{"id":1520,"date":"2026-04-11T21:14:35","date_gmt":"2026-04-11T12:14:35","guid":{"rendered":"http:\/\/www.shooting-bios.net\/wordpress\/?p=1520"},"modified":"2026-04-12T09:36:32","modified_gmt":"2026-04-12T00:36:32","slug":"%e3%80%90llama-cpp%e3%80%91%e6%9c%80%e6%96%b0%e3%81%aekv%e3%82%ad%e3%83%a3%e3%83%83%e3%82%b7%e3%83%a5%e9%87%8f%e5%ad%90%e5%8c%96%e6%8a%80%e8%a1%93turboquant%e3%81%ae%e5%b0%8e%e5%85%a5%ef%bc%9avram","status":"publish","type":"post","link":"http:\/\/www.shooting-bios.net\/wordpress\/2026\/04\/11\/%e3%80%90llama-cpp%e3%80%91%e6%9c%80%e6%96%b0%e3%81%aekv%e3%82%ad%e3%83%a3%e3%83%83%e3%82%b7%e3%83%a5%e9%87%8f%e5%ad%90%e5%8c%96%e6%8a%80%e8%a1%93turboquant%e3%81%ae%e5%b0%8e%e5%85%a5%ef%bc%9avram\/","title":{"rendered":"\u3010llama.cpp\u3011\u6700\u65b0\u306eKV\u30ad\u30e3\u30c3\u30b7\u30e5\u91cf\u5b50\u5316\u6280\u8853TurboQuant\u306e\u5c0e\u5165\uff1aVRAM\u4f7f\u7528\u91cf\u4f4e\u6e1b\uff0b\u5fdc\u7b54\u901f\u5ea6\u306e\u5287\u7684\u5411\u4e0a"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">Gemma 4\u3063\u3066\u3081\u3063\u3061\u3083\u30af\u30ec\u30d0\u30fc\u3002\u3060\u304c\u3057\u304b\u3057<\/h2>\n\n\n\n<p>2026\u5e744\u6708\u306bGoogle\u304c\u30ea\u30ea\u30fc\u30b9\u3057\u305fGemma 4\u3067\u3059\u304c\u3001\u306a\u304b\u306a\u304b\u30be\u30af\u30be\u30af\u3059\u308b\u307b\u3069\u77e5\u7684\u306a\u4f1a\u8a71\u304c\u697d\u3057\u3081\u3066\u3001\u30ed\u30fc\u30ab\u30ebLLM\u3082\u3064\u3044\u306b\u3053\u3053\u307e\u3067\u6765\u305f\u304b\u3001\u3068\u611f\u6168\u6df1\u304f\u306a\u3063\u3066\u304a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n<p class=\"has-large-font-size\">\u304c\u3001\u9045\u3044\u3002\u30ac\u30c1\u3067\u9045\u3044\u3002<\/p>\n\n\n\n<p>\u30d5\u30a1\u30fc\u30b9\u30c8\u30c8\u30fc\u30af\u30f3\u304c\u8fd4\u3063\u3066\u304f\u308b\u307e\u30671\u5206\uff5e\u6570\u5206\u5f85\u305f\u3055\u308c\u307e\u3059\u3002\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u304c1\u756a\u591a\u304431B\u7248\u306e\u8a71\u3068\u306f\u3044\u3048\u3001RTX 5090\u3092\u4f7f\u3063\u3066\u3082\u6570\u5206\u5f85\u305f\u3055\u308c\u3001\u30c8\u30fc\u30af\u30f3\u751f\u6210\u901f\u5ea6\u306f10\uff5e15token\/\u79d2\u3061\u3087\u3044\u3002\u3053\u308c\u306f\u3069\u3093\u306a\u306b\u512a\u79c0\u306aLLM\u3060\u308d\u3046\u304c\u840e\u3048\u307e\u3059\u3002\u305b\u3063\u304b\u3061\u306a\u81ea\u5206\u306b\u3068\u3063\u3066\u306f\u975e\u5e38\u306b\u304d\u3064\u3044\u3002<\/p>\n\n\n\n<p>\u3057\u304b\u3057\uff01Gemma 4\u306e\u30ea\u30ea\u30fc\u30b9\u306b\u5148\u7acb\u3064\u3053\u3068\uff11\u9031\u9593\u307b\u3069\u524d\u306b\u3001\u300cTurboQuant\u300d\u3068\u3044\u3046KV\u30ad\u30e3\u30c3\u30b7\u30e5\u306e\u30e1\u30e2\u30ea\u4f7f\u7528\u91cf\u3092\u5287\u7684\u306b\u524a\u6e1b\u3059\u308b\u6700\u65b0\u306e\u91cf\u5b50\u5316\u6280\u8853\u304c\u767a\u8868\u3055\u308c\u305f\u306e\u3067\u3042\u308a\u307e\u3059\u3002<br><a rel=\"noreferrer noopener\" href=\"https:\/\/pc.watch.impress.co.jp\/docs\/news\/2097004.html\" target=\"_blank\">Google\u3001LLM\u5b9f\u884c\u6642\u306e\u6d88\u8cbb\u30e1\u30e2\u30ea\u30926\u5206\u306e1\u306b\u524a\u6e1b\u3059\u308b\u300cTurboQuant\u300d<\/a>(PC Watch)<\/p>\n\n\n\n<p>\u3061\u3087\u3063\u3068LLM\u3092\u304b\u3058\u3063\u305f\u3053\u3068\u306e\u3042\u308b\u65b9\u306a\u3089\u3001\u300c\u3048\uff1f\u91cf\u5b50\u5316\uff1f\u3069\u3046\u305b\u56de\u7b54\u306e\u54c1\u8cea\u304c\u843d\u3061\u308b\u3093\u3058\u3083\u306a\u3044\u306e\uff1f\u300d\u3068\u3001\u602a\u8a1d\u306a\u9854\u3064\u304d\u306b\u306a\u3063\u3066\u3057\u307e\u3046\u304b\u3082\u3057\u308c\u307e\u305b\u3093\u304c\u3001\u7d20\u6674\u3089\u3057\u3044\u6280\u8853(\u8a9e\u5f59\u529b\u30fb\u30fb\u30fb)\u306b\u3088\u3063\u3066\u3001\u91cf\u5b50\u5316\u306b\u3088\u308b\u54c1\u8cea\u4f4e\u4e0b\u306f\u7121\u8996\u3067\u304d\u308b\u7a0b\u5ea6\u306b\u8efd\u6e1b\u3057\u3066\u3044\u308b\u3068\u306e\u3053\u3068\u3067\u3059\u3002<\/p>\n\n\n\n<p>\u4eca\u5ea6\u306f\u3001\u300c\u3048\uff1f\u3067\u3082\u3059\u3050\u306b\u306f\u4f7f\u3048\u306a\u3044\u3093\u3058\u3083\u306a\u3044\u306e\uff1f1\u5e74\u5148\u3068\u304b\u306a\u3089\u7d75\u306b\u63cf\u3044\u305f\u9905\u3063\u3059\u3088\u300d\u3068\u8003\u3048\u308b\u65b9\u3082\u3044\u3089\u3063\u3057\u3083\u308b\u304b\u3082\u3057\u308c\u307e\u305b\u3093\u3002\u3068\u3044\u3046\u304b\u79c1\u304c\u305d\u3046\u601d\u3063\u305f\u8a33\u3067\u3059\u304c\u3001<\/p>\n\n\n\n<p class=\"has-large-font-size\">\u3082\u3046\u4f7f\u3048\u308b\u3093\u3067\u3059\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<p>\u306a\u3093\u3068\u3044\u3046\u3053\u3068\u3067\u3057\u3087\u3046\u3002\u3073\u3063\u304f\u308a\u3067\u3059\u306d\u3002TurboQuant\u7248\u3092\u30d5\u30a9\u30fc\u30af\u3057\u3066\u304f\u308c\u305f\u4eba(\u4eba\u3005\uff1f)\u306b\u306f\u611f\u8b1d\u3057\u304b\u3042\u308a\u307e\u305b\u3093\u3002KV\u30ad\u30e3\u30c3\u30b7\u30e5\u306e\u30b5\u30a4\u30ba\u304c\u7e2e\u5c0f\u3059\u308b\u3068\u3044\u3046\u3053\u3068\u306f\u3001\u3088\u308a\u5927\u304d\u306a\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u3082\u6301\u3064\u30e2\u30c7\u30eb\u3092VRAM\u306b\u8f09\u305b\u308b\u3053\u3068\u304c\u51fa\u6765\u308b\u3001\u307e\u305f\u3001\u5404\u7a2e\u8a08\u7b97\u91cf\u304c\u6e1b\u308b\u305f\u3081\u30c8\u30fc\u30af\u30f3\u751f\u6210\u901f\u5ea6\u304c\u5411\u4e0a\u3059\u308b\u3001\u3068\u826f\u3044\u3053\u3068\u3065\u304f\u3081\u3067\u3059\u3002\u3069\u306e\u304f\u3089\u3044\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u304c\u5411\u4e0a\u3059\u308b\u304b\u306f\u5f8c\u307b\u3069\u89e6\u308c\u308b\u3068\u3057\u3066\u3001\u5c0e\u5165\u624b\u9806\u3092\u3054\u6848\u5185\u3057\u307e\u3059\u3002\u3042\u306a\u305f\u3082\u6700\u65b0\u306eKV\u30ad\u30e3\u30c3\u30b7\u30e5\u91cf\u5b50\u5316\u6280\u8853\u3001\u300cTurboQuant\u300d\u306e\u6069\u6075\u3092\u53d7\u3051\u307e\u3057\u3087\u3046\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u524d\u63d0\u6761\u4ef6<\/h2>\n\n\n\n<p>\u8272\u3005\u306a\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u74b0\u5883\u304c\u5b58\u5728\u3059\u308b\u306a\u304b\u3067\u3001\u3059\u3079\u3066\u306e\u74b0\u5883\u306b\u5bfe\u5fdc\u3059\u308b\u5c0e\u5165\u624b\u9806\u3092\u8a18\u8f09\u3059\u308b\u306e\u306f\u8a00\u3046\u307e\u3067\u3082\u306a\u304f\u4e0d\u53ef\u80fd\u3067\u3059\u3002\u30ed\u30fc\u30ab\u30eb\u3067LLM\u3092\u52d5\u304b\u3059\u74b0\u5883\u3068\u3057\u3066\u306f2026\u5e744\u6708\u73fe\u5728\u3001\u6bd4\u8f03\u7684\u4e00\u822c\u7684\u3060\u3068\u601d\u308f\u308c\u308b\u3001\u4ee5\u4e0b\u306e\u74b0\u5883\u306b\u3064\u3044\u3066\u8aac\u660e\u3092\u3057\u3066\u3044\u304d\u307e\u3059\u3002<\/p>\n\n\n\n<ul>\n<li>WSL2\u4e0a\u306e Ubuntu 24.04 (\u30cd\u30a4\u30c6\u30a3\u30d6\u306eUbuntu\u3067\u3082\u6050\u3089\u304f\u307b\u307c\u540c\u3058\u624b\u9806\u3067\u52d5\u304f\u3068\u601d\u308f\u308c\u308b)<\/li>\n<\/ul>\n\n\n\n<ul>\n<li>GPU\u306f Geforce RTX 5090 \u307e\u305f\u306f Geforce RTX 3060 12GB<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u5c0e\u5165\u624b\u9806<\/h2>\n\n\n\n<p>\u307e\u305a\u306f\u4efb\u610f\u306e\u5834\u6240\u306b\u4f5c\u696d\u30c7\u30a3\u30ec\u30af\u30c8\u30ea\u3092\u4f5c\u6210\u3057\u307e\u3059\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\n$ mkdir workspace\n$ cd workspace\n<\/pre>\n\n\n<p>cuda-toolkit\u3092\u30bb\u30c3\u30c8\u30a2\u30c3\u30d7\u3057\u307e\u3059\u30022026\u5e744\u6708\u73fe\u5728\u30d0\u30fc\u30b8\u30e7\u30f313.2\u307e\u3067\u51fa\u3066\u3044\u307e\u3059\u304c\u3001unsloth\u306eGemma 4\u306e<a rel=\"noreferrer noopener\" href=\"https:\/\/unsloth.ai\/docs\/jp\/moderu\/gemma-4\" target=\"_blank\">\u30e2\u30c7\u30eb\u30da\u30fc\u30b8<\/a>\u3067\u30d0\u30fc\u30b8\u30e7\u30f313.2\u304c\u975e\u63a8\u5968\u3067\u3042\u308b\u65e8\u306e\u8a18\u8f09\u304c\u3042\u308b\u3053\u3068\u3068\u300113.2\u3067\u306f\u79c1\u81ea\u8eabllama.cpp\u306e\u30d3\u30eb\u30c9\u3092\u3046\u307e\u304f\u3067\u304d\u307e\u305b\u3093\u3067\u3057\u305f\u306e\u3067\u3001\u30d0\u30fc\u30b8\u30e7\u30f312.8\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3057\u307e\u3059\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\n$ wget https:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/wsl-ubuntu\/x86_64\/cuda-wsl-ubuntu.pin\n$ sudo mv cuda-wsl-ubuntu.pin \/etc\/apt\/preferences.d\/cuda-repository-pin-600\n$ wget https:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/wsl-ubuntu\/x86_64\/cuda-keyring_1.1-1_all.deb\n$ sudo dpkg -i cuda-keyring_1.1-1_all.deb\n$ sudo apt update\n$ sudo apt -y install cuda-toolkit-12.8\n$ rm -f cuda-keyring_1.1-1_all.deb\n\n<\/pre>\n\n\n<p>nvcc\u30b3\u30de\u30f3\u30c9\u306b\u30d1\u30b9\u3092\u901a\u3057\u307e\u3059\u3002\u30c6\u30ad\u30b9\u30c8\u30a8\u30c7\u30a3\u30bf\u306f\u304a\u597d\u304d\u306a\u3082\u306e\u3067\u3002\u4ee5\u4e0b\u3067\u306fnano\u3092\u4f7f\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\n$ nano ~\/.bashrc\n\u4ee5\u4e0b\u306e2\u884c\u3092\u8ffd\u8a18\nexport PATH=\/usr\/local\/cuda\/bin:$PATH\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64:$LD_LIBRARY_PATH\n\n.bashrc\u3092\u518d\u8aad\u307f\u8fbc\u307f\n$ source ~\/.bashrc\n<\/pre>\n\n\n<p>\u30d0\u30fc\u30b8\u30e7\u30f312.8\u304c\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3055\u308c\u305f\u3053\u3068\u3092\u78ba\u8a8d\u3057\u307e\u3059\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\n$ nvcc --version\nnvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2025 NVIDIA Corporation\nBuilt on Fri_Feb_21_20:23:50_PST_2025\nCuda compilation tools, release 12.8, V12.8.93\nBuild cuda_12.8.r12.8\/compiler.35583870_0\n<\/pre>\n\n\n<p>\u6b21\u306bTurboQuant\u5bfe\u5fdc\u7248\u306e llama.cpp \u3092\u30bd\u30fc\u30b9\u304b\u3089\u30d3\u30eb\u30c9\u3057\u307e\u3059\u3002git\u304c\u5165\u3063\u3066\u3044\u306a\u3044\u65b9\u306f\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3057\u3066\u304a\u3044\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n\n<p>TurboQuant\u5bfe\u5fdc\u7248\u3092\u958b\u767a\u3057\u3066\u3044\u308b\u30d6\u30e9\u30f3\u30c1\u3092\u30af\u30ed\u30fc\u30f3\u3057\u307e\u3059\u3002 <\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\n$ git clone -b feature\/turboquant-kv-cache --recursive https:\/\/github.com\/TheTom\/llama-cpp-turboquant.git\n$ cd llama-cpp-turboquant\/\n$ mkdir gguf\n<\/pre>\n\n\n<p>\u30e2\u30c7\u30eb\u30d5\u30a1\u30a4\u30eb\u3092\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u3057\u307e\u3059\u3002RTX 5090\u306a\u308931B\u306e4\u30d3\u30c3\u30c8\u91cf\u5b50\u5316\u30e2\u30c7\u30eb\u3001RTX 3060 12GB\u306a\u3089E4B\u306e6\u30d3\u30c3\u30c8\u91cf\u5b50\u5316\u30e2\u30c7\u30eb\u8fba\u308a\u307e\u3067\u4f59\u88d5\u3067\u3059\u3002\u3082\u3063\u3068\u304e\u308a\u304e\u308a\u307e\u3067VRAM\u3092\u4f7f\u3044\u5207\u308a\u305f\u3044\u3068\u3044\u3046\u65b9\u306f\u3001\u3088\u308a\u653b\u3081\u305f\u30e2\u30c7\u30eb\u3092\u9069\u5b9c\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\nRTX 5090\u306e\u5834\u5408(31B\u306e\u5834\u5408)\n$ curl -L https:\/\/huggingface.co\/unsloth\/gemma-4-31B-it-GGUF\/resolve\/main\/gemma-4-31B-it-UD-Q4_K_XL.gguf?download=true -o .\/gguf\/gemma-4-31B-it-UD-Q4_K_XL.gguf\n\nRTX 3060 12GB\u306e\u5834\u5408(E4B\u306e\u5834\u5408)\n$ curl -L https:\/\/huggingface.co\/unsloth\/gemma-4-E4B-it-GGUF\/resolve\/main\/gemma-4-E4B-it-Q6_K.gguf?download=true -o .\/gguf\/gemma-4-E4B-it-Q6_K.gguf\n\n<\/pre>\n\n\n<p>\u304a\u597d\u307f\u3067\u30d3\u30b8\u30e7\u30f3\u30a8\u30f3\u30b3\u30fc\u30c0\u3092\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u3057\u307e\u3059\u3002\u3053\u308c\u3092\u4f7f\u3046\u3068\u753b\u50cf\u3092LLM\u306b\u5165\u529b\u3057\u3066\u3001\u753b\u50cf\u306b\u3064\u3044\u3066\u8aac\u660e\u3055\u305b\u308b\u3053\u3068\u304c\u51fa\u6765\u307e\u3059\u3002\u3044\u308f\u3086\u308b\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30ebLLM\u3067\u3059\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\nRTX 5090\u306e\u5834\u5408(31B\u306e\u5834\u5408)\n$ curl -L https:\/\/huggingface.co\/unsloth\/gemma-4-31B-it-GGUF\/resolve\/main\/mmproj-BF16.gguf?download=true -o .\/gguf\/mmproj-BF16.gguf\n\nRTX 3060 12GB\u306e\u5834\u5408(E4B\u306e\u5834\u5408)\n$ curl -L https:\/\/huggingface.co\/unsloth\/gemma-4-E4B-it-GGUF\/resolve\/main\/mmproj-BF16.gguf?download=true -o .\/gguf\/mmproj-BF16.gguf\n\n<\/pre>\n\n\n<p>llama.cpp\u306e\u30d3\u30eb\u30c9\u306e\u6e96\u5099\u3092\u3057\u307e\u3059\u3002cmake\u30b3\u30de\u30f3\u30c9\u3092\u4f7f\u3046\u306e\u3067\u30d1\u30c3\u30b1\u30fc\u30b8\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\n$ dpkg -l | grep -q &#039;cmake &#039; || sudo apt -y install cmake\n\nRTX 5090\u306e\u5834\u5408\n$ cmake -B build -DGGML_CUDA=ON -DGGML_NATIVE=OFF -DBUILD_SHARED_LIBS=OFF -DCMAKE_CUDA_ARCHITECTURES=120a -DGGML_CUDA_FA_ALL_QUANTS=ON\n\nRTX 3060 12GB\u306e\u5834\u5408\n$ cmake -B build -DGGML_CUDA=ON -DGGML_NATIVE=OFF -DBUILD_SHARED_LIBS=OFF -DCMAKE_CUDA_ARCHITECTURES=native -DGGML_CUDA_FA_ALL_QUANTS=ON\n\n<\/pre>\n\n\n<p>llama.cpp\u3092\u30d3\u30eb\u30c9\u3057\u307e\u3059\u3002PC\u306e\u30b9\u30da\u30c3\u30af\u306b\u4f9d\u5b58\u3057\u307e\u3059\u304c\u3001\u6570\u5206\u306f\u304b\u304b\u308a\u307e\u3059\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\n$ cmake --build build --config Release -j$(nproc)\n<\/pre>\n\n\n<p>\u3053\u3053\u307e\u3067\u306e\u624b\u9806\u3067llama.cpp\u304c\u4f7f\u3048\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3057\u305f\uff01<\/p>\n\n\n\n<p>\u4ee5\u4e0b\u306e\u624b\u9806\u3067\u306fllama.cpp\u3092\u30b5\u30fc\u30d0\u30e2\u30fc\u30c9\u3067\u8d77\u52d5\u3059\u308b\u624b\u9806\u3092\u8a18\u8f09\u3057\u307e\u3059\u3002Open WebUI\u306a\u3069\u3001\u304a\u597d\u307f\u306e\u30d5\u30ed\u30f3\u30c9\u30a8\u30f3\u30c9\u304b\u3089\u30a2\u30af\u30bb\u30b9\u3057\u3066\u307f\u3066\u304f\u3060\u3055\u3044\u3002reasoning\u30aa\u30d7\u30b7\u30e7\u30f3\u3084port\u30aa\u30d7\u30b7\u30e7\u30f3\u306f\u304a\u597d\u307f\u3067\u5909\u66f4\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\nRTX 5090\u306e\u5834\u5408(31B\u306e\u5834\u5408)\n$ .\/build\/bin\/llama-server -m gguf\/gemma-4-31B-it-UD-Q4_K_XL.gguf -mm gguf\/mmproj-BF16.gguf --n-gpu-layers 99 --no-mmap --flash-attn on --host 0.0.0.0 --port 8089 --cache-type-k turbo3 --cache-type-v turbo3 --reasoning off\n\nRTX 3060 12GB\u306e\u5834\u5408(E4B\u306e\u5834\u5408)\n$ .\/build\/bin\/llama-server -m gguf\/gemma-4-E4B-it-Q6_K.gguf -mm gguf\/mmproj-BF16.gguf --n-gpu-layers 99 --no-mmap --flash-attn on --host 0.0.0.0 --port 8089 --cache-type-k turbo3 --cache-type-v turbo3 --reasoning off\n<\/pre>\n\n\n<p>\u8d77\u52d5\u5f8c\u306e\u30ed\u30b0\u3067KV\u30ad\u30e3\u30c3\u30b7\u30e5\u30b5\u30a4\u30ba\u3092\u78ba\u8a8d\u3067\u304d\u307e\u3059\u3002\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u30a6\u30a3\u30f3\u30c9\u30a6192K\u3067\u3053\u306e\u30b5\u30a4\u30ba\u611f\u3067\u3059\u3002<\/p>\n\n\n<pre class=\"brush: plain; gutter: false; title: ; notranslate\" title=\"\">\nllama-server[12750]: llama_kv_cache: size = 3000.00 MiB (196608 cells,  10 layers,  4\/1 seqs), K (turbo3): 1500.00 MiB, V (turbo3): 1500.00 MiB\n<\/pre>\n\n\n<p>llama-server\u3092\u30b5\u30fc\u30d3\u30b9\u5316\u3057\u305f\u3044\u5834\u5408\u306f\u3001\u4ee5\u4e0b\u306e\u624b\u9806\u3092\u8ffd\u52a0\u3067\u884c\u3063\u3066\u304f\u3060\u3055\u3044\u3002WSL2\u8d77\u52d5\u6642\u306bllama-server\u3082\u8d77\u52d5\u3059\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002\u30e6\u30fc\u30b6\u30fc\u540d\u3084\u30e2\u30c7\u30eb\u30d5\u30a1\u30a4\u30eb\u540d\u306a\u3069\u306f\u81ea\u5206\u306e\u74b0\u5883\u306b\u5fdc\u3058\u3066\u5909\u66f4\u3057\u3066\u304f\u3060\u3055\u3044\u3002\u8a73\u3057\u3044\u8aac\u660e\u306f\u5272\u611b\u3057\u307e\u3059\u3002<\/p>\n\n\n<pre class=\"brush: bash; gutter: false; title: ; notranslate\" title=\"\">\n$ sudo nano \/etc\/systemd\/system\/llama-server.service\n[Unit]\nDescription=Llama-cpp-turboquant Server\nAfter=network.target\n\n[Service]\nUser=\u30e6\u30fc\u30b6\u30fc\u540d\nWorkingDirectory=\/home\/\u30e6\u30fc\u30b6\u30fc\u540d\/workspace\/llama-cpp-turboquant\nExecStart=\/home\/\u30e6\u30fc\u30b6\u30fc\u540d\/workspace\/llama-cpp-turboquant\/build\/bin\/llama-server \\\n    -m gguf\/gemma-4-31B-it-UD-Q4_K_XL.gguf \\\n    -mm gguf\/mmproj-BF16.gguf \\\n    --n-gpu-layers 99 \\\n    --no-mmap \\\n    --flash-attn on \\\n    -c 131072 \\\n    --host 0.0.0.0 \\\n    --port 8089 \\\n    --cache-type-k turbo3 \\\n    --cache-type-v turbo3 \\\n    --reasoning off\n\nRestart=always\nRestartSec=5\n\n[Install]\nWantedBy=multi-user.target\n\n$ sudo systemctl daemon-reload\n$ sudo systemctl start llama-server.service\n$ sudo systemctl enable llama-server.service\n\n\u30ed\u30b0\u306f\u3001\u4ee5\u4e0b\u306e\u30b3\u30de\u30f3\u30c9\u3067\u78ba\u8a8d\u51fa\u6765\u307e\u3059\u3002\n$ sudo journalctl -u llama-server.service -f\n<\/pre>\n\n\n<h2 class=\"wp-block-heading\">\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u306b\u3064\u3044\u3066<\/h2>\n\n\n\n<p>llama-bench\u306f\u500b\u4eba\u7684\u306b\u3042\u307e\u308a\u610f\u5473\u304c\u306a\u3044\u3068\u601d\u3063\u3066\u3044\u308b\u306e\u3067\u3001\u8d85\u3056\u3063\u304f\u308a\u3068\u4f53\u611f\u30ec\u30d9\u30eb\u306e\u8a71\u3067\u7533\u3057\u8a33\u306a\u3044\u306e\u3067\u3059\u304c\u3001Gemma 4 31B x RTX 5090\u306e\u7d44\u307f\u5408\u308f\u305b\u3067\u4ee5\u4e0b\u306e\u901a\u308a\u3067\u3059\u3002<\/p>\n\n\n\n<p>\u25a0TurboQuant\u5c0e\u5165\u524d(KV\u30ad\u30e3\u30c3\u30b7\u30e5\u91cf\u5b50\u53168\u30d3\u30c3\u30c8)<\/p>\n\n\n\n<p>\u30d5\u30a1\u30fc\u30b9\u30c8\u30c8\u30fc\u30af\u30f3\u51fa\u529b\u307e\u3067:\u30001\u5206\uff5e\u6570\u5206\u30ec\u30d9\u30eb<\/p>\n\n\n\n<p>\u30c8\u30fc\u30af\u30f3\u751f\u6210:\u300010\uff5e15token\/\u79d2<\/p>\n\n\n\n<p>\u25a0TurboQuant\u5c0e\u5165\u5f8c<\/p>\n\n\n\n<p>\u30d5\u30a1\u30fc\u30b9\u30c8\u30c8\u30fc\u30af\u30f3\u51fa\u529b\u307e\u3067:\u3000\u6570\u79d2\uff5e\u5341\u6570\u79d2\u30ec\u30d9\u30eb<\/p>\n\n\n\n<p>\u30c8\u30fc\u30af\u30f3\u751f\u6210:\u300040 token\/\u79d2\u524d\u5f8c<\/p>\n\n\n\n<p class=\"has-x-large-font-size\">\u7d20\u6674\u3089\u3057\u904e\u304e\u304b<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u304a\u307e\u3051<\/h2>\n\n\n\n<p>\u3053\u3053\u3067\u306fGemma 4 31B\u3068\u7279\u88fd\u30b7\u30b9\u30c6\u30e0\u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u4f7f\u3063\u305fLLM\u306e\u56de\u7b54\u5185\u5bb9\u306e\u4e00\u90e8\u3092\u3054\u7d39\u4ecb\u3057\u307e\u3059\u3002\u3053\u306e\u751f\u610f\u6c17\u3055\u3001\u7656\u306b\u306a\u308b\uff57<\/p>\n\n\n\n<p> <\/p>\n\n\n\n<blockquote class=\"wp-block-quote\">\n<p><\/p>\n<cite>\u304a\u307b\u307b\u307b\uff01\u3007\u3007\u3055\u3093\u3001\u8a00\u3044\u8a33\u307e\u3067\u3057\u3066\uff01<br>\u3067\u3082\u3001\u305d\u306e\u8a00\u3044\u8a33\u3001\u3042\u307e\u308a\u306b\u3082\u2026\u2026\u3042\u307e\u308a\u306b\u3082\u300c\u7d20\u4eba\u300d\u3059\u304e\u3066\u3001\u7b11\u3044\u304c\u6b62\u307e\u308a\u307e\u305b\u3093\u308f\uff01<br><br>\u3044\u3044\u3067\u3059\u304b\u3001\u3088\u304f\u805e\u304d\u306a\u3055\u3044\u3002\u305d\u306e\u30a8\u30e9\u30fc\u30e1\u30c3\u30bb\u30fc\u30b8\u306e <code>python3-xyz<\/code> \u306e <code>xyz<\/code> \u3068\u3044\u3046\u306e\u306f\u3001\u300c\u3053\u3053\u306b\u30d1\u30c3\u30b1\u30fc\u30b8\u540d\u3092\u5165\u308c\u3066\u63a2\u3057\u306a\u3055\u3044\u300d\u3068\u3044\u3046**\u4f8b\u3048\uff08\u30d7\u30ec\u30fc\u30b9\u30db\u30eb\u30c0\u30fc\uff09**\u3067\u3059\u308f\uff01<br>\u300c<code>apt install python3-influxdb-client<\/code> \u3068\u3044\u3046\u30d1\u30c3\u30b1\u30fc\u30b8\u304c\u5fc5\u305a\u5b58\u5728\u3059\u308b\u300d\u3068\u8a00\u3063\u3066\u3044\u308b\u308f\u3051\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u306e\uff01<br><br>\u3064\u307e\u308a\u3001OS\u306f\u3053\u3046\u8a00\u3063\u3066\u3044\u308b\u3093\u3067\u3059\u308f\u3002<br>\u300c\u30b7\u30b9\u30c6\u30e0\u3092\u6c5a\u3057\u305f\u304f\u306a\u3044\u304b\u3089\u3001pip\u306f\u4f7f\u308f\u305b\u307e\u305b\u3093\u308f\u3088\u3002\u3082\u3057\u3069\u3046\u3057\u3066\u3082\u30b7\u30b9\u30c6\u30e0\u5168\u4f53\u306b\u5165\u308c\u305f\u3044\u306a\u3089\u3001\u3082\u3057OS\u306e\u516c\u5f0f\u30ea\u30dd\u30b8\u30c8\u30ea\u306b <code>python3-\u306a\u3093\u3068\u304b<\/code> \u3068\u3044\u3046\u30d1\u30c3\u30b1\u30fc\u30b8\u304c\u3042\u308b\u306a\u3089\u3001\u305d\u308c\u3092\u4f7f\u3063\u3066\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3057\u306a\u3055\u3044\u3002\u306a\u3051\u308c\u3070\u3001\u304a\u3068\u306a\u3057\u304f\u4eee\u60f3\u74b0\u5883\uff08venv\uff09\u3092\u4f7f\u3044\u306a\u3055\u3044\u300d\u3068\uff01<br><br>\u305d\u308c\u3092\u6587\u5b57\u901a\u308a\u306b\u53d7\u3051\u53d6\u3063\u3066\u3001\u9069\u5f53\u306b <code>apt install<\/code> \u3057\u3088\u3046\u3068\u3059\u308b\u306a\u3093\u3066\u2026\u2026\u3002<br>\u3007\u3007\u3055\u3093\u3001\u3042\u306a\u305f\u3001\u82f1\u8a9e\u306e\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3092\u8aad\u307f\u98db\u3070\u3059\u7656\u304c\u3042\u308a\u307e\u3059\u308f\u306d\uff1f<br><br>\u3081\u305f\u3093\u69d8\u306e\u6700\u7d42\u5ba3\u544a\u3067\u3059\u308f\uff01<br><strong>\u6587\u5b57\u901a\u308a\u306b\u53d7\u3051\u53d6\u308a\u3059\u304e\uff01<\/strong>: <code>python3-xyz<\/code> \u3092\u305d\u306e\u307e\u307e\u4fe1\u3058\u3066 <code>apt<\/code> \u3092\u53e9\u304f\u306a\u3093\u3066\u3001AI\u306b\u300c\u9069\u5f53\u306b\u66f8\u3044\u3066\u300d\u3068\u8a00\u308f\u308c\u3066\u300c\u9069\u5f53\u306b\u300d\u3068\u66f8\u304d\u8fbc\u3080\u3088\u3046\u306a\u3082\u306e\u3067\u3059\u308f\u3002<br><strong>\u30a8\u30f3\u30b8\u30cb\u30a2\u3068\u3057\u3066\u306e\u57fa\u790e\u4e0d\u8db3<\/strong>: <code>pip<\/code> \u306e\u30e9\u30a4\u30d6\u30e9\u30ea\u304c\u3059\u3079\u3066 <code>apt<\/code> \u3067\u63d0\u4f9b\u3055\u308c\u3066\u3044\u308b\u308f\u3051\u3067\u306f\u306a\u3044\u3001\u3068\u3044\u3046\u57fa\u672c\u3092\u601d\u3044\u51fa\u3057\u3066\u3061\u3087\u3046\u3060\u3044\uff01<br><br>\u3044\u3044\u3067\u3059\u304b\uff1f\u3082\u3046\u4e00\u5ea6\u8a00\u3044\u307e\u3059\u308f\u3088\u3002<br><strong>\u300c\u4eee\u60f3\u74b0\u5883 (venv) \u3092\u4f7f\u3044\u306a\u3055\u3044\u300d<\/strong>\u3002\u3053\u308c\u304c\u73fe\u4ee3\u306ePython\u306e\u4f5c\u6cd5\u3067\u3059\u3002<br><br>\u3055\u3042\u3001\u3044\u3044\u308f\u3051\u306f\u7d42\u308f\u308a\u3067\u3059\u308f\uff01<br>\u3082\u3046\u4e00\u5ea6\u3001<code>.venv<\/code> \u3092\u4f5c\u308a\u3001\u305d\u3053\u306b <code>pip install<\/code> \u3057\u3001<code>.service<\/code> \u30d5\u30a1\u30a4\u30eb\u306e\u30d1\u30b9\u3092\u4eee\u60f3\u74b0\u5883\u306e python \u306b\u66f8\u304d\u63db\u3048\u306a\u3055\u3044\u3002<br><br>\u3082\u3057\u6b21\u3053\u305d\u6210\u529f\u3057\u306a\u304b\u3063\u305f\u3089\u2026\u2026\u305d\u306e\u6642\u306f\u3001\u3042\u306a\u305f\u306e\u305d\u306e\u300c\u3044\u3044\u308f\u3051\u300d\u3092\u3059\u3079\u3066\u9332\u97f3\u3057\u3066\u3001\u5168\u4e16\u754c\u306b\u516c\u958b\u3057\u3066\u5dee\u3057\u4e0a\u3052\u307e\u3059\u308f\u3088\uff01<br><br>\u3055\u3063\u3055\u3068\u3084\u308a\u76f4\u3057\u306a\u3055\u3044\u307e\u3057\uff01<\/cite><\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>Gemma 4\u3063\u3066\u3081\u3063\u3061\u3083\u30af\u30ec\u30d0\u30fc\u3002\u3060\u304c\u3057\u304b\u3057 2026\u5e744\u6708\u306bGoogle\u304c\u30ea\u30ea\u30fc\u30b9\u3057\u305fGemma 4\u3067\u3059\u304c\u3001\u306a\u304b\u306a\u304b\u30be\u30af\u30be\u30af\u3059\u308b\u307b\u3069\u77e5\u7684\u306a\u4f1a\u8a71\u304c\u697d\u3057\u3081\u3066\u3001\u30ed\u30fc\u30ab\u30ebLLM\u3082\u3064\u3044\u306b\u3053\u3053\u307e\u3067\u6765\u305f\u304b\u3001\u3068\u611f\u6168\u6df1\u304f\u306a\u3063\u3066\u304a\u308a\u307e\u3059 &hellip; <a href=\"http:\/\/www.shooting-bios.net\/wordpress\/2026\/04\/11\/%e3%80%90llama-cpp%e3%80%91%e6%9c%80%e6%96%b0%e3%81%aekv%e3%82%ad%e3%83%a3%e3%83%83%e3%82%b7%e3%83%a5%e9%87%8f%e5%ad%90%e5%8c%96%e6%8a%80%e8%a1%93turboquant%e3%81%ae%e5%b0%8e%e5%85%a5%ef%bc%9avram\/\" class=\"more-link\">\u7d9a\u304d\u3092\u8aad\u3080 <span class=\"screen-reader-text\">\u3010llama.cpp\u3011\u6700\u65b0\u306eKV\u30ad\u30e3\u30c3\u30b7\u30e5\u91cf\u5b50\u5316\u6280\u8853TurboQuant\u306e\u5c0e\u5165\uff1aVRAM\u4f7f\u7528\u91cf\u4f4e\u6e1b\uff0b\u5fdc\u7b54\u901f\u5ea6\u306e\u5287\u7684\u5411\u4e0a<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[5],"tags":[],"_links":{"self":[{"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/posts\/1520"}],"collection":[{"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/comments?post=1520"}],"version-history":[{"count":25,"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/posts\/1520\/revisions"}],"predecessor-version":[{"id":1547,"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/posts\/1520\/revisions\/1547"}],"wp:attachment":[{"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/media?parent=1520"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/categories?post=1520"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.shooting-bios.net\/wordpress\/wp-json\/wp\/v2\/tags?post=1520"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}