6216 ZeRO-1<\/td>\n<\/tr>\n<tr>\n<td>7B\/13B SFT&#xff0c;\u663e\u5b58\u7d27\u5f20<\/td>\n<td>ZeRO-2<\/td>\n<\/tr>\n<tr>\n<td>\u8bad\u7ec3\/\u5fae\u8c03\u8f83\u5927\u6a21\u578b&#xff0c;\u5355\u5361\u653e\u4e0d\u4e0b\u5b8c\u6574\u53c2\u6570<\/td>\n<td>ZeRO-3<\/td>\n<\/tr>\n<tr>\n<td>\u5355\u673a\u591a\u5361\u7f51\u7edc\u4e00\u822c<\/td>\n<td>\u5c3d\u91cf\u907f\u514d\u8fc7\u91cd ZeRO-3 \u901a\u4fe1<\/td>\n<\/tr>\n<tr>\n<td>\u53ea\u6709\u5c11\u91cf\u4efb\u52a1\u6570\u636e<\/td>\n<td>\u5148\u8003\u8651 LoRA\/QLoRA&#xff0c;\u800c\u4e0d\u662f\u5168\u53c2\u6570 ZeRO<\/td>\n<\/tr>\n<tr>\n<td>\u6781\u9650\u4f4e\u663e\u5b58<\/td>\n<td>ZeRO-3 &#043; offload&#xff0c;\u4f46\u63a5\u53d7\u541e\u5410\u4e0b\u964d<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>ZeRO \u9636\u6bb5\u8d8a\u9ad8&#xff0c;\u663e\u5b58\u8d8a\u7701&#xff0c;\u4f46\u901a\u4fe1\u548c\u7cfb\u7edf\u590d\u6742\u5ea6\u8d8a\u9ad8\u3002\u9009\u62e9\u65f6\u4e0d\u8981\u53ea\u770b\u201c\u80fd\u4e0d\u80fd\u8dd1\u201d&#xff0c;\u8fd8\u8981\u770b tokens\/s\u3001\u6545\u969c\u6062\u590d\u3001checkpoint \u65f6\u95f4\u548c\u8c03\u8bd5\u6210\u672c\u3002<\/p>\n<hr \/>\n<h3>12. \u5e38\u89c1\u95ee\u9898\u6392\u67e5<\/h3>\n<p>ZeRO \u8bad\u7ec3\u51fa\u95ee\u9898\u65f6&#xff0c;\u4e0d\u8981\u53ea\u76ef\u7740 OOM \u62a5\u9519\u3002\u53ef\u4ee5\u6309\u5c42\u6b21\u6392\u67e5&#xff1a;<\/p>\n<table>\n<tr>\u95ee\u9898\u53ef\u80fd\u539f\u56e0\u6392\u67e5\u65b9\u5411<\/tr>\n<tbody>\n<tr>\n<td>\u542f\u52a8\u5c31 OOM<\/td>\n<td>micro batch \u592a\u5927\u3001seq_len \u592a\u957f\u3001ZeRO \u9636\u6bb5\u4e0d\u591f<\/td>\n<td>\u964d micro batch\u3001\u5f00 checkpointing\u3001\u5347 ZeRO \u9636\u6bb5<\/td>\n<\/tr>\n<tr>\n<td>\u8bad\u7ec3\u4e2d\u9014 OOM<\/td>\n<td>\u957f\u6837\u672c\u3001\u5185\u5b58\u788e\u7247\u3001\u8bc4\u4f30\u9636\u6bb5\u5cf0\u503c<\/td>\n<td>\u68c0\u67e5\u957f\u5ea6\u5206\u5e03\u3001bucket\u3001eval batch<\/td>\n<\/tr>\n<tr>\n<td>\u541e\u5410\u5f88\u4f4e<\/td>\n<td>\u901a\u4fe1\u74f6\u9888\u3001offload\u3001\u6570\u636e\u52a0\u8f7d\u6162<\/td>\n<td>\u770b GPU \u5229\u7528\u7387\u3001\u7f51\u7edc\u3001CPU\u3001dataloader<\/td>\n<\/tr>\n<tr>\n<td>checkpoint \u5f88\u6162<\/td>\n<td>ZeRO-3 \u805a\u5408\u6743\u91cd\u3001\u78c1\u76d8\u5e26\u5bbd\u4e0d\u8db3<\/td>\n<td>\u8c03\u6574\u4fdd\u5b58\u9891\u7387\u3001\u5206\u7247\u4fdd\u5b58\u3001\u68c0\u67e5\u5b58\u50a8<\/td>\n<\/tr>\n<tr>\n<td>\u6062\u590d\u540e loss \u5f02\u5e38<\/td>\n<td>optimizer\/scheduler\/data state \u672a\u6062\u590d<\/td>\n<td>\u6838\u5bf9 checkpoint \u5b8c\u6574\u6027\u548c\u914d\u7f6e\u7248\u672c<\/td>\n<\/tr>\n<tr>\n<td>\u591a\u5361\u4e0d\u4e00\u81f4<\/td>\n<td>\u968f\u673a\u79cd\u5b50\u3001\u6570\u636e\u5206\u7247\u3001\u901a\u4fe1\u5f02\u5e38<\/td>\n<td>\u68c0\u67e5 rank \u65e5\u5fd7\u548c\u6570\u636e sampler<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5efa\u8bae\u8bad\u7ec3\u65e5\u5fd7\u81f3\u5c11\u8bb0\u5f55&#xff1a;<\/p>\n<p>1. \u6bcf\u6b65 loss \u548c learning rate&#xff1b;<br \/>\n2. tokens\/s \u548c samples\/s&#xff1b;<br \/>\n3. \u6bcf\u5f20 GPU \u663e\u5b58\u5cf0\u503c&#xff1b;<br \/>\n4. gradient norm&#xff1b;<br \/>\n5. ZeRO stage \u548c offload \u72b6\u6001&#xff1b;<br \/>\n6. checkpoint \u4fdd\u5b58\/\u6062\u590d\u8017\u65f6&#xff1b;<br \/>\n7. \u6570\u636e\u52a0\u8f7d\u8017\u65f6&#xff1b;<br \/>\n8. \u901a\u4fe1\u7b49\u5f85\u6216 GPU \u5229\u7528\u7387\u3002<\/p>\n<p>\u6ca1\u6709\u8fd9\u4e9b\u65e5\u5fd7&#xff0c;\u6392\u67e5 ZeRO \u95ee\u9898\u4f1a\u53d8\u6210\u731c\u914d\u7f6e\u3002<\/p>\n<hr \/>\n<h3>13. \u5e38\u89c1\u8bef\u533a<\/h3>\n<p>\u8bef\u533a\u4e00&#xff1a;ZeRO \u9636\u6bb5\u8d8a\u9ad8\u8d8a\u597d\u3002 ZeRO-3 \u66f4\u7701\u663e\u5b58&#xff0c;\u4f46\u901a\u4fe1\u590d\u6742\u3001\u541e\u5410\u53ef\u80fd\u4e0b\u964d\u3002\u663e\u5b58\u591f\u65f6&#xff0c;ZeRO-2 \u6216 DDP \u53ef\u80fd\u66f4\u597d\u3002<\/p>\n<p>\u8bef\u533a\u4e8c&#xff1a;\u7528\u4e86 DeepSpeed \u5c31\u4e0d\u7528\u7406\u89e3\u663e\u5b58\u3002 DeepSpeed \u5e2e\u4f60\u5207\u72b6\u6001&#xff0c;\u4f46\u4f60\u4ecd\u7136\u8981\u77e5\u9053\u53c2\u6570\u3001\u68af\u5ea6\u3001\u4f18\u5316\u5668\u3001\u6fc0\u6d3b\u503c\u5206\u522b\u5360\u591a\u5c11\u3002<\/p>\n<p>\u8bef\u533a\u4e09&#xff1a;Offload \u662f\u52a0\u901f\u624b\u6bb5\u3002 Offload \u4e3b\u8981\u662f\u7528\u901f\u5ea6\u6362\u5bb9\u91cf&#xff0c;\u901a\u5e38\u4f1a\u964d\u4f4e\u541e\u5410\u3002<\/p>\n<p>\u8bef\u533a\u56db&#xff1a;OOM \u4e00\u5b9a\u9760 ZeRO \u89e3\u51b3\u3002 OOM \u4e5f\u53ef\u80fd\u6765\u81ea seq_len\u3001micro batch\u3001\u6fc0\u6d3b\u503c\u3001\u8bc4\u4f30\u914d\u7f6e\u3001\u6570\u636e\u5f02\u5e38\u548c\u4e34\u65f6\u5f20\u91cf\u3002<\/p>\n<p>\u8bef\u533a\u4e94&#xff1a;\u8bad\u7ec3\u914d\u7f6e\u53ef\u4ee5\u76f4\u63a5\u590d\u5236\u3002 \u540c\u6837\u6a21\u578b\u5728\u4e0d\u540c GPU\u3001\u7f51\u7edc\u3001batch\u3001seq_len \u4e0b&#xff0c;\u6700\u4f73 ZeRO \u914d\u7f6e\u53ef\u80fd\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n<hr \/>\n<h3>14. \u5c0f\u7ed3<\/h3>\n<p>DeepSpeed ZeRO \u7684\u6838\u5fc3\u662f\u6d88\u9664\u6570\u636e\u5e76\u884c\u4e2d\u7684\u91cd\u590d\u8bad\u7ec3\u72b6\u6001\u3002ZeRO-1 \u5207\u4f18\u5316\u5668\u72b6\u6001&#xff0c;ZeRO-2 \u518d\u5207\u68af\u5ea6&#xff0c;ZeRO-3 \u8fde\u53c2\u6570\u4e5f\u5207\u3002\u5b83\u80fd\u8ba9\u66f4\u5927\u6a21\u578b\u5728\u6709\u9650\u663e\u5b58\u4e0b\u8bad\u7ec3&#xff0c;\u4f46\u4ee3\u4ef7\u662f\u901a\u4fe1\u3001checkpoint \u548c\u8c03\u8bd5\u590d\u6742\u5ea6\u4e0a\u5347\u3002\u5de5\u7a0b\u4e0a\u8981\u5148\u5efa\u7acb\u663e\u5b58\u8d26\u672c&#xff0c;\u518d\u6839\u636e\u74f6\u9888\u9009\u62e9\u9636\u6bb5&#xff0c;\u800c\u4e0d\u662f\u76f2\u76ee\u8ffd\u6c42\u6700\u9ad8\u9636\u6bb5\u3002<\/p>\n<h3>\u5927\u6a21\u578b\u89c6\u89d2<\/h3>\n<p>\u4ece\u4e13\u680f\u4e3b\u7ebf\u770b&#xff0c;\u524d\u9762\u51e0\u7bc7\u8bb2\u7684\u662f\u6570\u636e\u5982\u4f55\u8fdb\u5165\u8bad\u7ec3&#xff0c;\u8fd9\u4e00\u7bc7\u8bb2\u7684\u662f\u8bad\u7ec3\u72b6\u6001\u5982\u4f55\u653e\u8fdb\u786c\u4ef6\u3002ZeRO \u662f\u201c\u7528\u901a\u4fe1\u6362\u663e\u5b58\u201d\u7684\u5178\u578b\u5de5\u7a0b\u53d6\u820d&#xff0c;\u4e5f\u662f\u5927\u6a21\u578b\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\u548c\u6a21\u578b\u89c4\u6a21\u4e4b\u95f4\u6700\u91cd\u8981\u7684\u4e2d\u95f4\u5c42\u4e4b\u4e00\u3002<\/p>\n<h3>\u4e0b\u4e00\u7bc7<\/h3>\n<p>\u4e0b\u4e00\u7bc7\u4f1a\u7ee7\u7eed\u8ba8\u8bba\u8bad\u7ec3\u7a33\u5b9a\u6027&#xff1a;\u5373\u4f7f\u663e\u5b58\u95ee\u9898\u89e3\u51b3\u4e86&#xff0c;loss spike\u3001\u68af\u5ea6\u7206\u70b8\u3001\u6df7\u5408\u7cbe\u5ea6\u6ea2\u51fa\u548c\u574f batch \u4ecd\u7136\u53ef\u80fd\u8ba9\u8bad\u7ec3\u5931\u8d25\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u76ee\u5f551. \u5148\u628a\u95ee\u9898\u8bb2\u6e05\u695a&#xff1a;\u4e3a\u4ec0\u4e48\u201c\u80fd\u63a8\u7406\u201d\u4e0d\u7b49\u4e8e\u201c\u80fd\u8bad\u7ec3\u201d2. \u8bad\u7ec3\u663e\u5b58\u8d26\u672c&#xff1a;\u5230\u5e95\u54ea\u4e9b\u4e1c\u897f\u5360\u663e\u5b583. \u666e\u901a\u6570\u636e\u5e76\u884c\u4e3a\u4ec0\u4e48\u6d6a\u8d394. ZeRO-1&#xff1a;\u5148\u5207\u4f18\u5316\u5668\u72b6\u60015. ZeRO-2&#xff1a;\u4f18\u5316\u5668\u72b6\u6001\u548c\u68af\u5ea6\u90fd\u52076. ZeRO-3&#xff1a;\u8fde\u53c2\u6570\u4e5f\u52077. \u7528\u4ee3\u7801\u4f30\u7b97 ZeRO \u9636\u6bb5\u7684\u53c2\u6570\u72b6\u6001\u8282\u77018. DeepSpeed \u914d\u7f6e\u600e\u4e48\u8bfb9. ZeRO-3 \u914d\u7f6e\u548c\u51e0\u4e2a\u5173\u952e\u5f00\u517310. Offload&#xff1a;\u628a\u72b6\u6001\u653e\u5230 CPU <\/p>\n","protected":false},"author":2,"featured_media":116223,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[664,132],"topic":[],"class_list":["post-116225","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-664","tag-132"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6_DeepSpeed_ZeRO\u8be6\u89e3 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/116225.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6_DeepSpeed_ZeRO\u8be6\u89e3 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6587\u7ae0\u76ee\u5f551. \u5148\u628a\u95ee\u9898\u8bb2\u6e05\u695a&#xff1a;\u4e3a\u4ec0\u4e48\u201c\u80fd\u63a8\u7406\u201d\u4e0d\u7b49\u4e8e\u201c\u80fd\u8bad\u7ec3\u201d2. \u8bad\u7ec3\u663e\u5b58\u8d26\u672c&#xff1a;\u5230\u5e95\u54ea\u4e9b\u4e1c\u897f\u5360\u663e\u5b583. \u666e\u901a\u6570\u636e\u5e76\u884c\u4e3a\u4ec0\u4e48\u6d6a\u8d394. ZeRO-1&#xff1a;\u5148\u5207\u4f18\u5316\u5668\u72b6\u60015. ZeRO-2&#xff1a;\u4f18\u5316\u5668\u72b6\u6001\u548c\u68af\u5ea6\u90fd\u52076. ZeRO-3&#xff1a;\u8fde\u53c2\u6570\u4e5f\u52077. \u7528\u4ee3\u7801\u4f30\u7b97 ZeRO \u9636\u6bb5\u7684\u53c2\u6570\u72b6\u6001\u8282\u77018. DeepSpeed \u914d\u7f6e\u600e\u4e48\u8bfb9. ZeRO-3 \u914d\u7f6e\u548c\u51e0\u4e2a\u5173\u952e\u5f00\u517310. Offload&#xff1a;\u628a\u72b6\u6001\u653e\u5230 CPU\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/116225.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-20T15:08:50+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261020150846-6ad783fe4e804.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"5 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/116225.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/116225.html\",\"name\":\"\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6_DeepSpeed_ZeRO\u8be6\u89e3 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-20T15:08:50+00:00\",\"dateModified\":\"2026-10-20T15:08:50+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/116225.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/116225.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/116225.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6_DeepSpeed_ZeRO\u8be6\u89e3\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6_DeepSpeed_ZeRO\u8be6\u89e3 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/116225.html","og_locale":"zh_CN","og_type":"article","og_title":"\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6_DeepSpeed_ZeRO\u8be6\u89e3 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6587\u7ae0\u76ee\u5f551. \u5148\u628a\u95ee\u9898\u8bb2\u6e05\u695a&#xff1a;\u4e3a\u4ec0\u4e48\u201c\u80fd\u63a8\u7406\u201d\u4e0d\u7b49\u4e8e\u201c\u80fd\u8bad\u7ec3\u201d2. \u8bad\u7ec3\u663e\u5b58\u8d26\u672c&#xff1a;\u5230\u5e95\u54ea\u4e9b\u4e1c\u897f\u5360\u663e\u5b583. \u666e\u901a\u6570\u636e\u5e76\u884c\u4e3a\u4ec0\u4e48\u6d6a\u8d394. ZeRO-1&#xff1a;\u5148\u5207\u4f18\u5316\u5668\u72b6\u60015. ZeRO-2&#xff1a;\u4f18\u5316\u5668\u72b6\u6001\u548c\u68af\u5ea6\u90fd\u52076. ZeRO-3&#xff1a;\u8fde\u53c2\u6570\u4e5f\u52077. \u7528\u4ee3\u7801\u4f30\u7b97 ZeRO \u9636\u6bb5\u7684\u53c2\u6570\u72b6\u6001\u8282\u77018. DeepSpeed \u914d\u7f6e\u600e\u4e48\u8bfb9. ZeRO-3 \u914d\u7f6e\u548c\u51e0\u4e2a\u5173\u952e\u5f00\u517310. Offload&#xff1a;\u628a\u72b6\u6001\u653e\u5230 CPU","og_url":"https:\/\/www.wsisp.com\/helps\/116225.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-20T15:08:50+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261020150846-6ad783fe4e804.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"5 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/116225.html","url":"https:\/\/www.wsisp.com\/helps\/116225.html","name":"\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6_DeepSpeed_ZeRO\u8be6\u89e3 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-20T15:08:50+00:00","dateModified":"2026-10-20T15:08:50+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/116225.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/116225.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/116225.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6_DeepSpeed_ZeRO\u8be6\u89e3"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/116225","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=116225"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/116225\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/116223"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=116225"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=116225"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=116225"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=116225"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}