{"id":1949,"date":"2026-02-25T14:07:55","date_gmt":"2026-02-25T06:07:55","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/1949"},"modified":"2026-02-25T14:07:55","modified_gmt":"2026-02-25T06:07:55","slug":"%e4%bb%8e%e8%ae%ad%e7%bb%83%e5%88%b0%e6%8e%a8%e7%90%86%ef%bc%8cai%e5%b7%a5%e4%bd%9c%e6%b5%81%e5%85%a8%e9%9d%a2%e5%8a%a0%e9%80%9f%ef%bc%9a%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e6%b5%b7%e9%87%8f%e6%95%b0","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/1949","title":{"rendered":"\u4ece\u8bad\u7ec3\u5230\u63a8\u7406\uff0cAI\u5de5\u4f5c\u6d41\u5168\u9762\u52a0\u901f\uff1a\u661f\u5b87\u667a\u7b97\u6d77\u91cf\u6570\u636e\u96c6+\u9ad8\u901f\u5e76\u884c\u5b58\u50a8\u62c6\u89e3"},"content":{"rendered":"
\"\u4ece\u8bad\u7ec3\u5230\u63a8\u7406\uff0cAI\u5de5\u4f5c\u6d41\u5168\u9762\u52a0\u901f\uff1a\u661f\u5b87\u667a\u7b97\u6d77\u91cf\u6570\u636e\u96c6+\u9ad8\u901f\u5e76\u884c\u5b58\u50a8\u62c6\u89e3\"<\/figure>\n
\n

\u201c\u5f53\u6a21\u578b\u53c2\u6570\u7a81\u7834\u5343\u4ebf\uff0c\u8bad\u7ec3\u65f6\u95f4\u4ece\u6708\u7f29\u77ed\u5230\u5468\uff0c\u771f\u6b63\u7684\u6218\u573a\u5df2\u6084\u7136\u8f6c\u79fb\u5230\u63a8\u7406\u4fa7\u3002\u201d
\n\u2014\u20142024 \u5e74 5 \u6708\uff0c\u82f1\u4f1f\u8fbe GTC \u5927\u4f1a\u4e3b\u9898\u6f14\u8bb2<\/p>\n<\/blockquote>\n

\u5927\u6a21\u578b\u8bad\u7ec3\u540e\u65f6\u4ee3\uff0c\u63a8\u7406\u5de5\u4f5c\u6d41\u6210\u4e3a\u6027\u80fd\u74f6\u9888<\/h2>\n

\u8fc7\u53bb\u4e24\u5e74\uff0c\u5927\u6a21\u578b\u8bad\u7ec3\u4f9d\u9760\u5806 GPU\u3001\u62fc\u7b97\u529b\uff0c\u628a\u300c\u8bad\u7ec3\u65f6\u957f\u300d\u5377\u5230\u6781\u9650\u3002\u7136\u800c\uff0c\u5f53\u6a21\u578b\u8d70\u51fa\u5b9e\u9a8c\u5ba4\u3001\u8d70\u5411\u751f\u4ea7\u73af\u5883\uff0c\u5f00\u53d1\u8005\u53d1\u73b0\uff1a
\nCheckpoint \u52a8\u8f84\u4e0a\u767e GB\uff0c\u52a0\u8f7d\u4e00\u6b21\u5c31\u8981\u5341\u51e0\u5206\u949f\uff1b\u63a8\u7406\u8bf7\u6c42\u7a81\u53d1\u65f6\uff0c\u5e26\u5bbd\u88ab\u77ac\u95f4\u6253\u6ee1\uff0cP99 \u5ef6\u8fdf\u98d9\u5347\uff1b\u591a\u673a\u591a\u5361\u5e76\u884c\u63a8\u7406\uff0c\u6570\u636e\u5206\u7247\u53cd\u590d\u62f7\u8d1d\uff0cGPU \u7a7a\u8f6c\u7387\u9ad8\u8fbe 40%\u3002
\n\u4e00\u53e5\u8bdd\u2014\u2014\u8bad\u7ec3\u53ea\u662f\u4e0a\u534a\u573a\uff0c\u63a8\u7406\u624d\u662f\u51b3\u5b9a\u7528\u6237\u4f53\u9a8c\u4e0e\u6210\u672c\u7684\u751f\u6b7b\u6218<\/strong>\u3002 <\/p>\n

\u516c\u5f00+\u884c\u4e1a\u5b9a\u5236\u6570\u636e\u96c6\u955c\u50cf\uff0cTB\u7ea7\u6587\u672c\/\u56fe\u6587\u5bf9\u4e00\u952e\u6302\u8f7d<\/h2>\n

\u661f\u5b87\u667a\u7b97\u5728\u5e73\u53f0\u5c42\u7ed9\u51fa\u201c\u6570\u636e\u5148\u884c\u201d\u7684\u7b54\u6848\uff1a
\n1. \u516c\u5171\u8d44\u6e90\u5e93\u5df2\u5185\u7f6e Common Crawl\u3001LAION-5B\u3001RefinedWeb \u7b49 60+TB \u5f00\u6e90\u8bed\u6599<\/a>\uff0c\u65e0\u9700\u4e0b\u8f7d\uff0c\u5f00\u673a\u5373\u6302\u8f7d<\/strong>\uff1b
\n2. \u9488\u5bf9\u91d1\u878d\u3001\u533b\u7597\u3001\u6cd5\u5f8b\u4e09\u5927\u573a\u666f\uff0c\u63d0\u4f9b\u8131\u654f\u884c\u4e1a\u6570\u636e\u96c6\u955c\u50cf\uff0c\u652f\u6301\u53ea\u8bfb\u5feb\u7167\uff0c\u5408\u89c4\u4e0d\u51fa\u57df\uff1b
\n3. \u901a\u8fc7\u4e91\u786c\u76d8<\/a>\u4e0e\u4e91\u5b58\u50a8<\/a>\u53cc\u901a\u9053\uff0c\u7528\u6237\u53ef\u628a\u79c1\u6709\u6570\u636e\u79d2\u7ea7\u70ed\u63d2\u62d4\u5230\u4efb\u610f GPU\u4e91\u4e3b\u673a\uff0c\u8de8\u5b9e\u4f8b 0 \u6210\u672c\u8fc1\u79fb<\/strong>\u3002 <\/p>\n

\n

\u5b9e\u6d4b\uff1a\u5728\u661f\u5b87\u667a\u7b97 GPU\u670d\u52a1\u5668\u79df\u7528 \u5e73\u53f0\u542f\u52a8\u4e00\u53f0 8\u00d7A100 \u5b9e\u4f8b\uff0c\u6267\u884c cp -r \/public\/laion-5b\/part-001 ~\/data<\/code>\uff0c1.2 TB \u56fe\u6587\u5bf9 3 \u5206\u949f\u5b8c\u6210\u62f7\u8d1d\uff0c\u76f8\u5f53\u4e8e 6.7 GB\/s \u6301\u7eed\u541e\u5410\u3002 <\/p>\n<\/blockquote>\n

\u661f\u5b87\u667a\u7b97\u5e76\u884c\u6587\u4ef6\u7cfb\u7edf100Gbps RDMA\uff0cCheckpoint\u52a0\u8f7d\u901f\u5ea6\u63d0\u53475\u500d<\/h2>\n

\u63a8\u7406\u4fa7\u6700\u6015\u201c\u6362\u6a21\u578b\u201d\u2014\u2014\u5355\u6b21 Checkpoint \u52a0\u8f7d\u6162 1 \u5206\u949f\uff0c\u7ebf\u4e0a 1000 \u5e76\u53d1\u5c31\u7b49\u4e8e 1000 \u5206\u949f GPU \u7a7a\u8f6c\u3002
\n\u661f\u5b87\u667a\u7b97\u81ea\u7814\u5e76\u884c\u6587\u4ef6\u7cfb\u7edf\u57fa\u4e8e 100 Gbps RDMA \u7f51\u7edc\uff0c\u628a\u5143\u6570\u636e\u4e0e\u6570\u636e\u901a\u9053\u5206\u79bb\uff1a
\n– \u5355\u5ba2\u6237\u7aef\u5cf0\u503c 12 GB\/s\uff0c\u7ebf\u6027\u6269\u5c55\u81f3 20 \u8282\u70b9\u65f6\u603b\u5e26\u5bbd > 200 GB\/s\uff1b
\n– \u539f\u751f\u652f\u6301 PyTorch torch.load(..., mmap=True)<\/code>\uff0cGPU \u76f4\u63a5\u5185\u5b58\u6620\u5c04<\/strong>\uff0c\u7701\u6389 CPU \u62f7\u8d1d\u73af\u8282\uff1b
\n– \u4e0e GPU\u670d\u52a1\u5668\u79df\u7528<\/a> \u540c\u4e00 VPC \u5185\u4e92\u901a\uff0c\u5ef6\u8fdf\u4f4e\u4e8e 200 \u00b5s\uff0cCheckpoint 200 GB \u52a0\u8f7d\u8017\u65f6\u4ece 10 \u5206\u949f\u964d\u5230 110 \u79d2<\/strong>\uff0c\u771f\u6b63\u8ba9\u201c\u5f39\u6027\u6269\u7f29\u201d\u6210\u4e3a\u73b0\u5b9e\u3002 <\/p>\n

\u4ee3\u7801\u793a\u4f8b\uff1aPyTorch DataPipe+\u661f\u5b87\u667a\u7b97AI\u5e94\u7528\u6a21\u677f\uff0c30\u5206\u949f\u8dd1\u901a\u5fae\u8c03\u2192\u63a8\u7406\u2192\u4e0a\u7ebf<\/h2>\n

\u4e0b\u9762\u7528\u4e00\u4efd\u6700\u5c0f\u53ef\u8fd0\u884c\u4ee3\u7801\uff0c\u6f14\u793a\u5982\u4f55\u5728\u661f\u5b87\u667a\u7b97\u5e73\u53f0\u5b8c\u6210\u201c\u5c0f\u53c2\u6570\u6a21\u578b\u5fae\u8c03 + \u63a8\u7406\u670d\u52a1\u201d\u7aef\u5230\u7aef\u95ed\u73af\u3002 <\/p>\n

\u6b65\u9aa4 1\uff1a\u542f\u52a8\u5b9e\u4f8b<\/h3>\n

\u767b\u5f55 \u661f\u5b87\u667a\u7b97\u63a7\u5236\u53f0<\/a>\uff0c\u9009\u62e9\u300cAI \u5e94\u7528\u300d\u2192\u300cPyTorch 2.2\u300d\u955c\u50cf\uff0cGPU \u89c4\u683c 4\u00d7RTX 4090\uff0c\u65b0\u7528\u6237\u6ce8\u518c\u5373\u9001 10 \u5143\u4f53\u9a8c\u91d1<\/strong>\uff0c\u53ef\u8dd1\u901a\u5b8c\u6574\u793a\u4f8b\u3002 <\/p>\n

\u6b65\u9aa4 2\uff1a\u6570\u636e\u96c6\u4e00\u952e\u6302\u8f7d<\/h3>\n
# \u516c\u5171\u8d44\u6e90\u5df2\u6302\u8f7d\u81f3 \/public\uff0c\u76f4\u63a5\u8f6f\u94fe\u5373\u53ef\nln -s \/public\/refinedweb\/2024-06 ~\/data\/refinedweb\n<\/code><\/pre>\n
\u6b65\u9aa4 3\uff1a30 \u884c\u4ee3\u7801\u5fae\u8c03<\/h3>\n
# train.py\nimport torch, torchdata, os\nfrom transformers import AutoTokenizer, AutoModelForCausalLM\n\nmodel_name = "microsoft\/DialoGPT-medium"\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForCausalLM.from_pretrained(model_name).cuda()\n\n# \u4f7f\u7528\u661f\u5b87\u667a\u7b97 DataPipe \u52a0\u901f\u8bfb\u53d6\ndp = torchdata.datapipes.iter.FileLister("~\/data\/refinedweb")\ndp = dp.open_files(mode='rt').readlines().shuffle().batch(16)\n\nopt = torch.optim.AdamW(model.parameters(), 2e-5)\nfor epoch in range(1):\n    for batch in dp:\n        inputs = tokenizer(batch, return_tensors='pt', truncation=True, padding=True, max_length=512).to("cuda")\n        out = model(**inputs, labels=inputs.input_ids)\n        out.loss.backward()\n        opt.step(); opt.zero_grad()\n        if epoch == 0: break   # \u793a\u4f8b\u53ea\u8dd1 1 step\n\ntorch.save(model.state_dict(), "\/workspace\/chat-ckpt.pt")\n<\/code><\/pre>\n
\u6b65\u9aa4 4\uff1a\u5e76\u884c\u6587\u4ef6\u7cfb\u7edf\u79d2\u8f7d\u63a8\u7406<\/h3>\n
# infer.py\nimport time, torch\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\nmodel = AutoModelForCausalLM.from_pretrained("microsoft\/DialoGPT-medium").cuda()\n# \u661f\u5b87\u667a\u7b97 RDMA \u5e76\u884c\u6587\u4ef6\u7cfb\u7edf\uff0c110 \u79d2\u52a0\u8f7d 200 GB Checkpoint\nmodel.load_state_dict(torch.load("\/workspace\/chat-ckpt.pt", mmap=True))\nmodel.eval()\n\n# \u6784\u9020\u63a8\u7406\u670d\u52a1\nfrom flask import Flask, request, jsonify\napp = Flask(__name__)\n\n@app.route("\/chat", methods=["POST"])\ndef chat():\n    text = request.json["text"]\n    ids = tokenizer.encode(text, return_tensors='pt').cuda()\n    reply = model.generate(ids, max_length=60, pad_token_id=tokenizer.eos_token_id)\n    return jsonify({"reply": tokenizer.decode(reply[0], skip_special_tokens=True)})\n\napp.run(host="0.0.0.0", port=8000)\n<\/code><\/pre>\n
\u6b65\u9aa4 5\uff1a\u4e0a\u7ebf<\/h3>\n
\u63a7\u5236\u53f0\u6253\u5f00\u300c\u7aef\u53e3\u653e\u884c\u300d\u2192 8000\uff0c\u516c\u7f51\u53ef\u8c03\u7528\u7684 Chat API 30 \u5206\u949f\u5c31\u7eea<\/strong>\u3002  <\/p>\n
\u5199\u5728\u6700\u540e\uff1a\u628a\u201c\u7b97\u529b\u201d\u505a\u6210\u201c\u751f\u4ea7\u529b\u201d<\/h2>\n
\u4ece\u8bad\u7ec3\u5230\u63a8\u7406\uff0cAI \u5de5\u4f5c\u6d41\u50cf\u4e00\u6761\u9ad8\u901f\u73af\u8def\uff0c\u4efb\u4f55\u4e00\u6bb5\u9650\u901f\u90fd\u4f1a\u8ba9\u6574\u6761\u8def\u5835\u8f66\u3002
\n\u661f\u5b87\u667a\u7b97\u901a\u8fc7\u300cGPU\u670d\u52a1\u5668\u79df\u7528 + \u6d77\u91cf\u6570\u636e\u96c6 + 100Gbps RDMA \u5e76\u884c\u5b58\u50a8 + \u4e00\u952e AI \u5e94\u7528<\/strong>\u300d\u7684\u7ec4\u5408\u62f3\uff0c\u628a\u9650\u901f\u70b9\u9010\u4e00\u62c6\u6389\uff1a
\n– \u6570\u636e\u96c6\u4e0d\u7528\u4e0b\u8f7d\uff0cTB \u7ea7\u955c\u50cf\u79d2\u6302\u8f7d<\/strong>\uff1b
\n– Checkpoint \u52a0\u8f7d\u63d0\u901f 5 \u500d\uff0cGPU \u7a7a\u8f6c\u7387\u964d 40%<\/strong>\uff1b
\n– PyTorch DataPipe \u6a21\u677f\u76f4\u63a5\u8dd1\uff0c\u5fae\u8c03\u2192\u63a8\u7406\u2192\u4e0a\u7ebf 30 \u5206\u949f\u95ed\u73af<\/strong>\u3002  <\/p>\n
\u73b0\u5728\u6ce8\u518c \u661f\u5b87\u667a\u7b97<\/a>\uff0c\u5373\u53ef\u9886\u53d6 10 \u5143\u4f53\u9a8c\u91d1\uff0c0 \u6210\u672c\u8bd5\u9a7e RTX 4090 GPU\u4e91\u4e3b\u673a<\/strong>\uff0c\u4eb2\u624b\u4f53\u9a8c\u201c\u63a8\u7406\u98d9\u8f66\u201d\u7684\u5feb\u611f\u3002
\n\u8ba9\u7b97\u529b\u56de\u5f52\u521b\u65b0\uff0c\u628a\u7b49\u5f85\u7559\u7ed9\u5386\u53f2\u2014\u2014\u661f\u5b87\u667a\u7b97\uff0c\u4e0e\u4f60\u4e00\u8d77\u52a0\u901f AI \u7684\u4e0b\u4e00\u79d2\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u201c\u5f53\u6a21\u578b\u53c2\u6570\u7a81\u7834\u5343\u4ebf\uff0c\u8bad\u7ec3\u65f6\u95f4\u4ece\u6708\u7f29\u77ed\u5230\u5468\uff0c\u771f\u6b63\u7684\u6218\u573a\u5df2\u6084\u7136…<\/p>\n","protected":false},"author":2,"featured_media":1948,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-1949","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":48,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/1949","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/comments?post=1949"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/1949\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/1948"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=1949"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=1949"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=1949"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}