\u63a8\u8ad6\u30b3\u30b9\u30c8\u6700\u9069\u5316\u30d7\u30ec\u30a4\u30d6\u30c3\u30af<\/title>\n<style>\n body{font-family: -apple-system,BlinkMacSystemFont,Segoe UI,Roboto,Helvetica,Arial,\"Hiragino Kaku Gothic ProN\",\"Hiragino Sans\",\"Noto Sans CJK JP\",\"Yu Gothic UI\",\"YuGothic\",Meiryo,sans-serif; line-height:1.65; margin:40px; color:#111;}\n code, pre{font-family: ui-monospace,SFMono-Regular,Menlo,Monaco,Consolas,\"Liberation Mono\",\"Courier New\",monospace; font-size: 0.95em;}\n pre{background:#f6f8fa; padding:12px 16px; border-radius:8px; overflow:auto; border:1px solid #eaecef;}\n h1,h2,h3{line-height:1.35}\n table{border-collapse:collapse; margin:16px 0; width:100%}\n th,td{border:1px solid #ddd; padding:8px; vertical-align:top}\n th{background:#fafafa; text-align:left}\n .ok{color:#0a7}\n .warn{color:#e57}\n .mono{font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, \"Courier New\", monospace}\n<\/style>\n<\/head>\n<body>\n<header>\n \n <p class=\"lead\">\u524d\u63d0\uff1a\u30aa\u30f3\u30d7\u30ec\uff08\u4f8b\uff1aRTX 4070 Ti SUPER \u00d71\uff09\u307e\u305f\u306f\u30af\u30e9\u30a6\u30c9\u306e\u6df7\u5728\u3092\u60f3\u5b9a\u3002CUDA \u3060\u3051\u306b\u7e1b\u3089\u308c\u305a\u3001ROCm \/ OpenVINO \/ ONNX Runtime \/ TensorRT \/ Triton \/ vLLM \u7b49\u3092\u9069\u6750\u9069\u6240\u3067\u4f7f\u3044\u5206\u3051\u308b\u3002\u3053\u3053\u306b\u3042\u308b\u306e\u306f\u8abf\u67fb\u4e2d\u5185\u5bb9\u306e\u30e1\u30e2\u3067\u3059\u3002<\/p>\n<\/header>\n\n<main>\n<section>\n <h2>0. \u307e\u305a\u6c7a\u3081\u308b\u3053\u3068\uff083\u5206\u30c1\u30a7\u30c3\u30af\u30ea\u30b9\u30c8\uff09<\/h2>\n <div class=\"grid cols-2\">\n <div class=\"ok\">\n <strong>\u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u306e\u578b<\/strong>\n <ul>\n <li>\u753b\u50cf\u89e3\u6790\uff1a\u5206\u985e \/ \u691c\u51fa\uff08YOLO\/RT-DETR\u7cfb\uff09\/ \u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3<\/li>\n <li>\u751f\u6210\uff1a\u62e1\u6563\uff08SD\u7cfb\uff09\/ \u30c6\u30ad\u30b9\u30c8\u2192\u753b\u50cf \/ \u753b\u50cf\u2192\u753b\u50cf<\/li>\n <li>OCR\uff1a\u30ec\u30a4\u30a2\u30a6\u30c8\u89e3\u6790 + \u8a8d\u8b58\uff08PaddleOCR \u7b49\uff09<\/li>\n <li>\u691c\u7d22\uff1a\u57cb\u3081\u8fbc\u307f\u751f\u6210 + ANN \u691c\u7d22 + \uff08\u4efb\u610f\uff09\u518d\u30e9\u30f3\u30ad\u30f3\u30b0<\/li>\n <\/ul>\n <\/div>\n <div class=\"ok\">\n <strong>SLO\/\u5236\u7d04<\/strong>\n <ul>\n <li>\u30ec\u30a4\u30c6\u30f3\u30b7\uff08p95 ms\uff09\u30fb\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8\uff08req\/s\uff09\u30fbQoS\uff08Drop\u8a31\u5bb9\uff09<\/li>\n <li>\u7cbe\u5ea6\u3057\u304d\u3044\u5024\uff08mAP \/ F1 \/ CER \/ nDCG\uff09<\/li>\n <li>\u30b3\u30b9\u30c8\u4e0a\u9650\uff08\u5186\/1000 \u30ea\u30af\u30a8\u30b9\u30c8 or \u5186\/\u6642\uff09<\/li>\n <\/ul>\n <\/div>\n <div class=\"ok\">\n <strong>\u914d\u5099\u5148<\/strong>\n <ul>\n <li>CUDA\uff08NVIDIA\uff09\/ ROCm\uff08AMD\uff09\/ OpenVINO\uff08Intel CPU+iGPU\uff09<\/li>\n <li>\u30af\u30e9\u30a6\u30c9\u5c02\u7528 HW\uff08AWS Inferentia2 \/ TPU v5e \u306a\u3069\uff09<\/li>\n <\/ul>\n <\/div>\n <div class=\"ok\">\n <strong>\u30e2\u30c7\u30eb\u9078\u5b9a<\/strong>\n <ul>\n <li>\u84b8\u7559\u6e08\u307f\u30fb\u8efd\u91cf\u7cfb\uff08\u4f8b\uff1aYOLOv8n\/s, MobileSAM, Distil-Embedding\uff09<\/li>\n <li>\u91cf\u5b50\u5316\u524d\u63d0\uff08FP16\/INT8\/INT4\uff09\u3067\u518d\u5b66\u7fd2\u8a31\u5bb9\u304b<\/li>\n <\/ul>\n <\/div>\n <\/div>\n<\/section>\n\n<section>\n <h2>1. \u8abf\u9054\u6226\u7565\uff08\u30cf\u30fc\u30c9\u5225\u30fb\u8cbb\u7528\u5bfe\u52b9\u679c\u306e\u52d8\u6240\uff09<\/h2>\n <table>\n <thead><tr><th>\u533a\u5206<\/th><th>\u9069\u6027\u30ef\u30fc\u30af\u30ed\u30fc\u30c9<\/th><th>\u5f37\u307f<\/th><th>\u843d\u3068\u3057\u7a74<\/th><th>\u3056\u3063\u304f\u308a\u6307\u91dd<\/th><\/tr><\/thead>\n <tbody>\n <tr>\n <td><strong>NVIDIA\uff08CUDA\/TensorRT\uff09<\/strong><\/td>\n <td>\u753b\u50cf\u89e3\u6790\/\u751f\u6210, \u5927\u898f\u6a21LLM\u63a8\u8ad6, \u6df7\u5728<\/td>\n <td>\u6210\u719f\u30c4\u30fc\u30eb\u7fa4\uff08TensorRT, Triton, cuDNN\uff09\u3068\u9ad8\u52b9\u7387 FP16\/INT8<\/td>\n <td>\u88c5\u7f6e\u5358\u4fa1\u30fb\u6d88\u8cbb\u96fb\u529b\u9ad8\u3081\u3001\u5bfe\u4e2d\u8f38\u51fa\u898f\u5236\u306e\u5f71\u97ff<\/td>\n <td>\u5b66\u7fd2\u3082\u8996\u91ce\u306a\u3089\u7b2c\u4e00\u5019\u88dc\u3002\u63a8\u8ad6\u5c02\u7528\u306a\u3089 L4\/L40S\/RTX 4000 SFF \u3082\u691c\u8a0e<\/td>\n <\/tr>\n <tr>\n <td><strong>AMD\uff08ROCm\uff09<\/strong><\/td>\n <td>\u63a8\u8ad6\u5168\u822c\uff08\u7279\u306b\u753b\u50cf\/\u691c\u7d22\/\u4e00\u90e8LLM\uff09<\/td>\n <td>\u4fa1\u683c\u6027\u80fd\u6bd4\/\u5165\u624b\u6027\u3001ONNX\/Triton\/torch.compile \u3082\u6d3b\u7528\u53ef<\/td>\n <td>\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u9069\u5408\u6027\u5dee\u30fb\u4e00\u90e8\u30e2\u30c7\u30eb\u306e\u6700\u9069\u5316\u30ce\u30a6\u30cf\u30a6\u304c\u5fc5\u8981<\/td>\n <td>\u63a8\u8ad6\u30b3\u30b9\u30c8\u6700\u5c0f\u5316\u306e\u672c\u547d\u3002\u30af\u30e9\u30a6\u30c9\/\u30aa\u30f3\u30d7\u30ec\u6df7\u5728\u3067\u30ed\u30c3\u30af\u30a4\u30f3\u56de\u907f<\/td>\n <\/tr>\n <tr>\n <td><strong>Intel\uff08OpenVINO \/ CPU+iGPU\uff09<\/strong><\/td>\n <td>OCR\u30fb\u691c\u51fa\u30fb\u30af\u30e9\u30b7\u30ab\u30ebCV\u30fb\u8efd\u91cfLLM\/Embedding<\/td>\n <td>CPU\u3060\u3051\u3067INT8\u304c\u901f\u3044\u3002\u4f9d\u5b58\u30e9\u30a4\u30d6\u30e9\u30ea\u5c11\u306a\u304f\u904b\u7528\u304c\u8efd\u3044<\/td>\n <td>\u8d85\u5927\u898f\u6a21\u751f\u6210\u306f\u4e0d\u5411\u304d<\/td>\n <td>\u7701\u96fb\u529b\u30fb\u73fe\u5834Edge\u7528\u306b\u6700\u9069\u3002\u65e2\u8a2d\u30b5\u30fc\u30d0\u306e\u5ef6\u547d\u306b\u52b9\u304f<\/td>\n <\/tr>\n <tr>\n <td><strong>\u5c02\u7528HW\uff08Inferentia2\/TPU\u7b49\uff09<\/strong><\/td>\n <td>LLM\/Embedding \u5927\u91cf\u30d0\u30c3\u30c1 \/ \u691c\u7d22<\/td>\n <td>\u30af\u30e9\u30a6\u30c9\u3067\u306e\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8\u5358\u4fa1\u304c\u5b89\u3044<\/td>\n <td>\u5b66\u7fd2\/\u30e2\u30c7\u30eb\u4e92\u63db\u3084\u904b\u7528\u77e5\u898b\u304c\u5fc5\u8981<\/td>\n <td>\u30b9\u30d1\u30a4\u30af\u8ca0\u8377\u3084\u591c\u9593\u30d0\u30c3\u30c1\u306b\u9650\u5b9a\u3057\u3066\u6df7\u5728\u904b\u7528<\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n <p class=\"note\">\u610f\u601d\u6c7a\u5b9a\u306e\u8981\u70b9\uff1a<strong>\u63a8\u8ad6\u306f\u591a\u69d8\u5316\u30fb\u5206\u6563\uff08\u30de\u30eb\u30c1\u30d9\u30f3\u30c0\uff09<\/strong>\u304c\u57fa\u672c\u3002\u5b66\u7fd2\u306fNVIDIA\u5bc4\u308a\u3067\u3082\u3001\u63a8\u8ad6\u306f AMD\/Intel\/\u5c02\u7528HW \u3092\u6df7\u305c\u3066 <em>\u5186\/\u30ea\u30af\u30a8\u30b9\u30c8<\/em> \u3092\u6700\u5c0f\u5316\u3002<\/p>\n<\/section>\n\n<section>\n <h2>2. \u5b9f\u88c5\u30ec\u30b7\u30d4\uff08\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u5225\uff09<\/h2>\n <h3>2.1 ONNX Runtime\uff08\u5168\u65b9\u4f4d\u30fb\u307e\u305a\u306f\u3053\u308c\uff09<\/h3>\n <p>EP\uff08Execution Provider\uff09\u3092\u5dee\u3057\u66ff\u3048\u308b\u3060\u3051\u3067 CUDA \/ ROCm \/ OpenVINO \/ TensorRT \u3092\u5207\u308a\u66ff\u3048\u53ef\u80fd\u3002\u5358\u4e00\u30b3\u30fc\u30c9\u3067\u30de\u30eb\u30c1\u30d9\u30f3\u30c0\u5bfe\u5fdc\u3002<\/p>\n <pre><code class=\"language-bash\"># \u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u4f8b\uff08\u74b0\u5883\u306b\u5408\u308f\u305b\u3066\u9078\u629e\uff09\n# CUDA:\npip install onnxruntime-gpu # \u30d0\u30fc\u30b8\u30e7\u30f3\u306fCUDA\u5bfe\u5fdc\u8868\u3092\u5fc5\u305a\u78ba\u8a8d\n# ROCm:\npip install onnxruntime-rocm\n# OpenVINO(=CPU\/iGPU\u6700\u9069\u5316):\npip install onnxruntime-openvino\n<\/code><\/pre>\n <pre><code class=\"language-python\">import onnxruntime as ort\n\n# EP\u512a\u5148\u9806\u4f4d\u3092\u5217\u6319\uff08\u6700\u521d\u306b\u4f7f\u3048\u308b\u3082\u306e\u304c\u9078\u3070\u308c\u308b\uff09\nproviders = [\n (\"TensorrtExecutionProvider\", {\"trt_fp16_enable\": True, \"trt_int8_enable\": True}),\n \"CUDAExecutionProvider\",\n \"ROCmExecutionProvider\",\n (\"OpenVINOExecutionProvider\", {\"device_type\": \"CPU_FP32\"}),\n \"CPUExecutionProvider\"\n]\nso = ort.SessionOptions(); so.intra_op_num_threads = 0\nsess = ort.InferenceSession(\"model.onnx\", providers=providers, sess_options=so)\n<\/code><\/pre>\n\n <h4>\u91cf\u5b50\u5316\uff08PTQ\/QAT\uff09<\/h4>\n <pre><code class=\"language-bash\"># \u52d5\u7684\u91cf\u5b50\u5316\uff08Linear INT8\uff1a\u4e3b\u306bTransformer\u306eMatMul\uff09\npython -m onnxruntime.quantization.quantize_dynamic \\\n --model_input model.onnx --model_output model.int8.onnx \\\n --per_channel --optimize_model --activation_type qint8 --weight_type qint8\n\n# \u9759\u7684\u91cf\u5b50\u5316\uff08\u6821\u6b63\u30c7\u30fc\u30bf\u3067\u6821\u6b63\uff09\uff1a\u753b\u50cf\/\u691c\u51fa\u30e2\u30c7\u30eb\u306b\u6709\u52b9\npython -m onnxruntime.quantization.quantize \\\n --model_input model.onnx --model_output model.int8.onnx \\\n --quant_format QDQ --calibrate \\\n --data_folder .\/calib --per_channel --activation_type qint8 --weight_type qint8\n<\/code><\/pre>\n <p class=\"note\">Tip\uff1a<strong>INT8\u5316\u3067\u7cbe\u5ea6\u304c\u843d\u3061\u305f\u3089<\/strong>\u3001(1) activation \u3060\u3051 INT8\u3001weight \u306f FP16\u3001(2) \u4e8b\u524d\u306e <em>SmoothQuant<\/em> \u4e92\u63db\u306e\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u3001(3) \u6821\u6b63\u30c7\u30fc\u30bf\u3092\u201c\u672c\u756a\u5206\u5e03\u201d\u306b\u5bc4\u305b\u308b\u3002<\/p>\n\n <h3>2.2 TensorRT\uff08NVIDIA\u6700\u901f\u306e\u5b9a\u756a\uff09<\/h3>\n <pre><code class=\"language-bash\"># \u30a8\u30f3\u30b8\u30f3\u751f\u6210\uff08FP16\/INT8\uff09\ntrtexec --onnx=model.onnx --saveEngine=model_fp16.trt \\\n --fp16 --workspace=4096 --timingCacheFile=cache.txt\n\n# INT8\uff08\u8981\u6821\u6b63\uff09\uff1a\ntrtexec --onnx=model.onnx --saveEngine=model_int8.trt \\\n --int8 --calib=calib.cache --workspace=4096 --timingCacheFile=cache.txt\n<\/code><\/pre>\n <p>\u914d\u4fe1\u306f <strong>Triton Inference Server<\/strong> \u3067\u3002\u30e2\u30c7\u30eb\u30ea\u30dd\u30b8\u30c8\u30ea\u306b\u7f6e\u304f\u3060\u3051\u3067\u540c\u6642\u306b ONNX\/TensorRT\/PyTorch \u3092\u6271\u3048\u308b\u3002A\/B\u30c6\u30b9\u30c8\u3084\u52d5\u7684\u30d0\u30c3\u30c1\u3067 <em>\u5186\/\u30ea\u30af\u30a8\u30b9\u30c8<\/em> \u3092\u4e0b\u3052\u308b\u3002<\/p>\n\n <h3>2.3 ROCm \/ AMD\uff08PyTorch + ONNX\/Triton\uff09<\/h3>\n <pre><code class=\"language-bash\"># PyTorch(ROCm) \u4f8b\uff1a\u30a4\u30f3\u30c7\u30c3\u30af\u30b9URL\u306f\u74b0\u5883\u306b\u5408\u308f\u305b\u308b\npip install torch torchvision torchaudio --index-url https:\/\/download.pytorch.org\/whl\/rocm6.0\n\n# \u5b9f\u884c\u6642\u306b\u30bf\u30fc\u30b2\u30c3\u30c8GPU\u3092\u5236\u5fa1\nexport HIP_VISIBLE_DEVICES=0\n<\/code><\/pre>\n <pre><code class=\"language-python\">import torch\n# \u65e2\u5b58PyTorch\u30e2\u30c7\u30eb\u3092 ROCm \u3067 FP16 \u5b9f\u884c\nmodel.half().to(\"cuda\") # AMD\u3067\u3082 torch.cuda \u30c7\u30d0\u30a4\u30b9\u6271\u3044\nwith torch.inference_mode():\n y = model(x.half().to(\"cuda\"))\n<\/code><\/pre>\n <p>ONNX\u306b\u843d\u3068\u3057\u3066 onnxruntime-rocm \/ Triton\uff08TensorRT \u3067\u306f\u306a\u304f ONNX Backend\uff09\u3067\u904b\u7528\u3059\u308b\u3068\u7ba1\u7406\u304c\u697d\u3002<br>JIT\u3084 <code>torch.compile()<\/code>\uff08inductor\uff09\u3067 kernel fusion \u3092\u72d9\u3046\u3002<\/p>\n\n <h3>2.4 OpenVINO\uff08Intel CPU\/iGPU\uff09<\/h3>\n <pre><code class=\"language-bash\"># FP16 \u5909\u63db\uff08OVC\uff09\npython -m openvino.tools.ovc --input_model model.onnx --compress_to_fp16 \\\n --output_dir out_ov_fp16\n\n# PTQ\uff08POT\uff09\u8a2d\u5b9a\u4f8b\uff08YAML\/JSON\uff09\ncat <<'JSON' > pot.json\n{\n \"model\": {\"model_name\": \"detector\", \"model\": \"out_ov_fp16\/model.xml\"},\n \"engine\": {\"type\": \"simplified\", \"data_source\": \".\/calib\"},\n \"compression\": [{\"algorithm\": \"DefaultQuantization\", \"preset\": \"performance\"}]\n}\nJSON\npot -c pot.json -d out_ov_int8\n<\/code><\/pre>\n <pre><code class=\"language-python\">from openvino.runtime import Core\nie = Core()\ncomp = ie.compile_model(\"out_ov_int8\/detector.xml\", device_name=\"CPU\")\n<\/code><\/pre>\n <p class=\"note\">OpenVINO \u306f CPU \u3060\u3051\u3067\u3082 INT8 \u3067\u5065\u95d8\u3002OCR\u30fb\u5c0f\u578b\u691c\u51fa\u30fb\u30ec\u30a4\u30a2\u30a6\u30c8\u89e3\u6790\u306e <em>\u5186\/\u51e6\u7406<\/em> \u3092\u5287\u7684\u306b\u4e0b\u3052\u3084\u3059\u3044\u3002<\/p>\n<\/section>\n\n<section>\n <h2>3. \u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u5225\u306e\u6700\u9069\u5316\u30d1\u30bf\u30fc\u30f3<\/h2>\n <h3>3.1 \u753b\u50cf\u89e3\u6790\uff08\u691c\u51fa\/\u5206\u985e\/\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\uff09<\/h3>\n <ul>\n <li><strong>\u30e2\u30c7\u30eb\u9078\u629e<\/strong>\uff1a\u84b8\u7559\u30fb\u5c0f\u578b\uff08n\/s\/tiny\u7cfb\uff09\u3002NMS\/\u7f70\u5247\u306e\u95be\u5024\u306f\u5b9f\u30c7\u30fc\u30bf\u3067\u518d\u6700\u9069\u5316\u3002<\/li>\n <li><strong>\u524d\u51e6\u7406<\/strong>\uff1a\u5165\u529b\u89e3\u50cf\u5ea6\u3092 <em>\u52d5\u7684<\/em> \u306b\u4e0b\u3052\u308b\uff08\u30b9\u30b3\u30a2\u5b89\u5b9a\u7bc4\u56f2\u3067\uff09\u3002<\/li>\n <li><strong>\u30d0\u30c3\u30c1\u30f3\u30b0<\/strong>\uff1aTriton\u306e\u52d5\u7684\u30d0\u30c3\u30c1\u3067 p95 \u3092\u5d29\u3055\u305a\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8\u6700\u5927\u5316\u3002<\/li>\n <li><strong>\u91cf\u5b50\u5316<\/strong>\uff1aINT8\uff08QDQ\/PTQ\uff09\u2192\u843d\u3061\u308b\u30af\u30e9\u30b9\u3060\u3051\u84b8\u7559\u4ed8\u304d QAT\u3002<\/li>\n <li><strong>\u5f8c\u51e6\u7406\u9ad8\u901f\u5316<\/strong>\uff1aNMS \u3092 GPU\/EP \u5b9f\u88c5\u306b\u5bc4\u305b\u308b\uff08ONNX opset or TensorRT plugin\uff09\u3002<\/li>\n <\/ul>\n <pre><code class=\"language-bash\"># YOLO\u7cfb\u306eONNX\u6700\u9069\u5316\uff08\u4e0d\u8981\u30ce\u30fc\u30c9\u524a\u6e1b, NMS\u7d71\u5408\uff09\npython tools\/export_yolo_onnx.py --weights best.pt --opt --nms --img 960 960\n<\/code><\/pre>\n\n <h3>3.2 \u751f\u6210\uff08\u62e1\u6563\uff1aStable Diffusion \u7cfb\uff09<\/h3>\n <ul>\n <li><strong>\u89e3\u50cf\u5ea6\u53ef\u5909<\/strong>\uff1a\u65e2\u5b9a 1024 \u3092 768\/896 \u306b\u81ea\u52d5\u964d\u683c\uff08\u6587\u8108\u6b21\u7b2c\uff09<\/li>\n <li><strong>\u91cd\u307f\u84b8\u7559<\/strong>\uff1aLoRA\/\u884c\u5217\u5206\u89e3\u3067\u8efd\u91cf\u5316\u3001<em>VAE \u306e INT8\/FP16\u5316<\/em> \u306f\u4f53\u611f\u5dee\u5927<\/li>\n <li><strong>xFormers \/ FlashAttention \u76f8\u5f53<\/strong>\uff1a\u30e1\u30e2\u30ea\u5e2f\u57df\u524a\u6e1b<\/li>\n <li><strong>\u30d0\u30c3\u30c1\u751f\u6210<\/strong>\uff1a\u30ac\u30a4\u30c0\u30f3\u30b9\u6bd4\u3068\u30b9\u30c6\u30c3\u30d7\u6570\u306e\u30af\u30ea\u30c3\u30d4\u30f3\u30b0\uff08\u54c1\u8cea-\u901f\u5ea6 Pareto\uff09<\/li>\n <\/ul>\n <pre><code class=\"language-python\">pipe.enable_xformers_memory_efficient_attention()\npipe.vae.to(dtype=torch.float16)\npipe.unet.to(memory_format=torch.channels_last)\npipe = torch.compile(pipe) # ROCm\/CUDA \u3044\u305a\u308c\u3082\u6709\u52b9\u306a\u3053\u3068\u304c\u591a\u3044\n<\/code><\/pre>\n\n <h3>3.3 OCR\uff08\u30ec\u30a4\u30a2\u30a6\u30c8 + \u8a8d\u8b58\uff09<\/h3>\n <ul>\n <li><strong>\u30ec\u30a4\u30a2\u30a6\u30c8\u5206\u5272<\/strong>\uff1a\u6bb5\u843d\/\u8868\/\u56f3\u3092\u5148\u306b\u5207\u308a\u5206\u3051\u3066\u500b\u5225\u6700\u9069\uff08\u5c0f\u578b\u691c\u51fa+CRNN\/ViT\uff09<\/li>\n <li><strong>\u89e3\u50cf\u5ea6\u81ea\u52d5<\/strong>\uff1a\u7d30\u5b57\/\u5c0f\u6587\u5b57\u306f DPI \u3092\u4e0a\u3052\u3066\u5c40\u6240\u518d\u30b9\u30ad\u30e3\u30f3<\/li>\n <li><strong>OpenVINO INT8<\/strong>\uff1aCPU\u306e\u307f\u3067\u53ef\u3002Edge\/\u30b5\u30fc\u30d0\u6df7\u5728\u306b\u5f37\u3044<\/li>\n <li><strong>\u8a00\u8a9e\u30e2\u30fc\u30c9\u5206\u96e2<\/strong>\uff1a\u82f1\u6570\u5b57\/\u65e5\u672c\u8a9e\/\u4e2d\u97d3\u3067\u30e2\u30c7\u30eb\u5207\u66ff\uff08\u96c6\u5408\u77e5\u3088\u308a\u901f\u3044\uff09<\/li>\n <\/ul>\n <pre><code class=\"language-python\"># onnxruntime + OpenVINO EP \u3067 CPU INT8 \u3092\u4f7f\u3046\u4f8b\nsess = ort.InferenceSession(\"ocr_int8.onnx\", providers=[\"OpenVINOExecutionProvider\"])\n<\/code><\/pre>\n\n <h3>3.4 \u691c\u7d22\uff08\u57cb\u3081\u8fbc\u307f + ANN + \u518d\u30e9\u30f3\u30ad\u30f3\u30b0\uff09<\/h3>\n <ul>\n <li><strong>\u8efd\u91cf\u57cb\u3081\u8fbc\u307f<\/strong>\uff1a\u6b21\u5143\u6570 768\u2192384 \u3078\u84b8\u7559\u3002INT8\/FP16 \u3067\u5145\u5206\u3002<\/li>\n <li><strong>ANN<\/strong>\uff1aHNSW or IVF+PQ\uff08OPQ\uff09\u3067\u30e1\u30e2\u30ea\uff06\u30c7\u30a3\u30b9\u30af\u524a\u6e1b\u3002<\/li>\n <li><strong>\u518d\u30e9\u30f3\u30af<\/strong>\uff1aCross-Encoder \u306f <em>\u30c8\u30c3\u30d7K\u306e\u307f<\/em> \u306b\u9650\u5b9a\uff081\u301c5\u4ef6\uff09\u3002<\/li>\n <li><strong>\u30ad\u30e3\u30c3\u30b7\u30e5<\/strong>\uff1a\u983b\u51fa\u30af\u30a8\u30ea\/\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u57cb\u3081\u8fbc\u307f\u3092\u30e1\u30e2\u30ea\/SSD\u30ad\u30e3\u30c3\u30b7\u30e5\u3002<\/li>\n <\/ul>\n <pre><code class=\"language-python\"># FAISS: IVF+PQ \u306e\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u4f8b\nimport faiss\nimport numpy as np\n\nD = 384; nlist = 4096; m = 64; # 64x8bit = 512bit\/\u30d9\u30af\u30c8\u30eb\nquantizer = faiss.IndexFlatL2(D)\nindex = faiss.IndexIVFPQ(quantizer, D, nlist, m, 8)\nindex.train(train_vecs.astype('float32'))\nindex.add(base_vecs.astype('float32'))\nD,I = index.search(query_vecs.astype('float32'), 20)\n<\/code><\/pre>\n<\/section>\n\n<section>\n <h2>4. \u91cf\u5b50\u5316\u30fb\u84b8\u7559\u306e\u5b9f\u6226\u30ec\u30b7\u30d4<\/h2>\n <h3>4.1 PTQ\uff08\u5f8c\u4ed8\u3051\u91cf\u5b50\u5316\uff09<\/h3>\n <ul>\n <li>\u6821\u6b63\u30bb\u30c3\u30c8\u3092 <em>\u672c\u756a\u30c7\u30fc\u30bf\u5206\u5e03<\/em> \u304b\u3089\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\uff08\u8aa4\u5dee\u306f\u5206\u5e03\u30ba\u30ec\u304c\u4e3b\u56e0\uff09<\/li>\n <li>\u7cbe\u5ea6\u60aa\u5316\u304c\u5927\u304d\u3044\u5c64\u306f <em>\u6df7\u5408\u7cbe\u5ea6<\/em>\uff08INT8\u21d4FP16\uff09\u3067\u6551\u6e08<\/li>\n <\/ul>\n <h3>4.2 QAT\uff08\u91cf\u5b50\u5316\u5bfe\u5fdc\u5b66\u7fd2\uff09<\/h3>\n <pre><code class=\"language-python\"># PyTorch\uff1aFakeQuant \u3092\u7d44\u307f\u8fbc\u307f\u518d\u5b66\u7fd2\nfrom torch.ao.quantization import get_default_qat_qconfig, prepare_qat, convert\nmodel.qconfig = get_default_qat_qconfig(\"fbgemm\")\nprepare_qat(model, inplace=True)\n# \u3053\u3053\u3067\u5fae\u8abf\u6574\u5b66\u7fd2 ...\nmodel = convert(model.eval())\n<\/code><\/pre>\n <h3>4.3 \u84b8\u7559\uff08\u6559\u5e2b=\u5143\u30e2\u30c7\u30eb, \u751f\u5f92=\u8efd\u91cf\uff09<\/h3>\n <pre><code class=\"language-python\"># \u5178\u578b\u30ed\u30b9\uff1a\u03b1*SoftTarget(KL) + \u03b2*HardCE + \u03b3*FeatureMatch\nloss = alpha*kl(soft_student, soft_teacher) + beta*ce(y_s, y) + gamma*mse(f_s, f_t)\n<\/code><\/pre>\n <p class=\"note\">\u9375\u306f <strong>\u8a55\u4fa1\u306e\u81ea\u52d5\u5316<\/strong>\uff08\u7cbe\u5ea6\u2194\u901f\u5ea6\u306e Pareto \u66f2\u7dda\u3092\u65e5\u6b21\u3067\u66f4\u65b0\uff09\u3002\u7cbe\u5ea6\u304c\u95be\u5024\u3092\u5272\u3089\u306a\u3044\u6700\u3082\u5b89\u3044\u69cb\u6210\u3092\u5e38\u306b\u9078\u3076\u3002<\/p>\n<\/section>\n\n<section>\n <h2>5. \u914d\u4fe1\u30fb\u30aa\u30fc\u30b1\u30b9\u30c8\u30ec\u30fc\u30b7\u30e7\u30f3<\/h2>\n <h3>5.1 Triton Inference Server<\/h3>\n <ul>\n <li>ONNX\/TensorRT\/PyTorch\/Python backend \u3092\u5358\u4e00\u30a8\u30f3\u30c9\u30dd\u30a4\u30f3\u30c8\u3067\u675f\u306d\u308b<\/li>\n <li>\u52d5\u7684\u30d0\u30c3\u30c1 \/ \u540c\u6642\u5b9f\u884c \/ \u30e2\u30c7\u30eb\u9593 Ensemble \/ A\/B \u30c6\u30b9\u30c8<\/li>\n <\/ul>\n <pre><code class=\"language-ini\"># model_repository\/layout\uff08\u4f8b\uff09\nrepo\/\n detector\/\n 1\/model.onnx\n config.pbtxt # max_batch_size, dynamic_batching, instance_group\n ocr\/\n 1\/model.xml # OpenVINO\n config.pbtxt\n<\/code><\/pre>\n\n <h3>5.2 vLLM \/ Text Generation Inference\uff08LLM\/Reranker\uff09<\/h3>\n <ul>\n <li>PagedAttention \u7b49\u3067\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8\u6539\u5584\u3001Speculative Decoding \u6709\u52b9<\/li>\n <li>Embedding \u306f onnxruntime \/ openvino \u3067\u5225\u30e9\u30a4\u30f3\u306b\u9003\u304c\u3059<\/li>\n <\/ul>\n\n <h3>5.3 \u76e3\u8996\u30fbSLO<\/h3>\n <ul>\n <li>p50\/p95 \u30ec\u30a4\u30c6\u30f3\u30b7\u3001GPU\/CPU \u5229\u7528\u7387\u3001\u30d0\u30c3\u30c1\u30b5\u30a4\u30ba\u5206\u5e03\u3001\u30a8\u30e9\u30fc\u7387<\/li>\n <li>\u54c1\u8cea\u76e3\u8996\uff1amAP\/CER\/nDCG \u3092\u30aa\u30f3\u30e9\u30a4\u30f3\u8a55\u4fa1\uff08\u5c0f\u898f\u6a21\u30b5\u30f3\u30d7\u30eb\u3067\uff09<\/li>\n <\/ul>\n<\/section>\n\n<section>\n <h2>6. \u30b3\u30b9\u30c8\u5f0f\u3068\u610f\u601d\u6c7a\u5b9a\u30d5\u30ed\u30fc<\/h2>\n <pre><code>\u30b3\u30b9\u30c8\/\u30ea\u30af\u30a8\u30b9\u30c8 \u2252 (\u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u6642\u7d66 + \u96fb\u529b) \/ (\u6709\u52b9\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8 req\/s) + \u30b9\u30c8\u30ec\u30fc\u30b8\u8ee2\u9001\n\u6709\u52b9\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8 = f(\u52d5\u7684\u30d0\u30c3\u30c1, \u89e3\u50cf\u5ea6, \u91cf\u5b50\u5316, \u524d\u5f8c\u51e6\u7406\u306e\u4e26\u5217\u5316)\n<\/code><\/pre>\n <ol>\n <li>\u307e\u305a FP16 \u30d9\u30fc\u30b9\u30e9\u30a4\u30f3\uff08CUDA\/ROCm\/OV \u3044\u305a\u308c\u304b\uff09\u3092\u4f5c\u308b<\/li>\n <li>INT8 PTQ \u2192 \u7cbe\u5ea6\u78ba\u8a8d \u2192 \u3060\u3081\u306a\u3089\u5c64\u5358\u4f4d\u3067\u6df7\u5408\u7cbe\u5ea6<\/li>\n <li>\u524d\u51e6\u7406\u30c0\u30a6\u30f3\u30b5\u30f3\u30d7\u30eb\/\u52d5\u7684\u30d0\u30c3\u30c1\/\u30ad\u30e3\u30c3\u30b7\u30e5\u5c0e\u5165<\/li>\n <li>\u5fc5\u8981\u306a\u3089 QAT\/\u84b8\u7559\u3067\u7cbe\u5ea6\u56de\u5fa9<\/li>\n <li>\u30af\u30e9\u30a6\u30c9\u5c02\u7528HW\u3067\u591c\u9593\u30d0\u30c3\u30c1\u3092\u9003\u304c\u3059\uff08\u30b9\u30dd\u30c3\u30c8\u6d3b\u7528\uff09<\/li>\n <\/ol>\n<\/section>\n\n<section>\n <h2>7. \u5177\u4f53\u7684\u30c6\u30f3\u30d7\u30ec\uff08\u305d\u306e\u307e\u307e\u6d41\u7528OK\uff09<\/h2>\n <h3>7.1 REST \u63a8\u8ad6\uff08onnxruntime + \u52d5\u7684EP\uff09<\/h3>\n <pre><code class=\"language-python\">from fastapi import FastAPI\nimport onnxruntime as ort\nimport numpy as np\n\nproviders=[\"TensorrtExecutionProvider\",\"CUDAExecutionProvider\",\n \"ROCmExecutionProvider\",\"OpenVINOExecutionProvider\",\n \"CPUExecutionProvider\"]\nsess = ort.InferenceSession(\"model.onnx\", providers=providers)\n\napp = FastAPI()\n@app.post(\"\/infer\")\ndef infer(x: list[float]):\n a = np.array(x, dtype=np.float32)[None, :]\n y = sess.run(None, {sess.get_inputs()[0].name: a})[0]\n return {\"y\": y.tolist()}\n<\/code><\/pre>\n\n <h3>7.2 Triton\uff08\u52d5\u7684\u30d0\u30c3\u30c1\uff09<\/h3>\n <pre><code class=\"language-proto\"># config.pbtxt\uff08\u629c\u7c8b\uff09\nname: \"detector\"\nplatform: \"onnxruntime_onnx\"\nmax_batch_size: 32\ndynamic_batching { preferred_batch_size: [4, 8, 16] max_queue_delay_microseconds: 2000 }\ninstance_group [{ kind: KIND_GPU, count: 1 }]\n<\/code><\/pre>\n\n <h3>7.3 OpenVINO INT8 OCR<\/h3>\n <pre><code class=\"language-bash\">python -m openvino.tools.ovc --input_model ocr.onnx --compress_to_fp16 -o ov_fp16\ncat pot.json | jq . # \u4e0a\u306e\u4f8b\u3092\u53c2\u7167\npot -c pot.json -d ov_int8\n<\/code><\/pre>\n<\/section>\n\n<section>\n <h2>8. \u54c1\u8cea\u3092\u843d\u3068\u3055\u305a\u30b3\u30b9\u30c8\u3092\u843d\u3068\u3059\u300c\u5c0f\u30ef\u30b6\u300d\u96c6<\/h2>\n <ul>\n <li><strong>\u89e3\u50cf\u5ea6\u30a2\u30c0\u30d7\u30c8<\/strong>\uff1a\u691c\u51fa\u7d50\u679c\u304c\u5b89\u5b9a\u3059\u308b\u6700\u5c0f\u89e3\u50cf\u5ea6\u3092\u30aa\u30f3\u30e9\u30a4\u30f3\u3067\u63a2\u7d22\u30fb\u4fdd\u5b58<\/li>\n <li><strong>\u65e9\u671f\u6253\u3061\u5207\u308a<\/strong>\uff1a\u751f\u6210\u306f <em>\u30b9\u30c6\u30c3\u30d7\u6570\u4e0a\u9650<\/em> \u3092\u52d5\u7684\u5236\u5fa1\u3001LLM \u306f speculative<\/li>\n <li><strong>\u30ad\u30e3\u30c3\u30b7\u30e5<\/strong>\uff1a\u518d\u5165\u529b\u306e\u591a\u3044 OCR \u30da\u30fc\u30b8\u3084\u691c\u7d22\u30af\u30a8\u30ea\u3092 TTL \u30ad\u30e3\u30c3\u30b7\u30e5<\/li>\n <li><strong>\u5206\u5c90<\/strong>\uff1a\u8efd\u91cf\u30e2\u30c7\u30eb\u2192\u5fae\u5999\u306a\u3089\u91cd\u3044\u30e2\u30c7\u30eb\u3067\u518d\u5224\u5b9a\uff08Two-Stage\uff09<\/li>\n <li><strong>\u30b9\u30b1\u30fc\u30eb<\/strong>\uff1a\u958b\u5e97\u30fb\u663c\u30fb\u591c\u3067\u30ec\u30d7\u30ea\u30ab\u6570\u3092\u81ea\u52d5\u5207\u66ff\uff08HPA\/\u30b9\u30b1\u30b8\u30e5\u30fc\u30e9\uff09<\/li>\n <\/ul>\n<\/section>\n\n<section>\n <h2>9. \u65e2\u5b58\u8cc7\u7523\u3078\u306e\u9069\u7528\u9806\u5e8f\uff08\u6700\u77ed2\u9031\u9593\u30d7\u30e9\u30f3\uff09<\/h2>\n <ol>\n <li>\u73fe\u72b6\u306e p95, req\/s, mAP\/CER \u3092\u56fa\u5b9a\u5316\u8a08\u6e2c\uff08\u30d9\u30fc\u30b9\u30e9\u30a4\u30f3\uff09<\/li>\n <li>ONNX \u3078\u6b63\u898f\u5316 \u2192 onnxruntime \u3067 EP \u53ef\u5909\u306b<\/li>\n <li>INT8 PTQ \u2192 \u7cbe\u5ea6\u691c\u8a3c \u2192\u30c0\u30e1\u7b87\u6240\u306e\u307f\u6df7\u5408\u7cbe\u5ea6 \/ QAT<\/li>\n <li>Triton \u5c0e\u5165\uff08\u52d5\u7684\u30d0\u30c3\u30c1\u30fbA\/B\uff09\u2192 \u30b3\u30b9\u30c8\u6700\u5c0f\u8a2d\u5b9a\u3092\u63a1\u629e<\/li>\n <li>\u91cd\u3044\u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u3092\u591c\u9593\u3060\u3051\u30af\u30e9\u30a6\u30c9\u5c02\u7528HW\u3078\u9003\u304c\u3059<\/li>\n <\/ol>\n <p class=\"warn\">\u3088\u304f\u3042\u308b\u5931\u6557\uff1a<strong>\u6821\u6b63\u30c7\u30fc\u30bf\u304c\u30b7\u30e7\u30dc\u3044<\/strong>\uff08\u2192INT8\u304c\u30dc\u30ed\u30dc\u30ed\uff09\u3002<strong>\u524d\u51e6\u7406\u30fb\u5f8c\u51e6\u7406\u304cCPU\u306b\u6b8b\u7559<\/strong>\uff08\u2192GPU\u904a\u3076\uff09\u3002<strong>p50\u3060\u3051\u6539\u5584<\/strong>\uff08\u2192p95\u60aa\u5316\u3067\u4f53\u611f\u30de\u30a4\u30ca\u30b9\uff09\u3002<\/p>\n<\/section>\n\n<section>\n <h2>10. \u4ed8\u9332\uff1a\u7528\u8a9e\u3056\u3063\u304f\u308a<\/h2>\n <ul>\n <li><strong>EP\uff08Execution Provider\uff09<\/strong>\uff1aONNX Runtime \u306b\u304a\u3051\u308b\u5b9f\u884c\u30a8\u30f3\u30b8\u30f3\u306e\u5207\u66ff\u30d7\u30e9\u30b0\u30a4\u30f3<\/li>\n <li><strong>PTQ\/QAT<\/strong>\uff1a\u5f8c\u4ed8\u91cf\u5b50\u5316 \/ \u91cf\u5b50\u5316\u5bfe\u5fdc\u5b66\u7fd2<\/li>\n <li><strong>IVF+PQ\/OPQ<\/strong>\uff1a\u30d9\u30af\u30c8\u30eb\u5727\u7e2e\u306e\u5b9a\u756a\u3002\u5de8\u5927\u30b3\u30fc\u30d1\u30b9\u306e\u691c\u7d22\u3092\u5b89\u4fa1\u306b<\/li>\n <li><strong>Speculative Decoding<\/strong>\uff1a\u5c0f\u30e2\u30c7\u30eb\u3067\u5148\u8aad\u307f\u3057\u5927\u30e2\u30c7\u30eb\u3067\u691c\u8a3c\u3059\u308b\u9ad8\u901f\u5316<\/li>\n <\/ul>\n<\/section>\n\n<\/main>\n\n<footer>\n <p>\u6700\u7d42\u66f4\u65b0: 2025-09-01 \/ \u4f5c\u6210\u8005: \u30d3\u30fc\u30ca\u30ec\u30c3\u30b8\u30c7\u30b6\u30a4\u30f3<\/p>\n<\/footer>\n<\/body>\n<\/html>\n\n","protected":false},"excerpt":{"rendered":"<p>\u63a8\u8ad6\u30b3\u30b9\u30c8\u6700\u9069\u5316\u30d7\u30ec\u30a4\u30d6\u30c3\u30af \u524d\u63d0\uff1a\u30aa\u30f3\u30d7\u30ec\uff08\u4f8b\uff1aRTX 4070 Ti SUPER \u00d71\uff09\u307e\u305f\u306f\u30af\u30e9\u30a6\u30c9\u306e\u6df7\u5728\u3092\u60f3\u5b9a\u3002CUDA \u3060\u3051\u306b\u7e1b\u3089\u308c\u305a\u3001ROCm \/ OpenVINO \/ ONNX Runtime \/ Tens […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"saved_in_kubio":false,"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"_uf_show_specific_survey":0,"_uf_disable_surveys":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"sns_share_botton_hide":"","vkExUnit_sns_title":"","_vk_print_noindex":"","sitemap_hide":"","vkExUnit_EyeCatch_disable":"","_veu_custom_css":"","veu_display_promotion_alert":"common","vkexunit_cta_each_option":"","_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_post_was_ever_published":false},"categories":[5,68,24],"tags":[],"class_list":["post-1633","post","type-post","status-publish","format-standard","hentry","category-ai","category-68","category-24"],"aioseo_notices":[],"veu_head_title_object":{"title":"","add_site_title":""},"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/posts\/1633","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1633"}],"version-history":[{"count":3,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/posts\/1633\/revisions"}],"predecessor-version":[{"id":1636,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/posts\/1633\/revisions\/1636"}],"wp:attachment":[{"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1633"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1633"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1633"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}

{"id":1633,"date":"2025-09-01T10:06:20","date_gmt":"2025-09-01T01:06:20","guid":{"rendered":"https:\/\/beeknowledge.co.jp\/?p=1633"},"modified":"2025-09-01T10:06:21","modified_gmt":"2025-09-01T01:06:21","slug":"%e6%8e%a8%e8%ab%96%e3%82%b3%e3%82%b9%e3%83%88%e6%9c%80%e9%81%a9%e5%8c%96%e3%83%97%e3%83%ac%e3%82%a4%e3%83%96%e3%83%83%e3%82%af%ef%bc%88%e7%94%bb%e5%83%8f%e8%a7%a3%e6%9e%90%e3%83%bb%e7%94%9f%e6%88%90","status":"publish","type":"post","link":"https:\/\/beeknowledge.co.jp\/?p=1633","title":{"rendered":"\u63a8\u8ad6\u30b3\u30b9\u30c8\u6700\u9069\u5316\u30d7\u30ec\u30a4\u30d6\u30c3\u30af\uff08\u753b\u50cf\u89e3\u6790\u30fb\u751f\u6210\u30fbOCR\u30fb\u691c\u7d22\uff09\u8abf\u9054\u6226\u7565 \u00d7 \u5b9f\u88c5\u30ec\u30b7\u30d4 \u00d7 \u91cf\u5b50\u5316\u30fb\u6700\u9069\u5316 \u8a2d\u5b9a\u4f8b"},"content":{"rendered":"\n\n\n\n\n\u63a8\u8ad6\u30b3\u30b9\u30c8\u6700\u9069\u5316\u30d7\u30ec\u30a4\u30d6\u30c3\u30af<\/title>\n<style>\n body{font-family: -apple-system,BlinkMacSystemFont,Segoe UI,Roboto,Helvetica,Arial,\"Hiragino Kaku Gothic ProN\",\"Hiragino Sans\",\"Noto Sans CJK JP\",\"Yu Gothic UI\",\"YuGothic\",Meiryo,sans-serif; line-height:1.65; margin:40px; color:#111;}\n code, pre{font-family: ui-monospace,SFMono-Regular,Menlo,Monaco,Consolas,\"Liberation Mono\",\"Courier New\",monospace; font-size: 0.95em;}\n pre{background:#f6f8fa; padding:12px 16px; border-radius:8px; overflow:auto; border:1px solid #eaecef;}\n h1,h2,h3{line-height:1.35}\n table{border-collapse:collapse; margin:16px 0; width:100%}\n th,td{border:1px solid #ddd; padding:8px; vertical-align:top}\n th{background:#fafafa; text-align:left}\n .ok{color:#0a7}\n .warn{color:#e57}\n .mono{font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, \"Courier New\", monospace}\n<\/style>\n<\/head>\n<body>\n<header>\n \n <p class=\"lead\">\u524d\u63d0\uff1a\u30aa\u30f3\u30d7\u30ec\uff08\u4f8b\uff1aRTX 4070 Ti SUPER \u00d71\uff09\u307e\u305f\u306f\u30af\u30e9\u30a6\u30c9\u306e\u6df7\u5728\u3092\u60f3\u5b9a\u3002CUDA \u3060\u3051\u306b\u7e1b\u3089\u308c\u305a\u3001ROCm \/ OpenVINO \/ ONNX Runtime \/ TensorRT \/ Triton \/ vLLM \u7b49\u3092\u9069\u6750\u9069\u6240\u3067\u4f7f\u3044\u5206\u3051\u308b\u3002\u3053\u3053\u306b\u3042\u308b\u306e\u306f\u8abf\u67fb\u4e2d\u5185\u5bb9\u306e\u30e1\u30e2\u3067\u3059\u3002<\/p>\n<\/header>\n\n<main>\n<section>\n <h2>0. \u307e\u305a\u6c7a\u3081\u308b\u3053\u3068\uff083\u5206\u30c1\u30a7\u30c3\u30af\u30ea\u30b9\u30c8\uff09<\/h2>\n <div class=\"grid cols-2\">\n <div class=\"ok\">\n <strong>\u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u306e\u578b<\/strong>\n <ul>\n <li>\u753b\u50cf\u89e3\u6790\uff1a\u5206\u985e \/ \u691c\u51fa\uff08YOLO\/RT-DETR\u7cfb\uff09\/ \u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3<\/li>\n <li>\u751f\u6210\uff1a\u62e1\u6563\uff08SD\u7cfb\uff09\/ \u30c6\u30ad\u30b9\u30c8\u2192\u753b\u50cf \/ \u753b\u50cf\u2192\u753b\u50cf<\/li>\n <li>OCR\uff1a\u30ec\u30a4\u30a2\u30a6\u30c8\u89e3\u6790 + \u8a8d\u8b58\uff08PaddleOCR \u7b49\uff09<\/li>\n <li>\u691c\u7d22\uff1a\u57cb\u3081\u8fbc\u307f\u751f\u6210 + ANN \u691c\u7d22 + \uff08\u4efb\u610f\uff09\u518d\u30e9\u30f3\u30ad\u30f3\u30b0<\/li>\n <\/ul>\n <\/div>\n <div class=\"ok\">\n <strong>SLO\/\u5236\u7d04<\/strong>\n <ul>\n <li>\u30ec\u30a4\u30c6\u30f3\u30b7\uff08p95 ms\uff09\u30fb\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8\uff08req\/s\uff09\u30fbQoS\uff08Drop\u8a31\u5bb9\uff09<\/li>\n <li>\u7cbe\u5ea6\u3057\u304d\u3044\u5024\uff08mAP \/ F1 \/ CER \/ nDCG\uff09<\/li>\n <li>\u30b3\u30b9\u30c8\u4e0a\u9650\uff08\u5186\/1000 \u30ea\u30af\u30a8\u30b9\u30c8 or \u5186\/\u6642\uff09<\/li>\n <\/ul>\n <\/div>\n <div class=\"ok\">\n <strong>\u914d\u5099\u5148<\/strong>\n <ul>\n <li>CUDA\uff08NVIDIA\uff09\/ ROCm\uff08AMD\uff09\/ OpenVINO\uff08Intel CPU+iGPU\uff09<\/li>\n <li>\u30af\u30e9\u30a6\u30c9\u5c02\u7528 HW\uff08AWS Inferentia2 \/ TPU v5e \u306a\u3069\uff09<\/li>\n <\/ul>\n <\/div>\n <div class=\"ok\">\n <strong>\u30e2\u30c7\u30eb\u9078\u5b9a<\/strong>\n <ul>\n <li>\u84b8\u7559\u6e08\u307f\u30fb\u8efd\u91cf\u7cfb\uff08\u4f8b\uff1aYOLOv8n\/s, MobileSAM, Distil-Embedding\uff09<\/li>\n <li>\u91cf\u5b50\u5316\u524d\u63d0\uff08FP16\/INT8\/INT4\uff09\u3067\u518d\u5b66\u7fd2\u8a31\u5bb9\u304b<\/li>\n <\/ul>\n <\/div>\n <\/div>\n<\/section>\n\n<section>\n <h2>1. \u8abf\u9054\u6226\u7565\uff08\u30cf\u30fc\u30c9\u5225\u30fb\u8cbb\u7528\u5bfe\u52b9\u679c\u306e\u52d8\u6240\uff09<\/h2>\n <table>\n <thead><tr><th>\u533a\u5206<\/th><th>\u9069\u6027\u30ef\u30fc\u30af\u30ed\u30fc\u30c9<\/th><th>\u5f37\u307f<\/th><th>\u843d\u3068\u3057\u7a74<\/th><th>\u3056\u3063\u304f\u308a\u6307\u91dd<\/th><\/tr><\/thead>\n <tbody>\n <tr>\n <td><strong>NVIDIA\uff08CUDA\/TensorRT\uff09<\/strong><\/td>\n <td>\u753b\u50cf\u89e3\u6790\/\u751f\u6210, \u5927\u898f\u6a21LLM\u63a8\u8ad6, \u6df7\u5728<\/td>\n <td>\u6210\u719f\u30c4\u30fc\u30eb\u7fa4\uff08TensorRT, Triton, cuDNN\uff09\u3068\u9ad8\u52b9\u7387 FP16\/INT8<\/td>\n <td>\u88c5\u7f6e\u5358\u4fa1\u30fb\u6d88\u8cbb\u96fb\u529b\u9ad8\u3081\u3001\u5bfe\u4e2d\u8f38\u51fa\u898f\u5236\u306e\u5f71\u97ff<\/td>\n <td>\u5b66\u7fd2\u3082\u8996\u91ce\u306a\u3089\u7b2c\u4e00\u5019\u88dc\u3002\u63a8\u8ad6\u5c02\u7528\u306a\u3089 L4\/L40S\/RTX 4000 SFF \u3082\u691c\u8a0e<\/td>\n <\/tr>\n <tr>\n <td><strong>AMD\uff08ROCm\uff09<\/strong><\/td>\n <td>\u63a8\u8ad6\u5168\u822c\uff08\u7279\u306b\u753b\u50cf\/\u691c\u7d22\/\u4e00\u90e8LLM\uff09<\/td>\n <td>\u4fa1\u683c\u6027\u80fd\u6bd4\/\u5165\u624b\u6027\u3001ONNX\/Triton\/torch.compile \u3082\u6d3b\u7528\u53ef<\/td>\n <td>\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u9069\u5408\u6027\u5dee\u30fb\u4e00\u90e8\u30e2\u30c7\u30eb\u306e\u6700\u9069\u5316\u30ce\u30a6\u30cf\u30a6\u304c\u5fc5\u8981<\/td>\n <td>\u63a8\u8ad6\u30b3\u30b9\u30c8\u6700\u5c0f\u5316\u306e\u672c\u547d\u3002\u30af\u30e9\u30a6\u30c9\/\u30aa\u30f3\u30d7\u30ec\u6df7\u5728\u3067\u30ed\u30c3\u30af\u30a4\u30f3\u56de\u907f<\/td>\n <\/tr>\n <tr>\n <td><strong>Intel\uff08OpenVINO \/ CPU+iGPU\uff09<\/strong><\/td>\n <td>OCR\u30fb\u691c\u51fa\u30fb\u30af\u30e9\u30b7\u30ab\u30ebCV\u30fb\u8efd\u91cfLLM\/Embedding<\/td>\n <td>CPU\u3060\u3051\u3067INT8\u304c\u901f\u3044\u3002\u4f9d\u5b58\u30e9\u30a4\u30d6\u30e9\u30ea\u5c11\u306a\u304f\u904b\u7528\u304c\u8efd\u3044<\/td>\n <td>\u8d85\u5927\u898f\u6a21\u751f\u6210\u306f\u4e0d\u5411\u304d<\/td>\n <td>\u7701\u96fb\u529b\u30fb\u73fe\u5834Edge\u7528\u306b\u6700\u9069\u3002\u65e2\u8a2d\u30b5\u30fc\u30d0\u306e\u5ef6\u547d\u306b\u52b9\u304f<\/td>\n <\/tr>\n <tr>\n <td><strong>\u5c02\u7528HW\uff08Inferentia2\/TPU\u7b49\uff09<\/strong><\/td>\n <td>LLM\/Embedding \u5927\u91cf\u30d0\u30c3\u30c1 \/ \u691c\u7d22<\/td>\n <td>\u30af\u30e9\u30a6\u30c9\u3067\u306e\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8\u5358\u4fa1\u304c\u5b89\u3044<\/td>\n <td>\u5b66\u7fd2\/\u30e2\u30c7\u30eb\u4e92\u63db\u3084\u904b\u7528\u77e5\u898b\u304c\u5fc5\u8981<\/td>\n <td>\u30b9\u30d1\u30a4\u30af\u8ca0\u8377\u3084\u591c\u9593\u30d0\u30c3\u30c1\u306b\u9650\u5b9a\u3057\u3066\u6df7\u5728\u904b\u7528<\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n <p class=\"note\">\u610f\u601d\u6c7a\u5b9a\u306e\u8981\u70b9\uff1a<strong>\u63a8\u8ad6\u306f\u591a\u69d8\u5316\u30fb\u5206\u6563\uff08\u30de\u30eb\u30c1\u30d9\u30f3\u30c0\uff09<\/strong>\u304c\u57fa\u672c\u3002\u5b66\u7fd2\u306fNVIDIA\u5bc4\u308a\u3067\u3082\u3001\u63a8\u8ad6\u306f AMD\/Intel\/\u5c02\u7528HW \u3092\u6df7\u305c\u3066 <em>\u5186\/\u30ea\u30af\u30a8\u30b9\u30c8<\/em> \u3092\u6700\u5c0f\u5316\u3002<\/p>\n<\/section>\n\n<section>\n <h2>2. \u5b9f\u88c5\u30ec\u30b7\u30d4\uff08\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u5225\uff09<\/h2>\n <h3>2.1 ONNX Runtime\uff08\u5168\u65b9\u4f4d\u30fb\u307e\u305a\u306f\u3053\u308c\uff09<\/h3>\n <p>EP\uff08Execution Provider\uff09\u3092\u5dee\u3057\u66ff\u3048\u308b\u3060\u3051\u3067 CUDA \/ ROCm \/ OpenVINO \/ TensorRT \u3092\u5207\u308a\u66ff\u3048\u53ef\u80fd\u3002\u5358\u4e00\u30b3\u30fc\u30c9\u3067\u30de\u30eb\u30c1\u30d9\u30f3\u30c0\u5bfe\u5fdc\u3002<\/p>\n <pre><code class=\"language-bash\"># \u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u4f8b\uff08\u74b0\u5883\u306b\u5408\u308f\u305b\u3066\u9078\u629e\uff09\n# CUDA:\npip install onnxruntime-gpu # \u30d0\u30fc\u30b8\u30e7\u30f3\u306fCUDA\u5bfe\u5fdc\u8868\u3092\u5fc5\u305a\u78ba\u8a8d\n# ROCm:\npip install onnxruntime-rocm\n# OpenVINO(=CPU\/iGPU\u6700\u9069\u5316):\npip install onnxruntime-openvino\n<\/code><\/pre>\n <pre><code class=\"language-python\">import onnxruntime as ort\n\n# EP\u512a\u5148\u9806\u4f4d\u3092\u5217\u6319\uff08\u6700\u521d\u306b\u4f7f\u3048\u308b\u3082\u306e\u304c\u9078\u3070\u308c\u308b\uff09\nproviders = [\n (\"TensorrtExecutionProvider\", {\"trt_fp16_enable\": True, \"trt_int8_enable\": True}),\n \"CUDAExecutionProvider\",\n \"ROCmExecutionProvider\",\n (\"OpenVINOExecutionProvider\", {\"device_type\": \"CPU_FP32\"}),\n \"CPUExecutionProvider\"\n]\nso = ort.SessionOptions(); so.intra_op_num_threads = 0\nsess = ort.InferenceSession(\"model.onnx\", providers=providers, sess_options=so)\n<\/code><\/pre>\n\n <h4>\u91cf\u5b50\u5316\uff08PTQ\/QAT\uff09<\/h4>\n <pre><code class=\"language-bash\"># \u52d5\u7684\u91cf\u5b50\u5316\uff08Linear INT8\uff1a\u4e3b\u306bTransformer\u306eMatMul\uff09\npython -m onnxruntime.quantization.quantize_dynamic \\\n --model_input model.onnx --model_output model.int8.onnx \\\n --per_channel --optimize_model --activation_type qint8 --weight_type qint8\n\n# \u9759\u7684\u91cf\u5b50\u5316\uff08\u6821\u6b63\u30c7\u30fc\u30bf\u3067\u6821\u6b63\uff09\uff1a\u753b\u50cf\/\u691c\u51fa\u30e2\u30c7\u30eb\u306b\u6709\u52b9\npython -m onnxruntime.quantization.quantize \\\n --model_input model.onnx --model_output model.int8.onnx \\\n --quant_format QDQ --calibrate \\\n --data_folder .\/calib --per_channel --activation_type qint8 --weight_type qint8\n<\/code><\/pre>\n <p class=\"note\">Tip\uff1a<strong>INT8\u5316\u3067\u7cbe\u5ea6\u304c\u843d\u3061\u305f\u3089<\/strong>\u3001(1) activation \u3060\u3051 INT8\u3001weight \u306f FP16\u3001(2) \u4e8b\u524d\u306e <em>SmoothQuant<\/em> \u4e92\u63db\u306e\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u3001(3) \u6821\u6b63\u30c7\u30fc\u30bf\u3092\u201c\u672c\u756a\u5206\u5e03\u201d\u306b\u5bc4\u305b\u308b\u3002<\/p>\n\n <h3>2.2 TensorRT\uff08NVIDIA\u6700\u901f\u306e\u5b9a\u756a\uff09<\/h3>\n <pre><code class=\"language-bash\"># \u30a8\u30f3\u30b8\u30f3\u751f\u6210\uff08FP16\/INT8\uff09\ntrtexec --onnx=model.onnx --saveEngine=model_fp16.trt \\\n --fp16 --workspace=4096 --timingCacheFile=cache.txt\n\n# INT8\uff08\u8981\u6821\u6b63\uff09\uff1a\ntrtexec --onnx=model.onnx --saveEngine=model_int8.trt \\\n --int8 --calib=calib.cache --workspace=4096 --timingCacheFile=cache.txt\n<\/code><\/pre>\n <p>\u914d\u4fe1\u306f <strong>Triton Inference Server<\/strong> \u3067\u3002\u30e2\u30c7\u30eb\u30ea\u30dd\u30b8\u30c8\u30ea\u306b\u7f6e\u304f\u3060\u3051\u3067\u540c\u6642\u306b ONNX\/TensorRT\/PyTorch \u3092\u6271\u3048\u308b\u3002A\/B\u30c6\u30b9\u30c8\u3084\u52d5\u7684\u30d0\u30c3\u30c1\u3067 <em>\u5186\/\u30ea\u30af\u30a8\u30b9\u30c8<\/em> \u3092\u4e0b\u3052\u308b\u3002<\/p>\n\n <h3>2.3 ROCm \/ AMD\uff08PyTorch + ONNX\/Triton\uff09<\/h3>\n <pre><code class=\"language-bash\"># PyTorch(ROCm) \u4f8b\uff1a\u30a4\u30f3\u30c7\u30c3\u30af\u30b9URL\u306f\u74b0\u5883\u306b\u5408\u308f\u305b\u308b\npip install torch torchvision torchaudio --index-url https:\/\/download.pytorch.org\/whl\/rocm6.0\n\n# \u5b9f\u884c\u6642\u306b\u30bf\u30fc\u30b2\u30c3\u30c8GPU\u3092\u5236\u5fa1\nexport HIP_VISIBLE_DEVICES=0\n<\/code><\/pre>\n <pre><code class=\"language-python\">import torch\n# \u65e2\u5b58PyTorch\u30e2\u30c7\u30eb\u3092 ROCm \u3067 FP16 \u5b9f\u884c\nmodel.half().to(\"cuda\") # AMD\u3067\u3082 torch.cuda \u30c7\u30d0\u30a4\u30b9\u6271\u3044\nwith torch.inference_mode():\n y = model(x.half().to(\"cuda\"))\n<\/code><\/pre>\n <p>ONNX\u306b\u843d\u3068\u3057\u3066 onnxruntime-rocm \/ Triton\uff08TensorRT \u3067\u306f\u306a\u304f ONNX Backend\uff09\u3067\u904b\u7528\u3059\u308b\u3068\u7ba1\u7406\u304c\u697d\u3002<br>JIT\u3084 <code>torch.compile()<\/code>\uff08inductor\uff09\u3067 kernel fusion \u3092\u72d9\u3046\u3002<\/p>\n\n <h3>2.4 OpenVINO\uff08Intel CPU\/iGPU\uff09<\/h3>\n <pre><code class=\"language-bash\"># FP16 \u5909\u63db\uff08OVC\uff09\npython -m openvino.tools.ovc --input_model model.onnx --compress_to_fp16 \\\n --output_dir out_ov_fp16\n\n# PTQ\uff08POT\uff09\u8a2d\u5b9a\u4f8b\uff08YAML\/JSON\uff09\ncat <<'JSON' > pot.json\n{\n \"model\": {\"model_name\": \"detector\", \"model\": \"out_ov_fp16\/model.xml\"},\n \"engine\": {\"type\": \"simplified\", \"data_source\": \".\/calib\"},\n \"compression\": [{\"algorithm\": \"DefaultQuantization\", \"preset\": \"performance\"}]\n}\nJSON\npot -c pot.json -d out_ov_int8\n<\/code><\/pre>\n <pre><code class=\"language-python\">from openvino.runtime import Core\nie = Core()\ncomp = ie.compile_model(\"out_ov_int8\/detector.xml\", device_name=\"CPU\")\n<\/code><\/pre>\n <p class=\"note\">OpenVINO \u306f CPU \u3060\u3051\u3067\u3082 INT8 \u3067\u5065\u95d8\u3002OCR\u30fb\u5c0f\u578b\u691c\u51fa\u30fb\u30ec\u30a4\u30a2\u30a6\u30c8\u89e3\u6790\u306e <em>\u5186\/\u51e6\u7406<\/em> \u3092\u5287\u7684\u306b\u4e0b\u3052\u3084\u3059\u3044\u3002<\/p>\n<\/section>\n\n<section>\n <h2>3. \u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u5225\u306e\u6700\u9069\u5316\u30d1\u30bf\u30fc\u30f3<\/h2>\n <h3>3.1 \u753b\u50cf\u89e3\u6790\uff08\u691c\u51fa\/\u5206\u985e\/\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\uff09<\/h3>\n <ul>\n <li><strong>\u30e2\u30c7\u30eb\u9078\u629e<\/strong>\uff1a\u84b8\u7559\u30fb\u5c0f\u578b\uff08n\/s\/tiny\u7cfb\uff09\u3002NMS\/\u7f70\u5247\u306e\u95be\u5024\u306f\u5b9f\u30c7\u30fc\u30bf\u3067\u518d\u6700\u9069\u5316\u3002<\/li>\n <li><strong>\u524d\u51e6\u7406<\/strong>\uff1a\u5165\u529b\u89e3\u50cf\u5ea6\u3092 <em>\u52d5\u7684<\/em> \u306b\u4e0b\u3052\u308b\uff08\u30b9\u30b3\u30a2\u5b89\u5b9a\u7bc4\u56f2\u3067\uff09\u3002<\/li>\n <li><strong>\u30d0\u30c3\u30c1\u30f3\u30b0<\/strong>\uff1aTriton\u306e\u52d5\u7684\u30d0\u30c3\u30c1\u3067 p95 \u3092\u5d29\u3055\u305a\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8\u6700\u5927\u5316\u3002<\/li>\n <li><strong>\u91cf\u5b50\u5316<\/strong>\uff1aINT8\uff08QDQ\/PTQ\uff09\u2192\u843d\u3061\u308b\u30af\u30e9\u30b9\u3060\u3051\u84b8\u7559\u4ed8\u304d QAT\u3002<\/li>\n <li><strong>\u5f8c\u51e6\u7406\u9ad8\u901f\u5316<\/strong>\uff1aNMS \u3092 GPU\/EP \u5b9f\u88c5\u306b\u5bc4\u305b\u308b\uff08ONNX opset or TensorRT plugin\uff09\u3002<\/li>\n <\/ul>\n <pre><code class=\"language-bash\"># YOLO\u7cfb\u306eONNX\u6700\u9069\u5316\uff08\u4e0d\u8981\u30ce\u30fc\u30c9\u524a\u6e1b, NMS\u7d71\u5408\uff09\npython tools\/export_yolo_onnx.py --weights best.pt --opt --nms --img 960 960\n<\/code><\/pre>\n\n <h3>3.2 \u751f\u6210\uff08\u62e1\u6563\uff1aStable Diffusion \u7cfb\uff09<\/h3>\n <ul>\n <li><strong>\u89e3\u50cf\u5ea6\u53ef\u5909<\/strong>\uff1a\u65e2\u5b9a 1024 \u3092 768\/896 \u306b\u81ea\u52d5\u964d\u683c\uff08\u6587\u8108\u6b21\u7b2c\uff09<\/li>\n <li><strong>\u91cd\u307f\u84b8\u7559<\/strong>\uff1aLoRA\/\u884c\u5217\u5206\u89e3\u3067\u8efd\u91cf\u5316\u3001<em>VAE \u306e INT8\/FP16\u5316<\/em> \u306f\u4f53\u611f\u5dee\u5927<\/li>\n <li><strong>xFormers \/ FlashAttention \u76f8\u5f53<\/strong>\uff1a\u30e1\u30e2\u30ea\u5e2f\u57df\u524a\u6e1b<\/li>\n <li><strong>\u30d0\u30c3\u30c1\u751f\u6210<\/strong>\uff1a\u30ac\u30a4\u30c0\u30f3\u30b9\u6bd4\u3068\u30b9\u30c6\u30c3\u30d7\u6570\u306e\u30af\u30ea\u30c3\u30d4\u30f3\u30b0\uff08\u54c1\u8cea-\u901f\u5ea6 Pareto\uff09<\/li>\n <\/ul>\n <pre><code class=\"language-python\">pipe.enable_xformers_memory_efficient_attention()\npipe.vae.to(dtype=torch.float16)\npipe.unet.to(memory_format=torch.channels_last)\npipe = torch.compile(pipe) # ROCm\/CUDA \u3044\u305a\u308c\u3082\u6709\u52b9\u306a\u3053\u3068\u304c\u591a\u3044\n<\/code><\/pre>\n\n <h3>3.3 OCR\uff08\u30ec\u30a4\u30a2\u30a6\u30c8 + \u8a8d\u8b58\uff09<\/h3>\n <ul>\n <li><strong>\u30ec\u30a4\u30a2\u30a6\u30c8\u5206\u5272<\/strong>\uff1a\u6bb5\u843d\/\u8868\/\u56f3\u3092\u5148\u306b\u5207\u308a\u5206\u3051\u3066\u500b\u5225\u6700\u9069\uff08\u5c0f\u578b\u691c\u51fa+CRNN\/ViT\uff09<\/li>\n <li><strong>\u89e3\u50cf\u5ea6\u81ea\u52d5<\/strong>\uff1a\u7d30\u5b57\/\u5c0f\u6587\u5b57\u306f DPI \u3092\u4e0a\u3052\u3066\u5c40\u6240\u518d\u30b9\u30ad\u30e3\u30f3<\/li>\n <li><strong>OpenVINO INT8<\/strong>\uff1aCPU\u306e\u307f\u3067\u53ef\u3002Edge\/\u30b5\u30fc\u30d0\u6df7\u5728\u306b\u5f37\u3044<\/li>\n <li><strong>\u8a00\u8a9e\u30e2\u30fc\u30c9\u5206\u96e2<\/strong>\uff1a\u82f1\u6570\u5b57\/\u65e5\u672c\u8a9e\/\u4e2d\u97d3\u3067\u30e2\u30c7\u30eb\u5207\u66ff\uff08\u96c6\u5408\u77e5\u3088\u308a\u901f\u3044\uff09<\/li>\n <\/ul>\n <pre><code class=\"language-python\"># onnxruntime + OpenVINO EP \u3067 CPU INT8 \u3092\u4f7f\u3046\u4f8b\nsess = ort.InferenceSession(\"ocr_int8.onnx\", providers=[\"OpenVINOExecutionProvider\"])\n<\/code><\/pre>\n\n <h3>3.4 \u691c\u7d22\uff08\u57cb\u3081\u8fbc\u307f + ANN + \u518d\u30e9\u30f3\u30ad\u30f3\u30b0\uff09<\/h3>\n <ul>\n <li><strong>\u8efd\u91cf\u57cb\u3081\u8fbc\u307f<\/strong>\uff1a\u6b21\u5143\u6570 768\u2192384 \u3078\u84b8\u7559\u3002INT8\/FP16 \u3067\u5145\u5206\u3002<\/li>\n <li><strong>ANN<\/strong>\uff1aHNSW or IVF+PQ\uff08OPQ\uff09\u3067\u30e1\u30e2\u30ea\uff06\u30c7\u30a3\u30b9\u30af\u524a\u6e1b\u3002<\/li>\n <li><strong>\u518d\u30e9\u30f3\u30af<\/strong>\uff1aCross-Encoder \u306f <em>\u30c8\u30c3\u30d7K\u306e\u307f<\/em> \u306b\u9650\u5b9a\uff081\u301c5\u4ef6\uff09\u3002<\/li>\n <li><strong>\u30ad\u30e3\u30c3\u30b7\u30e5<\/strong>\uff1a\u983b\u51fa\u30af\u30a8\u30ea\/\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u57cb\u3081\u8fbc\u307f\u3092\u30e1\u30e2\u30ea\/SSD\u30ad\u30e3\u30c3\u30b7\u30e5\u3002<\/li>\n <\/ul>\n <pre><code class=\"language-python\"># FAISS: IVF+PQ \u306e\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u4f8b\nimport faiss\nimport numpy as np\n\nD = 384; nlist = 4096; m = 64; # 64x8bit = 512bit\/\u30d9\u30af\u30c8\u30eb\nquantizer = faiss.IndexFlatL2(D)\nindex = faiss.IndexIVFPQ(quantizer, D, nlist, m, 8)\nindex.train(train_vecs.astype('float32'))\nindex.add(base_vecs.astype('float32'))\nD,I = index.search(query_vecs.astype('float32'), 20)\n<\/code><\/pre>\n<\/section>\n\n<section>\n <h2>4. \u91cf\u5b50\u5316\u30fb\u84b8\u7559\u306e\u5b9f\u6226\u30ec\u30b7\u30d4<\/h2>\n <h3>4.1 PTQ\uff08\u5f8c\u4ed8\u3051\u91cf\u5b50\u5316\uff09<\/h3>\n <ul>\n <li>\u6821\u6b63\u30bb\u30c3\u30c8\u3092 <em>\u672c\u756a\u30c7\u30fc\u30bf\u5206\u5e03<\/em> \u304b\u3089\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\uff08\u8aa4\u5dee\u306f\u5206\u5e03\u30ba\u30ec\u304c\u4e3b\u56e0\uff09<\/li>\n <li>\u7cbe\u5ea6\u60aa\u5316\u304c\u5927\u304d\u3044\u5c64\u306f <em>\u6df7\u5408\u7cbe\u5ea6<\/em>\uff08INT8\u21d4FP16\uff09\u3067\u6551\u6e08<\/li>\n <\/ul>\n <h3>4.2 QAT\uff08\u91cf\u5b50\u5316\u5bfe\u5fdc\u5b66\u7fd2\uff09<\/h3>\n <pre><code class=\"language-python\"># PyTorch\uff1aFakeQuant \u3092\u7d44\u307f\u8fbc\u307f\u518d\u5b66\u7fd2\nfrom torch.ao.quantization import get_default_qat_qconfig, prepare_qat, convert\nmodel.qconfig = get_default_qat_qconfig(\"fbgemm\")\nprepare_qat(model, inplace=True)\n# \u3053\u3053\u3067\u5fae\u8abf\u6574\u5b66\u7fd2 ...\nmodel = convert(model.eval())\n<\/code><\/pre>\n <h3>4.3 \u84b8\u7559\uff08\u6559\u5e2b=\u5143\u30e2\u30c7\u30eb, \u751f\u5f92=\u8efd\u91cf\uff09<\/h3>\n <pre><code class=\"language-python\"># \u5178\u578b\u30ed\u30b9\uff1a\u03b1*SoftTarget(KL) + \u03b2*HardCE + \u03b3*FeatureMatch\nloss = alpha*kl(soft_student, soft_teacher) + beta*ce(y_s, y) + gamma*mse(f_s, f_t)\n<\/code><\/pre>\n <p class=\"note\">\u9375\u306f <strong>\u8a55\u4fa1\u306e\u81ea\u52d5\u5316<\/strong>\uff08\u7cbe\u5ea6\u2194\u901f\u5ea6\u306e Pareto \u66f2\u7dda\u3092\u65e5\u6b21\u3067\u66f4\u65b0\uff09\u3002\u7cbe\u5ea6\u304c\u95be\u5024\u3092\u5272\u3089\u306a\u3044\u6700\u3082\u5b89\u3044\u69cb\u6210\u3092\u5e38\u306b\u9078\u3076\u3002<\/p>\n<\/section>\n\n<section>\n <h2>5. \u914d\u4fe1\u30fb\u30aa\u30fc\u30b1\u30b9\u30c8\u30ec\u30fc\u30b7\u30e7\u30f3<\/h2>\n <h3>5.1 Triton Inference Server<\/h3>\n <ul>\n <li>ONNX\/TensorRT\/PyTorch\/Python backend \u3092\u5358\u4e00\u30a8\u30f3\u30c9\u30dd\u30a4\u30f3\u30c8\u3067\u675f\u306d\u308b<\/li>\n <li>\u52d5\u7684\u30d0\u30c3\u30c1 \/ \u540c\u6642\u5b9f\u884c \/ \u30e2\u30c7\u30eb\u9593 Ensemble \/ A\/B \u30c6\u30b9\u30c8<\/li>\n <\/ul>\n <pre><code class=\"language-ini\"># model_repository\/layout\uff08\u4f8b\uff09\nrepo\/\n detector\/\n 1\/model.onnx\n config.pbtxt # max_batch_size, dynamic_batching, instance_group\n ocr\/\n 1\/model.xml # OpenVINO\n config.pbtxt\n<\/code><\/pre>\n\n <h3>5.2 vLLM \/ Text Generation Inference\uff08LLM\/Reranker\uff09<\/h3>\n <ul>\n <li>PagedAttention \u7b49\u3067\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8\u6539\u5584\u3001Speculative Decoding \u6709\u52b9<\/li>\n <li>Embedding \u306f onnxruntime \/ openvino \u3067\u5225\u30e9\u30a4\u30f3\u306b\u9003\u304c\u3059<\/li>\n <\/ul>\n\n <h3>5.3 \u76e3\u8996\u30fbSLO<\/h3>\n <ul>\n <li>p50\/p95 \u30ec\u30a4\u30c6\u30f3\u30b7\u3001GPU\/CPU \u5229\u7528\u7387\u3001\u30d0\u30c3\u30c1\u30b5\u30a4\u30ba\u5206\u5e03\u3001\u30a8\u30e9\u30fc\u7387<\/li>\n <li>\u54c1\u8cea\u76e3\u8996\uff1amAP\/CER\/nDCG \u3092\u30aa\u30f3\u30e9\u30a4\u30f3\u8a55\u4fa1\uff08\u5c0f\u898f\u6a21\u30b5\u30f3\u30d7\u30eb\u3067\uff09<\/li>\n <\/ul>\n<\/section>\n\n<section>\n <h2>6. \u30b3\u30b9\u30c8\u5f0f\u3068\u610f\u601d\u6c7a\u5b9a\u30d5\u30ed\u30fc<\/h2>\n <pre><code>\u30b3\u30b9\u30c8\/\u30ea\u30af\u30a8\u30b9\u30c8 \u2252 (\u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u6642\u7d66 + \u96fb\u529b) \/ (\u6709\u52b9\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8 req\/s) + \u30b9\u30c8\u30ec\u30fc\u30b8\u8ee2\u9001\n\u6709\u52b9\u30b9\u30eb\u30fc\u30d7\u30c3\u30c8 = f(\u52d5\u7684\u30d0\u30c3\u30c1, \u89e3\u50cf\u5ea6, \u91cf\u5b50\u5316, \u524d\u5f8c\u51e6\u7406\u306e\u4e26\u5217\u5316)\n<\/code><\/pre>\n <ol>\n <li>\u307e\u305a FP16 \u30d9\u30fc\u30b9\u30e9\u30a4\u30f3\uff08CUDA\/ROCm\/OV \u3044\u305a\u308c\u304b\uff09\u3092\u4f5c\u308b<\/li>\n <li>INT8 PTQ \u2192 \u7cbe\u5ea6\u78ba\u8a8d \u2192 \u3060\u3081\u306a\u3089\u5c64\u5358\u4f4d\u3067\u6df7\u5408\u7cbe\u5ea6<\/li>\n <li>\u524d\u51e6\u7406\u30c0\u30a6\u30f3\u30b5\u30f3\u30d7\u30eb\/\u52d5\u7684\u30d0\u30c3\u30c1\/\u30ad\u30e3\u30c3\u30b7\u30e5\u5c0e\u5165<\/li>\n <li>\u5fc5\u8981\u306a\u3089 QAT\/\u84b8\u7559\u3067\u7cbe\u5ea6\u56de\u5fa9<\/li>\n <li>\u30af\u30e9\u30a6\u30c9\u5c02\u7528HW\u3067\u591c\u9593\u30d0\u30c3\u30c1\u3092\u9003\u304c\u3059\uff08\u30b9\u30dd\u30c3\u30c8\u6d3b\u7528\uff09<\/li>\n <\/ol>\n<\/section>\n\n<section>\n <h2>7. \u5177\u4f53\u7684\u30c6\u30f3\u30d7\u30ec\uff08\u305d\u306e\u307e\u307e\u6d41\u7528OK\uff09<\/h2>\n <h3>7.1 REST \u63a8\u8ad6\uff08onnxruntime + \u52d5\u7684EP\uff09<\/h3>\n <pre><code class=\"language-python\">from fastapi import FastAPI\nimport onnxruntime as ort\nimport numpy as np\n\nproviders=[\"TensorrtExecutionProvider\",\"CUDAExecutionProvider\",\n \"ROCmExecutionProvider\",\"OpenVINOExecutionProvider\",\n \"CPUExecutionProvider\"]\nsess = ort.InferenceSession(\"model.onnx\", providers=providers)\n\napp = FastAPI()\n@app.post(\"\/infer\")\ndef infer(x: list[float]):\n a = np.array(x, dtype=np.float32)[None, :]\n y = sess.run(None, {sess.get_inputs()[0].name: a})[0]\n return {\"y\": y.tolist()}\n<\/code><\/pre>\n\n <h3>7.2 Triton\uff08\u52d5\u7684\u30d0\u30c3\u30c1\uff09<\/h3>\n <pre><code class=\"language-proto\"># config.pbtxt\uff08\u629c\u7c8b\uff09\nname: \"detector\"\nplatform: \"onnxruntime_onnx\"\nmax_batch_size: 32\ndynamic_batching { preferred_batch_size: [4, 8, 16] max_queue_delay_microseconds: 2000 }\ninstance_group [{ kind: KIND_GPU, count: 1 }]\n<\/code><\/pre>\n\n <h3>7.3 OpenVINO INT8 OCR<\/h3>\n <pre><code class=\"language-bash\">python -m openvino.tools.ovc --input_model ocr.onnx --compress_to_fp16 -o ov_fp16\ncat pot.json | jq . # \u4e0a\u306e\u4f8b\u3092\u53c2\u7167\npot -c pot.json -d ov_int8\n<\/code><\/pre>\n<\/section>\n\n<section>\n <h2>8. \u54c1\u8cea\u3092\u843d\u3068\u3055\u305a\u30b3\u30b9\u30c8\u3092\u843d\u3068\u3059\u300c\u5c0f\u30ef\u30b6\u300d\u96c6<\/h2>\n <ul>\n <li><strong>\u89e3\u50cf\u5ea6\u30a2\u30c0\u30d7\u30c8<\/strong>\uff1a\u691c\u51fa\u7d50\u679c\u304c\u5b89\u5b9a\u3059\u308b\u6700\u5c0f\u89e3\u50cf\u5ea6\u3092\u30aa\u30f3\u30e9\u30a4\u30f3\u3067\u63a2\u7d22\u30fb\u4fdd\u5b58<\/li>\n <li><strong>\u65e9\u671f\u6253\u3061\u5207\u308a<\/strong>\uff1a\u751f\u6210\u306f <em>\u30b9\u30c6\u30c3\u30d7\u6570\u4e0a\u9650<\/em> \u3092\u52d5\u7684\u5236\u5fa1\u3001LLM \u306f speculative<\/li>\n <li><strong>\u30ad\u30e3\u30c3\u30b7\u30e5<\/strong>\uff1a\u518d\u5165\u529b\u306e\u591a\u3044 OCR \u30da\u30fc\u30b8\u3084\u691c\u7d22\u30af\u30a8\u30ea\u3092 TTL \u30ad\u30e3\u30c3\u30b7\u30e5<\/li>\n <li><strong>\u5206\u5c90<\/strong>\uff1a\u8efd\u91cf\u30e2\u30c7\u30eb\u2192\u5fae\u5999\u306a\u3089\u91cd\u3044\u30e2\u30c7\u30eb\u3067\u518d\u5224\u5b9a\uff08Two-Stage\uff09<\/li>\n <li><strong>\u30b9\u30b1\u30fc\u30eb<\/strong>\uff1a\u958b\u5e97\u30fb\u663c\u30fb\u591c\u3067\u30ec\u30d7\u30ea\u30ab\u6570\u3092\u81ea\u52d5\u5207\u66ff\uff08HPA\/\u30b9\u30b1\u30b8\u30e5\u30fc\u30e9\uff09<\/li>\n <\/ul>\n<\/section>\n\n<section>\n <h2>9. \u65e2\u5b58\u8cc7\u7523\u3078\u306e\u9069\u7528\u9806\u5e8f\uff08\u6700\u77ed2\u9031\u9593\u30d7\u30e9\u30f3\uff09<\/h2>\n <ol>\n <li>\u73fe\u72b6\u306e p95, req\/s, mAP\/CER \u3092\u56fa\u5b9a\u5316\u8a08\u6e2c\uff08\u30d9\u30fc\u30b9\u30e9\u30a4\u30f3\uff09<\/li>\n <li>ONNX \u3078\u6b63\u898f\u5316 \u2192 onnxruntime \u3067 EP \u53ef\u5909\u306b<\/li>\n <li>INT8 PTQ \u2192 \u7cbe\u5ea6\u691c\u8a3c \u2192\u30c0\u30e1\u7b87\u6240\u306e\u307f\u6df7\u5408\u7cbe\u5ea6 \/ QAT<\/li>\n <li>Triton \u5c0e\u5165\uff08\u52d5\u7684\u30d0\u30c3\u30c1\u30fbA\/B\uff09\u2192 \u30b3\u30b9\u30c8\u6700\u5c0f\u8a2d\u5b9a\u3092\u63a1\u629e<\/li>\n <li>\u91cd\u3044\u30ef\u30fc\u30af\u30ed\u30fc\u30c9\u3092\u591c\u9593\u3060\u3051\u30af\u30e9\u30a6\u30c9\u5c02\u7528HW\u3078\u9003\u304c\u3059<\/li>\n <\/ol>\n <p class=\"warn\">\u3088\u304f\u3042\u308b\u5931\u6557\uff1a<strong>\u6821\u6b63\u30c7\u30fc\u30bf\u304c\u30b7\u30e7\u30dc\u3044<\/strong>\uff08\u2192INT8\u304c\u30dc\u30ed\u30dc\u30ed\uff09\u3002<strong>\u524d\u51e6\u7406\u30fb\u5f8c\u51e6\u7406\u304cCPU\u306b\u6b8b\u7559<\/strong>\uff08\u2192GPU\u904a\u3076\uff09\u3002<strong>p50\u3060\u3051\u6539\u5584<\/strong>\uff08\u2192p95\u60aa\u5316\u3067\u4f53\u611f\u30de\u30a4\u30ca\u30b9\uff09\u3002<\/p>\n<\/section>\n\n<section>\n <h2>10. \u4ed8\u9332\uff1a\u7528\u8a9e\u3056\u3063\u304f\u308a<\/h2>\n <ul>\n <li><strong>EP\uff08Execution Provider\uff09<\/strong>\uff1aONNX Runtime \u306b\u304a\u3051\u308b\u5b9f\u884c\u30a8\u30f3\u30b8\u30f3\u306e\u5207\u66ff\u30d7\u30e9\u30b0\u30a4\u30f3<\/li>\n <li><strong>PTQ\/QAT<\/strong>\uff1a\u5f8c\u4ed8\u91cf\u5b50\u5316 \/ \u91cf\u5b50\u5316\u5bfe\u5fdc\u5b66\u7fd2<\/li>\n <li><strong>IVF+PQ\/OPQ<\/strong>\uff1a\u30d9\u30af\u30c8\u30eb\u5727\u7e2e\u306e\u5b9a\u756a\u3002\u5de8\u5927\u30b3\u30fc\u30d1\u30b9\u306e\u691c\u7d22\u3092\u5b89\u4fa1\u306b<\/li>\n <li><strong>Speculative Decoding<\/strong>\uff1a\u5c0f\u30e2\u30c7\u30eb\u3067\u5148\u8aad\u307f\u3057\u5927\u30e2\u30c7\u30eb\u3067\u691c\u8a3c\u3059\u308b\u9ad8\u901f\u5316<\/li>\n <\/ul>\n<\/section>\n\n<\/main>\n\n<footer>\n <p>\u6700\u7d42\u66f4\u65b0: 2025-09-01 \/ \u4f5c\u6210\u8005: \u30d3\u30fc\u30ca\u30ec\u30c3\u30b8\u30c7\u30b6\u30a4\u30f3<\/p>\n<\/footer>\n<\/body>\n<\/html>\n\n","protected":false},"excerpt":{"rendered":"<p>\u63a8\u8ad6\u30b3\u30b9\u30c8\u6700\u9069\u5316\u30d7\u30ec\u30a4\u30d6\u30c3\u30af \u524d\u63d0\uff1a\u30aa\u30f3\u30d7\u30ec\uff08\u4f8b\uff1aRTX 4070 Ti SUPER \u00d71\uff09\u307e\u305f\u306f\u30af\u30e9\u30a6\u30c9\u306e\u6df7\u5728\u3092\u60f3\u5b9a\u3002CUDA \u3060\u3051\u306b\u7e1b\u3089\u308c\u305a\u3001ROCm \/ OpenVINO \/ ONNX Runtime \/ Tens […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"saved_in_kubio":false,"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"_uf_show_specific_survey":0,"_uf_disable_surveys":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"sns_share_botton_hide":"","vkExUnit_sns_title":"","_vk_print_noindex":"","sitemap_hide":"","vkExUnit_EyeCatch_disable":"","_veu_custom_css":"","veu_display_promotion_alert":"common","vkexunit_cta_each_option":"","_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_post_was_ever_published":false},"categories":[5,68,24],"tags":[],"class_list":["post-1633","post","type-post","status-publish","format-standard","hentry","category-ai","category-68","category-24"],"aioseo_notices":[],"veu_head_title_object":{"title":"","add_site_title":""},"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/posts\/1633","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1633"}],"version-history":[{"count":3,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/posts\/1633\/revisions"}],"predecessor-version":[{"id":1636,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=\/wp\/v2\/posts\/1633\/revisions\/1636"}],"wp:attachment":[{"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1633"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1633"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/beeknowledge.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1633"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}