{"id":854,"date":"2023-04-09T17:25:00","date_gmt":"2023-04-09T08:25:00","guid":{"rendered":"https:\/\/staka.jp\/wordpress\/?p=854"},"modified":"2023-04-09T17:25:00","modified_gmt":"2023-04-09T08:25:00","slug":"pile%e3%81%ae%e6%96%87%e5%ad%97%e6%a7%8b%e6%88%90%ef%bc%88%e3%81%aa%e3%81%9ccerebras-gpt%e3%81%a7%e6%97%a5%e6%9c%ac%e8%aa%9e%e3%81%8c%e4%bd%bf%e3%81%88%e3%82%8b%e3%81%ae%e3%81%8b%ef%bc%9f%ef%bc%89","status":"publish","type":"post","link":"https:\/\/staka.jp\/wordpress\/?p=854","title":{"rendered":"The Pile\u306e\u69cb\u6210\uff08\u306a\u305cCerebras-GPT\u3067\u65e5\u672c\u8a9e\u304c\u4f7f\u3048\u308b\u306e\u304b\uff1f\uff09"},"content":{"rendered":"\n

ChatGPT\u304c\u76db\u308a\u4e0a\u304c\u308b\u4e2d\u3001\u30aa\u30fc\u30d7\u30f3\u30e9\u30a4\u30bb\u30f3\u30b9\u306aLLM\uff08\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb\uff09\u958b\u767a\u3082\u884c\u308f\u308c\u3066\u3044\u308b\u3002\u305d\u306e\u4e2d\u3067Cerebras-GPT\uff08Cerebras Systems Releases Seven New GPT Models Trained on CS-2 Wafer-Scale Systems – Cerebras<\/a>\uff09\u304c\u65e5\u672c\u8a9e\u3092\u4f7f\u3048\u308b\u3068\u805e\u3044\u3066\u8a66\u3057\u3066\u307f\u305f\u3002<\/p>\n\n\n\n

Colab Pro +\u3067A100 GPU\u3092\u4f7f\u3046\u3068\u3001cerebras\/Cerebras-GPT-6.7B \u00b7 Hugging Face<\/a>\u306b\u305d\u3063\u30666.7B\u306e\u30e2\u30c7\u30eb\u3092\u52d5\u4f5c\u3055\u305b\u308b\u3053\u3068\u304c\u3067\u304d\u308b[1]\u3002<\/p>\n\n\n\n

import torch\ndevice = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')\n\nfrom transformers import AutoTokenizer, AutoModelForCausalLM\nfrom transformers import pipeline\n\ntokenizer = AutoTokenizer.from_pretrained(\"cerebras\/Cerebras-GPT-6.7B\")\nmodel = AutoModelForCausalLM.from_pretrained(\"cerebras\/Cerebras-GPT-6.7B\")\n\npipe = pipeline(\"text-generation\", model=model, tokenizer=tokenizer, device=device)\n\ntext = \"\u732b\u3068\u66ae\u3089\u3057\u306610\u5e74\u306b\u306a\u308b\u4eba\u306b\u805e\u3044\u305f\u3001\u732b\u3068\u4e00\u7dd2\u306b\u66ae\u3089\u3059\u3068\u304d\u306b\u5fc5\u8981\u306a\u3082\u306e\u7b2c1\u4f4d\u306f\"\ngenerated_text = pipe(text, max_length=500, do_sample=False, no_repeat_ngram_size=2)[0]\nprint(generated_text['generated_text'])<\/code><\/pre>\n\n\n\n
\u4e0a\u8a18\u306e\u51fa\u529b\u306f\u300c\u732b\u3068\u66ae\u3089\u3057\u306610\u5e74\u306b\u306a\u308b\u4eba\u306b\u805e\u3044\u305f\u3001\u732b\u3068\u4e00\u7dd2\u306b\u66ae\u3089\u3059\u3068\u304d\u306b\u5fc5\u8981\u306a\u3082\u306e\u7b2c1\u4f4d\u306f\u300c\u7b11\u9854\u300d\u3002\u7d9a\u3044\u3066\u300c\u604b\u611b\u300d\u3001\u300c\u81ea\u5206\u306e\u3053\u3068\u3092\u77e5\u3063\u3066\u3044\u308b\u300d\u3068\u3044\u3046\u7d50\u679c\u306b\u3002<\/code>\u300d[2]\u3001\u306a\u304a\u30012.7B\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u5834\u5408\u306f\u300c\u732b\u3068\u66ae\u3089\u3057\u306610\u5e74\u306b\u306a\u308b\u4eba\u306b\u805e\u3044\u305f\u3001\u732b\u3068\u4e00\u7dd2\u306b\u66ae\u3089\u3059\u3068\u304d\u306b\u5fc5\u8981\u306a\u3082\u306e\u7b2c1\u4f4d\u306f\u3001\u305d\u308c\u305e\u308c\u306e\u4eba\u9593\u304c\u7b11\u3063\u3066\u3044\u308b\u3068\u3044\u3046\u3053\u3068\u3067\u3059\u3002<\/code>\u300d[2]\u3060\u3063\u305f\u3002<\/p>\n\n\n\n
model = AutoModelForCausalLM.from_pretrained(\"cerebras\/Cerebras-GPT-13B\", torch_dtype=torch.float16)<\/code><\/pre>\n\n\n\n
\u3068\u3057\u3066Cerebras-GPT-13B\u3092\u4f7f\u3063\u305f\u6642\u306e\u7d50\u679c\u306f\u300c\u732b\u3068\u66ae\u3089\u3057\u306610\u5e74\u306b\u306a\u308b\u4eba\u306b\u805e\u3044\u305f\u3001\u732b\u3068\u4e00\u7dd2\u306b\u66ae\u3089\u3059\u3068\u304d\u306b\u5fc5\u8981\u306a\u3082\u306e\u7b2c1\u4f4d\u306f\u3001\u300c\u7b11\u9854\u300d\u3002\u7d50\u679c\u3001\u305d\u308c\u306f\u79c1\u306e\u8a00\u8449\u3067\u306f\u306a\u304f\u3001\u81ea\u5206\u306e\u3082\u306e\u3067\u3042\u308b\u3002<\/code>\u300d[2]\u3060\u3063\u305f\u3002<\/p>\n\n\n\n
\u65e5\u672c\u8a9e\u3092\u4f7f\u3048\u308b\u3068\u8a00\u3063\u3066\u3082\u6027\u80fd\u306f\u304b\u306a\u308a\u5fae\u5999\u306a\u611f\u3058\u3002
\u203b\u4ee5\u964d\u3001\u7406\u7531\u3092\u63a2\u3063\u3066\u3044\u304f\u304c\u672c\u6765\u4f7f\u3048\u306a\u3044\u306f\u305a\u306e\u65e5\u672c\u8a9e\u3063\u307d\u3044\u6587\u5b57\u5217\u304c\u8868\u793a\u3055\u308c\u308b\u306e\u306f\u7d50\u69cb\u3059\u3054\u3044\u3068\u3044\u3046\u306e\u304c\u5168\u4f53\u7684\u306a\u611f\u60f3\u3002<\/p>\n\n\n\n
The Pile<\/h2>\n\n\n\n
Cerebras-GPT\u30b7\u30ea\u30fc\u30ba\u306f\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3068\u3057\u3066The Pile<\/a>[3]\u3092\u7528\u3044\u3066\u3044\u308b\u3002\u8ad6\u6587\u306b\u66f8\u304b\u308c\u3066\u3044\u308b\u901a\u308aThe Pile\u306f\u82f1\u8a9e\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3067\u3042\u308b\u3002\u305d\u306e\u305f\u3081\u57fa\u672c\u7684\u306b\u306f\u65e5\u672c\u8a9e\u3092\u4f7f\u3048\u306a\u3044\u306f\u305a\u306a\u306e\u3060\u304c\u3001The Pile\u3092\u7528\u3044\u3066\u5b66\u7fd2\u3055\u308c\u305fLLM\u306f\u65e5\u672c\u8a9e\u3092\u3042\u308b\u7a0b\u5ea6\u89e3\u91c8\u3057\u3066\u3044\u308b\u306e\u3067\u306f\uff1f\u3068\u601d\u3046\u52d5\u4f5c\u3092\u3059\u308b\u3002\u3053\u308c\u306fThe Pile\u69cb\u7bc9\u3067\u4f7f\u308f\u308c\u305f\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u610f\u56f3\u305b\u305a\u65e5\u672c\u8a9e\u304c\u542b\u307e\u308c\u3066\u3044\u308b\u305f\u3081\u3068\u8a00\u308f\u308c\u3066\u3044\u308b\u3002\u524d\u7f6e\u304d\u304c\u9577\u304f\u306a\u3063\u305f\u304c\u3001\u672c\u8a18\u4e8b\u3067\u306f\u3053\u308c\u304c\u672c\u5f53\u306a\u306e\u304b\u691c\u8a3c\u3059\u308b\u3002<\/strong><\/p>\n\n\n\n
\u691c\u8a3c\u65b9\u6cd5<\/h2>\n\n\n\n
The Pile\u3092\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u3057\u542b\u307e\u308c\u308b\u6587\u5b57\u3092\u6587\u5b57\u5225\u306b\u30ab\u30a6\u30f3\u30c8\u3057\u3001\u4e0b\u8a18\u6761\u4ef6\u306e\u6587\u5b57\u3092\u65e5\u672c\u8a9e\u3068\u8003\u3048\u3066\u65e5\u672c\u8a9e\u306e\u69cb\u6210\u6bd4\u7387\u3092\u7b97\u51fa\u3057\u305f\u3002CJK\u306e\u540d\u306e\u901a\u308a\u65e5\u672c\u8a9e\u3067\u306f\u306a\u3044\u7528\u4f8b\u306e\u6f22\u5b57\u3092\u542b\u3081\u3066\u6570\u3048\u3066\u304a\u308a\u660e\u3089\u304b\u306b\u65e5\u672c\u8a9e\u5206\u3092\u591a\u304f\u30ab\u30a6\u30f3\u30c8\u3057\u3066\u3044\u308b<\/strong>\u3068\u3044\u3046\u554f\u984c\u304c\u3042\u308b\u306e\u3060\u304c\u3001\u3053\u3053\u3067\u306f\u3044\u3063\u305f\u3093\u7121\u8996\u3059\u308b\u3053\u3068\u306b\u3059\u308b[4]\u3002<\/p>\n\n\n\n