{"id":4722,"date":"2024-04-18T22:49:14","date_gmt":"2024-04-18T22:49:14","guid":{"rendered":"https:\/\/game.intel.com\/?p=4722"},"modified":"2024-05-29T21:16:37","modified_gmt":"2024-05-29T21:16:37","slug":"wield-the-power-of-llms-on-intel-arc-gpus","status":"publish","type":"post","link":"https:\/\/game.intel.com\/jp\/stories\/wield-the-power-of-llms-on-intel-arc-gpus\/","title":{"rendered":"\u30a4\u30f3\u30c6\u30eb\u00ae Arc\u2122 GPU\u3067LLM\u306e\u30d1\u30ef\u30fc\u3092\u767a\u63ee\u3059\u308b"},"content":{"rendered":"
\n

\u30a4\u30f3\u30c6\u30eb\u00ae Arc\u2122 GPU\u3092\u4f7f\u7528\u3057\u3066\u3001\u3055\u307e\u3056\u307e\u306aLLM\u3092\u30ed\u30fc\u30ab\u30eb\u3067\u7c21\u5358\u306b\u5b9f\u884c\u53ef\u80fd<\/h3>\n<\/blockquote>\n\n\n\n

\u30b8\u30a7\u30cd\u30ec\u30fc\u30c6\u30a3\u30d6AI\u306f\u3001\u30b3\u30f3\u30c6\u30f3\u30c4\u5236\u4f5c\u306e\u53ef\u80fd\u6027\u3092\u5927\u304d\u304f\u5909\u3048\u305f\u3002\u3053\u306e\u30c6\u30af\u30ce\u30ed\u30b8\u30fc\u306f\u3001\u3053\u308c\u307e\u3067\u60f3\u50cf\u3082\u3067\u304d\u306a\u304b\u3063\u305f\u753b\u50cf\u3001\u52d5\u753b\u3001\u6587\u7ae0\u3092\u63d0\u4f9b\u3059\u308b\u53ef\u80fd\u6027\u3092\u79d8\u3081\u3066\u3044\u308b\u3002\u5927\u898f\u6a21\u306a\u8a00\u8a9e\u30e2\u30c7\u30eb\uff08LLM\uff09\u306fAI\u306e\u6642\u4ee3\u306b\u5927\u304d\u306a\u8a71\u984c\u3092\u547c\u3093\u3067\u304a\u308a\u3001\u8ab0\u3067\u3082\u6b4c\u306e\u6b4c\u8a5e\u3092\u751f\u6210\u3057\u305f\u308a\u3001\u8907\u96d1\u306a\u7269\u7406\u5b66\u306e\u8cea\u554f\u306b\u5bfe\u3059\u308b\u7b54\u3048\u3092\u5f97\u305f\u308a\u3001\u30b9\u30e9\u30a4\u30c9\u30fb\u30d7\u30ec\u30bc\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306e\u30a2\u30a6\u30c8\u30e9\u30a4\u30f3\u3092\u8d77\u8349\u3057\u305f\u308a\u3059\u308b\u65b9\u6cd5\u3092\u4fc3\u3059\u3053\u3068\u304c\u3067\u304d\u308b\u3002\u305d\u3057\u3066\u3053\u308c\u3089\u306eAI\u6a5f\u80fd\u306f\u3001\u3082\u306f\u3084\u30af\u30e9\u30a6\u30c9\u3084\u30b5\u30d6\u30b9\u30af\u30ea\u30d7\u30b7\u30e7\u30f3\u30fb\u30b5\u30fc\u30d3\u30b9\u306b\u63a5\u7d9a\u3059\u308b\u5fc5\u8981\u306f\u306a\u3044\u3002\u81ea\u5206\u306ePC\u4e0a\u3067\u30ed\u30fc\u30ab\u30eb\u306b\u52d5\u4f5c\u3055\u305b\u308b\u3053\u3068\u304c\u3067\u304d\u3001\u30e2\u30c7\u30eb\u3092\u5b8c\u5168\u306b\u5236\u5fa1\u3057\u3066\u51fa\u529b\u3092\u30ab\u30b9\u30bf\u30de\u30a4\u30ba\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u3002<\/p>\n\n\n\n

\u3053\u306e\u8a18\u4e8b\u3067\u306f\u3001Intel\u00ae Arc\u2122 A770 16GB\u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30b9\u30ab\u30fc\u30c9\u3092\u642d\u8f09\u3057\u305fPC\u3067\u3001\u4e00\u822c\u7684\u306a\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb\uff08LLM\uff09\u3092\u30bb\u30c3\u30c8\u30a2\u30c3\u30d7\u3057\u3066\u5b9f\u9a13\u3059\u308b\u65b9\u6cd5\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002\u3053\u306e\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb\u3067\u306fMistral-7B-Instruct LLM\u3092\u4f7f\u7528\u3057\u307e\u3059\u304c\u3001Phi2\u3084Llama2\u306a\u3069\u304a\u597d\u307f\u306ePyTorch LLM\u3067\u3082\u540c\u3058\u624b\u9806\u3067\u4f7f\u7528\u3067\u304d\u307e\u3059\u3002\u3082\u3061\u308d\u3093\u3001\u6700\u65b0\u306eLlama3\u30e2\u30c7\u30eb\u3067\u3082\u4f7f\u3048\u307e\u3059\uff01<\/p>\n\n\n\n

IPEX-LLM<\/h2>\n\n\n\n

\u540c\u3058\u30d9\u30fc\u30b9\u30fb\u30a4\u30f3\u30b9\u30c8\u30ec\u30fc\u30b7\u30e7\u30f3\u3092\u4f7f\u3063\u3066\u3055\u307e\u3056\u307e\u306a\u30e2\u30c7\u30eb\u3092\u8d70\u3089\u305b\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u306e\u306f\u3001\u6b21\u306e\u3088\u3046\u306a\u7406\u7531\u304b\u3089\u3060\u3002 IPEX-LLM<\/a>PyTorch \u7528\u306e LLM \u30e9\u30a4\u30d6\u30e9\u30ea\u3067\u3059\u3002\u3053\u308c\u306f PyTorch \u7528\u30a4\u30f3\u30c6\u30eb\u00ae \u30a8\u30af\u30b9\u30c6\u30f3\u30b7\u30e7\u30f3<\/a> IPEX-LLM\u306f\u3001\u6700\u65b0\u306eLLM\u6700\u9069\u5316\u3068\u30ed\u30fc\u30d3\u30c3\u30c8(INT4\/FP4\/INT8\/FP8)\u91cd\u307f\u5727\u7e2e\u3092\u542b\u307f\u3001Intel\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u306e\u305f\u3081\u306e\u3059\u3079\u3066\u306e\u6700\u65b0\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u6700\u9069\u5316\u3092\u5099\u3048\u3066\u3044\u307e\u3059\u3002IPEX-LLM\u306fXe<\/sup>-Arc A\u30b7\u30ea\u30fc\u30ba\u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30b9\u30ab\u30fc\u30c9\u306e\u3088\u3046\u306aIntel\u30c7\u30a3\u30b9\u30af\u30ea\u30fc\u30c8GPU\u4e0a\u3067XMX AI\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30b7\u30e7\u30f3\u3092\u30b3\u30a2\u3068\u3057\u3001\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u5411\u4e0a\u3002Windows Subsystem for Linux\u30d0\u30fc\u30b8\u30e7\u30f32\u3001\u30cd\u30a4\u30c6\u30a3\u30d6Windows\u74b0\u5883\u3001\u30cd\u30a4\u30c6\u30a3\u30d6Linux\u4e0a\u3067Intel Arc A\u30b7\u30ea\u30fc\u30ba\u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30b9\u3092\u30b5\u30dd\u30fc\u30c8\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

\u307e\u305f\u3001\u3053\u308c\u3089\u306f\u3059\u3079\u3066\u30cd\u30a4\u30c6\u30a3\u30d6\u306ePyTorch\u306a\u306e\u3067\u3001PyTorch\u30e2\u30c7\u30eb\u3068\u5165\u529b\u30c7\u30fc\u30bf\u3092\u7c21\u5358\u306b\u5165\u308c\u66ff\u3048\u3066\u3001\u9ad8\u6027\u80fd\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30b7\u30e7\u30f3\u3092\u5099\u3048\u305f\u30a4\u30f3\u30c6\u30ebArc GPU\u4e0a\u3067\u5b9f\u884c\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u3002\u3053\u306e\u5b9f\u9a13\u306f\u3001\u6027\u80fd\u6bd4\u8f03\u306a\u3057\u306b\u306f\u5b8c\u4e86\u3057\u307e\u305b\u3093\u3067\u3057\u305f\u3002\u4ee5\u4e0b\u306eIntel Arc\u7528\u306e\u30a4\u30f3\u30b9\u30c8\u30e9\u30af\u30b7\u30e7\u30f3\u3068\u3001\u7af6\u5408\u88fd\u54c1\u7528\u306e\u4e00\u822c\u7684\u306b\u5165\u624b\u53ef\u80fd\u306a\u30a4\u30f3\u30b9\u30c8\u30e9\u30af\u30b7\u30e7\u30f3\u3092\u4f7f\u7528\u3057\u3066\u3001\u540c\u3058\u3088\u3046\u306a\u4fa1\u683c\u30bb\u30b0\u30e1\u30f3\u30c8\u306b\u4f4d\u7f6e\u3059\u308b2\u3064\u306e\u30c7\u30a3\u30b9\u30af\u30ea\u30fc\u30c8GPU\u3092\u898b\u3066\u307f\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n\n\n\n

\u4f8b\u3048\u3070\u3001IPEX-LLM\u30e9\u30a4\u30d6\u30e9\u30ea\u3092\u4f7f\u7528\u3057\u3066Mistral 7B\u30e2\u30c7\u30eb\u3092\u5b9f\u884c\u3059\u308b\u5834\u5408\u3001Arc A770 16GB\u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30ab\u30fc\u30c9\u306f\u3001CUDA\u3092\u4f7f\u7528\u3057\u3066GeForce RTX 4060 8GB\u3088\u308a\u3082\u6bce\u79d270\u30c8\u30fc\u30af\u30f3\uff08TPS\uff09\u3001\u3064\u307e\u308a70%\u591a\u3044TPS\u3092\u51e6\u7406\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002\u3053\u308c\u306f\u4f55\u3092\u610f\u5473\u3059\u308b\u306e\u3067\u3057\u3087\u3046\u304b\uff1f\u4e00\u822c\u7684\u306a\u7d4c\u9a13\u5247\u3067\u306f\u30011\u30c8\u30fc\u30af\u30f3\u306f1\u5358\u8a9e\u306e0.75\u306b\u76f8\u5f53\u3057\u3001\u826f\u3044\u6bd4\u8f03\u5bfe\u8c61\u306f \u4eba\u9593\u306e\u5e73\u5747\u8aad\u66f8\u901f\u5ea6\u306f1\u79d2\u9593\u306b4\u8a9e<\/a> \u307e\u305f\u306f5.3 TPS\u3067\u3059\u3002Arc A770 16GB\u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30ab\u30fc\u30c9\u306f\u3001\u4e00\u822c\u4eba\u304c\u6587\u5b57\u3092\u8aad\u3080\u3088\u308a\u3082\u306f\u308b\u304b\u306b\u901f\u304f\u6587\u5b57\u3092\u751f\u6210\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u308b\uff01<\/p>\n\n\n\n

\u5f53\u793e\u306e\u5185\u90e8\u30c6\u30b9\u30c8\u306b\u3088\u308b\u3068\u3001Arc A770 16GB\u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30b9\u30ab\u30fc\u30c9\u306f\u3001RTX 4060\u3068\u6bd4\u8f03\u3057\u3066\u3001\u5e45\u5e83\u3044\u30e2\u30c7\u30eb\u3067\u3053\u306e\u80fd\u529b\u3092\u767a\u63ee\u3057\u3001\u7af6\u4e89\u529b\u306e\u3042\u308b\u3001\u3042\u308b\u3044\u306f\u30c8\u30c3\u30d7\u30af\u30e9\u30b9\u306e\u6027\u80fd\u3092\u767a\u63ee\u3067\u304d\u308b\u3053\u3068\u304c\u5206\u304b\u3063\u3066\u304a\u308a\u3001Intel Arc\u30b0\u30e9\u30d5\u30a3\u30c3\u30af\u30b9\u306f\u3001\u30ed\u30fc\u30ab\u30ebLLM\u5b9f\u884c\u306b\u6700\u9069\u306a\u9078\u629e\u80a2\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u305d\u308c\u3067\u306f\u3001Arc A\u30b7\u30ea\u30fc\u30baGPU\u3067LLM\u3092\u59cb\u3081\u308b\u305f\u3081\u306e\u30bb\u30c3\u30c8\u30a2\u30c3\u30d7\u624b\u9806\u3092\u8aac\u660e\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

\u30a4\u30f3\u30b9\u30c8\u30ec\u30fc\u30b7\u30e7\u30f3\u30fb\u30a4\u30f3\u30b9\u30c8\u30e9\u30af\u30b7\u30e7\u30f3<\/h2>\n\n\n\n

\u74b0\u5883\u69cb\u7bc9\u306b\u3064\u3044\u3066\u306f\u3001\u3053\u3061\u3089\u306e\u30da\u30fc\u30b8\u3082\u53c2\u8003\u306b\u3057\u3066\u307b\u3057\u3044\uff1a \u30a4\u30f3\u30c6\u30ebGPU\u3092\u642d\u8f09\u3057\u305fWindows\u306bIPEX-LLM\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3059\u308b - IPEX-LLM\u6700\u65b0\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8<\/a><\/p>\n\n\n\n

1.\u30c7\u30d0\u30a4\u30b9\u30de\u30cd\u30fc\u30b8\u30e3\u3067\u7d71\u5408GPU\u3092\u7121\u52b9\u306b\u3059\u308b\u3002<\/p>\n\n\n\n

2.\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u3068\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb \u30a2\u30ca\u30b3\u30f3\u30c0<\/a>.<\/p>\n\n\n\n

3.\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u304c\u5b8c\u4e86\u3057\u305f\u3089\u3001\u30b9\u30bf\u30fc\u30c8\u30e1\u30cb\u30e5\u30fc\u3092\u958b\u304d\u3001Anaconda Prompt \u3092\u691c\u7d22\u3057\u3001\u7ba1\u7406\u8005\u3068\u3057\u3066\u5b9f\u884c\u3057\u3001\u4ee5\u4e0b\u306e\u30b3\u30de\u30f3\u30c9\u3092\u4f7f\u7528\u3057\u3066\u4eee\u60f3\u74b0\u5883\u3092\u4f5c\u6210\u3057\u307e\u3059\u3002\u5404\u30b3\u30de\u30f3\u30c9\u306f\u5225\u3005\u306b\u5165\u529b\u3057\u307e\u3059\uff1a<\/p>\n\n\n\n

conda create -n llm python=3.10.6\n\nllm \u3092\u8d77\u52d5\u3059\u308b\n\nlibuv \u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3059\u308b\n\npip install dpcpp-cpp-rt==2024.0.2 mkl-dpcpp==2024.0.0 onednn==2024.0.0 gradio\n\npip install --pre --upgrade ipex-llm[xpu] --extra-index-url https:\/\/pytorch-extension.intel.com\/release-whl\/stable\/xpu\/us\/\n\npip install transformers==4.38.0<\/code><\/pre>\n\n\n\n
4.4.demo.py\u3068\u3044\u3046\u540d\u524d\u306e\u30c6\u30ad\u30b9\u30c8\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3092\u4f5c\u6210\u3057\u3001C:\u21c4UsersYour_Username\u21c4Documents\u307e\u305f\u306f\u4efb\u610f\u306e\u30c7\u30a3\u30ec\u30af\u30c8\u30ea\u306b\u4fdd\u5b58\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
5.\u597d\u304d\u306a\u30a8\u30c7\u30a3\u30bf\u3067demo.py\u3092\u958b\u304d\u3001\u4ee5\u4e0b\u306e\u30b3\u30fc\u30c9\u30b5\u30f3\u30d7\u30eb\u3092\u30b3\u30d4\u30fc\u3057\u3066\u304f\u3060\u3055\u3044\uff1a<\/p>\n\n\n\n
from transformers import AutoTokenizer\nfrom ipex_llm.transformers import AutoModelForCausalLM\n\u30a4\u30f3\u30dd\u30fc\u30c8\u30c8\u30fc\u30c1\n\u30a4\u30f3\u30dd\u30fc\u30c8 intel_extension_for_pytorch\n\ndevice = \"xpu\" #\u30e2\u30c7\u30eb\u3092\u30ed\u30fc\u30c9\u3059\u308b\u30c7\u30d0\u30a4\u30b9\u3002\n\nmodel_id = \"mistralai\/Mistral-7B-Instruct-v0.2\" #\u62b1\u304d\u3064\u304d\u9854\u306e\u30e2\u30c7\u30ebID\n\ntokenizer = AutoTokenizer.from_pretrained(model_id)\nmodel = AutoModelForCausalLM.from_pretrained(model_id, load_in_4bit=True, torch_dtype=torch.float16)\nmodel = model.to(\u30c7\u30d0\u30a4\u30b9)\n\n\u30e1\u30c3\u30bb\u30fc\u30b8 = [\n    {\"role\"\uff1a\"user\", \"content\"\uff1a\u300c\u597d\u304d\u306a\u8abf\u5473\u6599\u306f\u4f55\u3067\u3059\u304b\uff1f\uff5d\n    {\"role\"\uff1a\"assistant\", \"content\"\uff1a\"\u307e\u3042\u3001\u79c1\u306f\u65b0\u9bae\u306a\u30ec\u30e2\u30f3\u6c41\u3092\u3088\u304f\u7d5e\u308b\u306e\u304c\u597d\u304d\u3067\u3059\u306d\u3002\u30ad\u30c3\u30c1\u30f3\u3067\u3069\u3093\u306a\u6599\u7406\u3092\u4f5c\u308b\u3068\u304d\u3067\u3082\u3001\u30d4\u30ea\u30c3\u3068\u3057\u305f\u98a8\u5473\u3092\u30d7\u30e9\u30b9\u3057\u3066\u304f\u308c\u308b\u3093\u3067\u3059\uff01\"}\u3001\n    {\"role\"\uff1a\"user\", \"content\"\uff1a\u300c\u30de\u30e8\u30cd\u30fc\u30ba\u306e\u30ec\u30b7\u30d4\u306f\u3042\u308a\u307e\u3059\u304b\uff1f\uff5d\n]\n\nencodeds = tokenizer.apply_chat_template(messages, return_tensors=\"pt\")\n\nmodel_inputs = encodeds.to(\u30c7\u30d0\u30a4\u30b9)\nmodel.to(device)\n\ngenerated_ids = model.generate(model_inputs, max_new_tokens=1000, do_sample=True)\ndecoded = tokenizer.batch_decode(generated_ids)\nprint(decoded[0])<\/code><\/pre>\n\n\n\n
\u30b5\u30f3\u30d7\u30eb\u30b3\u30fc\u30c9\u304b\u3089\u30d3\u30eb\u30c9\u3057\u305f\u30b3\u30fc\u30c9 \u3053\u306e\u30ea\u30dd\u30b8\u30c8\u30ea<\/a>.<\/em><\/p>\n\n\n\n
6.demo.py \u3092\u4fdd\u5b58\u3057\u307e\u3059\u3002Anaconda \u3067\u3001cd \u30b3\u30de\u30f3\u30c9\u3092\u4f7f\u3063\u3066 demo.py \u304c\u3042\u308b\u30c7\u30a3\u30ec\u30af\u30c8\u30ea\u306b\u79fb\u52d5\u3057\u3001Anaconda \u30d7\u30ed\u30f3\u30d7\u30c8\u3067\u4ee5\u4e0b\u306e\u30b3\u30de\u30f3\u30c9\u3092\u5b9f\u884c\u3057\u307e\u3059\uff1a<\/p>\n\n\n\n
python demo.py<\/code><\/pre>\n\n\n\n
\u30de\u30e8\u30cd\u30fc\u30ba\u4f5c\u308a\u306e\u7d20\u6575\u306a\u30ec\u30b7\u30d4\u304c\u624b\u306b\u5165\u308b\uff01<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u30e2\u30c7\u30eb\u30c1\u30a7\u30f3\u30b8<\/h2>\n\n\n\n
\u4e0a\u3067\u8a2d\u5b9a\u3057\u305f\u306e\u3068\u540c\u3058\u74b0\u5883\u3092\u4f7f\u3044\u3001demo.py\u3067\u30cf\u30ae\u30f3\u30b0\u30fb\u30d5\u30a7\u30a4\u30b9\u306e\u30e2\u30c7\u30ebid\u3092\u7f6e\u304d\u63db\u3048\u308b\u3053\u3068\u3067\u3001llama2-7B-chat-hf, llama3-8B-it, phi-2, gemma-7B-i, stablelm2 \u306a\u3069\u3001\u30cf\u30ae\u30f3\u30b0\u30fb\u30d5\u30a7\u30a4\u30b9\u3067\u4eba\u6c17\u306e\u3042\u308b\u4ed6\u306e\u30e2\u30c7\u30eb\u3092\u8a66\u3059\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
model_id = \"mistralai\/Mistral-7B-Instruct-v0.2\" #\u30cf\u30ae\u30f3\u30b0\u30d5\u30a7\u30a4\u30b9\u306e\u30e2\u30c7\u30ebID\n\n\u306b\n\nmodel_id = \"stabilityai\/stablelm-2-zephyr-1_6b\" # huggingface \u30e2\u30c7\u30ebID<\/code><\/pre>\n\n\n\n
\u3082\u3057 demo.py \u306e\u8d77\u52d5\u6642\u306b\u30a8\u30e9\u30fc\u304c\u767a\u751f\u3057\u305f\u5834\u5408\u306f\u3001\u4ee5\u4e0b\u306e\u624b\u9806\u3067 transformers \u3092\u30a2\u30c3\u30d7\u30b0\u30ec\u30fc\u30c9\/\u30c0\u30a6\u30f3\u30b0\u30ec\u30fc\u30c9\u3057\u3066\u304f\u3060\u3055\u3044\uff1a<\/p>\n\n\n\n
    \n
  1. \u30a2\u30ca\u30b3\u30f3\u30c0\u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u958b\u304f<\/li>\n\n\n\n
  2. \u30b3\u30f3\u30c0\u30fb\u30a2\u30af\u30c6\u30a3\u30d9\u30fc\u30c8llm<\/li>\n\n\n\n
  3. pip install transformers==4.37.0<\/li>\n<\/ol>\n\n\n\n
    \u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc\u306e\u30d0\u30fc\u30b8\u30e7\u30f3\u3092\u78ba\u8a8d\uff1a<\/strong><\/p>\n\n\n\n
    \u30e2\u30c7\u30ebID<\/th>\u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc \u30d1\u30c3\u30b1\u30fc\u30b8\u7248<\/th><\/tr><\/thead>
    \u30e1\u30bf\u30e9\u30de\/\u30e9\u30de2-7b-\u30c1\u30e3\u30c3\u30c8-hf<\/td>4.37.0<\/td><\/tr>
    \u30e1\u30bf\u30e9\u30de\/\u30e1\u30bf\u30e9\u30de-3-8B-\u30a4\u30f3\u30b9\u30c8\u30e9\u30af\u30c8<\/td>4.37.0<\/td><\/tr>
    \u5b89\u5b9aai\/\u5b89\u5b9alm-2-\u30bc\u30d5\u30a1\u30fc-1_6b<\/td>4.38.0<\/td><\/tr>
    mistralai\/Mistral-7B-Instruct-v0.2<\/td>4.38.0<\/td><\/tr>
    \u30de\u30a4\u30af\u30ed\u30bd\u30d5\u30c8\/\u30d5\u30a1\u30a42<\/td>4.38.0<\/td><\/tr>
    google\/gemma-7b-it<\/td>4.38.1<\/td><\/tr>
    THUDM\/chatglm3-6b<\/td>4.38.0<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    \u30e1\u30e2\u30ea\u8981\u4ef6\u306f\u30e2\u30c7\u30eb\u3084\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u306b\u3088\u3063\u3066\u7570\u306a\u308b\u5834\u5408\u304c\u3042\u308a\u307e\u3059\u3002IPEX-LLM \u3067\u52d5\u4f5c\u3059\u308b Intel Arc A750 8GB \u3067\u306f\u3001Llama-2-7B-chat-hf, Mistral-7B-Instruct-v0.2, phi-2 \u307e\u305f\u306f chatglm3-6B \u306e\u4f7f\u7528\u3092\u304a\u52e7\u3081\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
    ChatBot WebUI\u306e\u5b9f\u88c5<\/h2>\n\n\n\n
    \u305d\u308c\u3067\u306f\u3001Gradio\u30c1\u30e3\u30c3\u30c8\u30dc\u30c3\u30c8\u306e\u30a6\u30a7\u30d6\u30a4\u3092\u5b9f\u88c5\u3057\u3066\u3001\u30a6\u30a7\u30d6\u30d6\u30e9\u30a6\u30b6\u3092\u4f7f\u3063\u3066\u3088\u308a\u826f\u3044\u4f53\u9a13\u3092\u63d0\u4f9b\u3057\u307e\u3057\u3087\u3046\u3002LLM\u3068\u306e\u5bfe\u8a71\u578b\u30c1\u30e3\u30c3\u30c8\u30dc\u30c3\u30c8\u306e\u5b9f\u88c5\u306b\u95a2\u3059\u308b\u8a73\u7d30\u306f\u3001\u4ee5\u4e0b\u3092\u3054\u89a7\u304f\u3060\u3055\u3044\u3002 https:\/\/www.gradio.app\/guides\/creating-a-chatbot-fast<\/a><\/p>\n\n\n\n
    1.\u597d\u304d\u306a\u30c6\u30ad\u30b9\u30c8\u30a8\u30c7\u30a3\u30bf\u3067 chatbot_gradio.py \u3068\u3044\u3046\u540d\u524d\u306e\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3092\u4f5c\u6210\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
    2.\u4ee5\u4e0b\u306e\u30b3\u30fc\u30c9\u30b9\u30cb\u30da\u30c3\u30c8\u3092\u30b3\u30d4\u30fc\u3057\u3066chatbot_gradio.py\u306b\u8cbc\u308a\u4ed8\u3051\u307e\u3059\uff1a<\/p>\n\n\n\n
    gradio\u3092gr\u3068\u3057\u3066\u30a4\u30f3\u30dd\u30fc\u30c8\u3059\u308b\n\u30a4\u30f3\u30dd\u30fc\u30c8\u30c8\u30fc\u30c1\n\u30a4\u30f3\u30dd\u30fc\u30c8 intel_extension_for_pytorch\nfrom ipex_llm.transformers import AutoModelForCausalLM\nfrom transformers import AutoTokenizer, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer\nfrom threading import Thread\n\nmodel_id = \"mistralai\/Mistral-7B-Instruct-v0.2\"\n\ntokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)\nmodel = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, optimize_model=True, load_in_4bit=True, torch_dtype=torch.float16)\nmodel = model.half()\nmodel = model.to(\"xpu\")\nclass StopOnTokens(StoppingCriteria)\uff1a\n    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool\uff1a\n        stop_ids = [29, 0\uff3d\n        for stop_id in stop_ids\uff1a\n            if input_ids[0][-1] == stop_id\uff1a\n                return True\n        \u507d\u3092\u8fd4\u3059\n\ndef predict(message, history)\uff1a\n    stop = StopOnTokens()\n    history_format = [].\n    for human, assistant in history\uff1a\n        history_format.append({\"role\": \"user\", \"content\": human })\n        history_format.append({\"role\": \"assistant\", \"content\":assistant})\n    history_format.append({\"role\": \"user\", \"content\": \u30e1\u30c3\u30bb\u30fc\u30b8})\n\n    prompt = tokenizer.apply_chat_template(history_format, tokenize=False, add_generation_prompt=True)\n    model_inputs = tokenizer(prompt, return_tensors=\"pt\").to(\"xpu\")\n    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)\n    generate_kwargs = dict(\n        model_inputs\u3001\n        streamer=streamer\u3001\n        max_new_tokens=300\u3001\n        do_sample=True\u3001\n        top_p=0.95\u3001\n        top_k=20\u3001\n        temperature=0.8\u3001\n        num_beams=1\u3001\n        pad_token_id=tokenizer.eos_token_id\u3001\n        stopping_criteria=StoppingCriteriaList([stop])\n        )\n    t = Thread(target=model.generate, kwargs=generate_kwargs)\n    t.start()\n\n    partial_message = \"\"\n    for new_token in streamer\uff1a\n        if new_token != '<'\uff1a\n            partial_message += new_token\n            yield partial_message\n\ngr.ChatInterface(predict).launch()<\/code><\/pre>\n\n\n\n
    3.\u65b0\u3057\u3044 anaconda \u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u958b\u304d\u3001\u4ee5\u4e0b\u306e\u30b3\u30de\u30f3\u30c9\u3092\u5165\u529b\u3059\u308b\uff1a<\/p>\n\n\n\n