{"id":761,"date":"2026-04-12T20:17:11","date_gmt":"2026-04-12T11:17:11","guid":{"rendered":"https:\/\/blog.gurees.net\/?p=761"},"modified":"2026-04-12T20:44:49","modified_gmt":"2026-04-12T11:44:49","slug":"gemma-4%e3%83%99%e3%83%b3%e3%83%81%e3%83%9e%e3%83%bc%e3%82%af","status":"publish","type":"post","link":"https:\/\/blog.gurees.net\/?p=761","title":{"rendered":"Gemma 4\u3092llama.cpp\u3067\u30d9\u30f3\u30c1\uff08CUDA\u3068Vulkan\u306e\u5dee\u3082\u898b\u3066\u307f\u308b\uff09"},"content":{"rendered":"\n

Gemma 4\u304c\u30ea\u30ea\u30fc\u30b9\u3055\u308c\u3066\u4ee5\u964d\u3001\u30ed\u30fc\u30ab\u30ebLLM\u754c\u9688\u304c\u307e\u305f\u5c11\u3057\u6d3b\u6c17\u3065\u3044\u3066\u304d\u305f\u5370\u8c61\u304c\u3042\u308a\u307e\u3059\u3002
\u7279\u306b\u4eca\u56de\u306e\u30e2\u30c7\u30eb\u306f\u3001\u30b7\u30b9\u30c6\u30e0\u30d7\u30ed\u30f3\u30d7\u30c8\u3078\u306e\u8ffd\u5f93\u6027\u304c\u304b\u306a\u308a\u826f\u304f\u3001\u30ed\u30fc\u30eb\u30d7\u30ec\u30a4\u3084\u6307\u793a\u306e\u4e00\u8cab\u6027\u304c\u5d29\u308c\u306b\u304f\u3044\u70b9\u304c\u76ee\u7acb\u3061\u307e\u3059\u3002<\/p>\n\n\n\n

\u3055\u3089\u306b\u3001Apache License\u3067\u516c\u958b\u3055\u308c\u3066\u3044\u308b\u3053\u3068\u3082\u3042\u308a\u3001\u5546\u7528\u5229\u7528\u3084\u7d44\u307f\u8fbc\u307f\u7528\u9014\u3067\u3082\u6271\u3044\u3084\u3059\u304f\u3001\u300c\u624b\u5143\u3067\u52d5\u304b\u3059\u524d\u63d0\u306e\u30e2\u30c7\u30eb\u300d\u3068\u3057\u3066\u306e\u9b45\u529b\u304c\u4e00\u6bb5\u3068\u5897\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u305d\u3046\u3057\u305f\u6d41\u308c\u3082\u3042\u308a\u3001\u5b9f\u969b\u306e\u6027\u80fd\u304c\u3069\u306e\u7a0b\u5ea6\u51fa\u308b\u306e\u304b\u6c17\u306b\u306a\u3063\u3066\u304d\u305f\u306e\u3067\u3001\u4eca\u56de\u306fllama.cpp\u3092\u4f7f\u3063\u3066\u3044\u304f\u3064\u304b\u306e\u69cb\u6210\u3067\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u3092\u53d6\u3063\u3066\u307f\u307e\u3057\u305f\u3002
\u4eca\u56de\u306e\u7d50\u679c\u3068\u3057\u3066\u306f\u3001\u7279\u306bRTX\u7cfb\u306e\u30d0\u30e9\u30f3\u30b9\u306e\u826f\u3055\u304c\u76ee\u7acb\u3064\u5f62\u306b\u306a\u308a\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

GPU\u6027\u80fd\u5dee\u3068\u91cf\u5b50\u5316\u306e\u5f71\u97ff\u3092\u898b\u308b\u76ee\u7684\u3067\u3001\u4e3b\u306bGPU\u3068\u30e2\u30c7\u30eb\u306e\u7d44\u307f\u5408\u308f\u305b\u3092\u5909\u3048\u3066\u6bd4\u8f03\u3057\u3066\u3044\u307e\u3059\u3002
\u5f15\u6570\u306f\u300c-ngl 99 -fa 1 -p 512 -n 128 -r 5\u300d\u3067\u56fa\u5b9a\u3067\u3059\u3002<\/p>\n\n\n\n

GPU<\/td>\u30d0\u30c3\u30af\u30a8\u30f3\u30c9<\/td>\u30e2\u30c7\u30eb<\/td><\/tr>
P106-100<\/td>CUDA12.2<\/td>gemma-4-E4B-it-Q5_K_M<\/td><\/tr>
RTX3060<\/td>CUDA13.0<\/td>gemma-4-E4B-it-Q5_K_M<\/td><\/tr>
RTX3060<\/td>CUDA13.0<\/td>gemma-4-26B-A4B-it-UD-IQ1_M<\/td><\/tr>
Jetson Xavier NX<\/td>CUDA 11.4<\/td>gemma-4-E4B-it-Q5_K_M<\/td><\/tr>
Jetson Xavier NX<\/td>CUDA 11.4<\/td>gemma-4-E2B-it-Q5_K_M<\/td><\/tr>
2400G<\/td>Vulkan<\/td>gemma-4-E4B-it-Q5_K_M<\/td><\/tr>
2400G<\/td>Vulkan<\/td>gemma-4-E2B-it-Q5_K_M<\/td><\/tr>
2400G<\/td>Vulkan<\/td>gemma-4-26B-A4B-it-UD-IQ1_M<\/td><\/tr>
MI25<\/td>Vulkan<\/td>gemma-4-E4B-it-Q5_K_M<\/td><\/tr>
MI25<\/td>Vulkan<\/td>gemma-4-26B-A4B-it-UD-IQ1_M<\/td><\/tr>
MI25<\/td>Vulkan<\/td>gemma-4-26B-A4B-it-UD-IQ4_NL<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n\n\n\n\n

OS\u306f\u57fa\u672c\u7684\u306b\u306fLinux\u3067\u3059\u304c\u3001Ubuntu\u3068RockyLinux\u304c\u6df7\u5728\u3057\u3066\u308b\u74b0\u5883\u6bce\u306bCPU\u7b49\u3082\u30d0\u30e9\u30d0\u30e9\u306a\u306e\u3067\u3001\u3042\u304f\u307e\u3067\u53c2\u8003\u3067\u3059\u3002<\/p>\n\n\n\n

GPU<\/td>PowerLimit<\/td>BackEnd<\/td>Model<\/td>pp512<\/td>tg128<\/td>Version<\/td><\/tr>
P106-100<\/td>120W<\/td>CUDA12.2<\/td>gemma-4-E4B-it-Q5_K_M<\/td>536.59<\/td>26.33<\/td>ff5ef8278<\/td><\/tr>
P106-100<\/td>90W<\/td>CUDA12.2<\/td>gemma-4-E4B-it-Q5_K_M<\/td>503.32<\/td>25.82<\/td>ff5ef8278<\/td><\/tr>
P106-100<\/td>75W<\/td>CUDA12.2<\/td>gemma-4-E4B-it-Q5_K_M<\/td>463.42<\/td>24.78<\/td>ff5ef8278<\/td><\/tr>
P106-100<\/td>60W<\/td>CUDA12.2<\/td>gemma-4-E4B-it-Q5_K_M<\/td>396.50<\/td>21.89<\/td>ff5ef8278<\/td><\/tr>
RTX3060<\/td>100W<\/td>CUDA13.0<\/td>gemma-4-E4B-it-Q5_K_M<\/td>2303.12<\/td>64.50<\/td>a29e4c0b7<\/td><\/tr>
RTX3060<\/td>100W<\/td>CUDA13.0<\/td>gemma-4-26B-A4B-it-UD-IQ1_M<\/td>1528.09<\/td>81.94<\/td>a29e4c0b7<\/td><\/tr>
Jetson Xavier NX<\/td>15W<\/td>CUDA 11.4<\/td>gemma-4-E4B-it-Q5_K_M<\/td>20.33<\/td>4.03<\/td>5d3a4a7da<\/td><\/tr>
Jetson Xavier NX<\/td>15W<\/td>CUDA 11.4<\/td>gemma-4-E2B-it-Q5_K_M<\/td>206.51<\/td>15.17<\/td>5d3a4a7da<\/td><\/tr>
2400G<\/td>45W<\/td>Vulkan<\/td>gemma-4-E4B-it-Q5_K_M<\/td>41.89<\/td>8.95<\/td>f51fd36d7<\/td><\/tr>
2400G<\/td>45W<\/td>Vulkan<\/td>gemma-4-E2B-it-Q5_K_M<\/td>151.47<\/td>17.91<\/td>f51fd36d7<\/td><\/tr>
2400G<\/td>45W<\/td>Vulkan<\/td>gemma-4-26B-A4B-it-UD-IQ1_M<\/td>44.73<\/td>12.43<\/td>f51fd36d7<\/td><\/tr>
MI25<\/td>100W<\/td>Vulkan<\/td>gemma-4-E4B-it-Q5_K_M<\/td>454.77<\/td>37.44<\/td>85d482e6b<\/td><\/tr>
MI25<\/td>100W<\/td>Vulkan<\/td>gemma-4-26B-A4B-it-UD-IQ1_M<\/td>426.85<\/td>53.07<\/td>85d482e6b<\/td><\/tr>
MI25<\/td>100W<\/td>Vulkan<\/td>gemma-4-26B-A4B-it-UD-IQ4_NL<\/td>411.37<\/td>41.50<\/td>85d482e6b<\/td><\/tr>
MI25<\/td>170W<\/td>Vulkan<\/td>gemma-4-26B-A4B-it-UD-IQ4_NL<\/td>576.82<\/td>52.40<\/td>85d482e6b<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

\u5168\u4f53\u7684\u306b\u898b\u3066\u3001GPU\u306e\u4e16\u4ee3\u5dee\u306f\u305d\u306e\u307e\u307e\u7d20\u76f4\u306b\u6027\u80fd\u5dee\u3068\u3057\u3066\u51fa\u3066\u3044\u308b\u3082\u306e\u306e\u3001\u91cf\u5b50\u5316\u3084\u30e2\u30c7\u30eb\u30b5\u30a4\u30ba\u306e\u5f71\u97ff\u3082\u304b\u306a\u308a\u5927\u304d\u3044\u5370\u8c61\u3067\u3059\u3002
\u7279\u306bIQ1\u7cfb\u306f\u30b5\u30a4\u30ba\u306e\u308f\u308a\u306b\u901f\u5ea6\u304c\u51fa\u3066\u3044\u3066\u3001\u7528\u9014\u6b21\u7b2c\u3067\u306f\u5341\u5206\u5b9f\u7528\u7684\u306b\u6271\u3048\u305d\u3046\u3067\u3059\u3002<\/p>\n\n\n\n

PowerLimit\u306b\u3064\u3044\u3066\u3082\u78ba\u8a8d\u3057\u3066\u307f\u307e\u3057\u305f\u304c\u3001\u6975\u7aef\u306b\u7d5e\u3089\u306a\u3044\u9650\u308a\u306f\u305d\u3053\u307e\u3067\u5927\u304d\u304f\u6027\u80fd\u304c\u843d\u3061\u308b\u308f\u3051\u3067\u306f\u306a\u304f\u3001\u6d88\u8cbb\u96fb\u529b\u3068\u306e\u30d0\u30e9\u30f3\u30b9\u3092\u53d6\u308a\u306a\u304c\u3089\u5e38\u7528\u3059\u308b\u306e\u3082\u73fe\u5b9f\u7684\u305d\u3046\u3067\u3059\u3002<\/p>\n\n\n\n

Jetson\u3084iGPU\u306e\u3088\u3046\u306a\u30ea\u30bd\u30fc\u30b9\u304c\u9650\u3089\u308c\u308b\u74b0\u5883\u3067\u3082\u3001\u5c0f\u3055\u3044\u30e2\u30c7\u30eb\u3092\u9078\u3079\u3070\u305d\u308c\u306a\u308a\u306b\u52d5\u304f\u3053\u3068\u306f\u78ba\u8a8d\u3067\u304d\u305f\u306e\u3067\u3001\u7528\u9014\u3092\u5272\u308a\u5207\u308c\u3070\u9078\u629e\u80a2\u306b\u306f\u306a\u308a\u305d\u3046\u3067\u3059\u3002<\/p>\n\n\n\n

\u4e00\u65b9\u3067AMD\u7cfb\uff08Vulkan\uff09\u306f\u3001tg128\uff08\u751f\u6210\u5074\uff09\u306f\u30b9\u30da\u30c3\u30af\u901a\u308a\u306eTFLOPS\u306b\u8fd1\u3044\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u3067\u305d\u308c\u306a\u308a\u306b\u5065\u95d8\u3057\u3066\u3044\u308b\u3082\u306e\u306e\u3001pp512\uff08\u30d7\u30ed\u30f3\u30d7\u30c8\u51e6\u7406\uff09\u306fCUDA\u306b\u5bfe\u3057\u3066\u304b\u306a\u308a\u5dee\u304c\u51fa\u308b\u7d50\u679c\u306b\u306a\u308a\u307e\u3057\u305f\u3002
\u5b9f\u969b\u306e\u4f53\u611f\u3067\u3082\u3001\u751f\u6210\u4e2d\u3088\u308a\u3082\u6700\u521d\u306e\u5fdc\u7b54\u307e\u3067\u306e\u5f85\u3061\u6642\u9593\u306b\u5dee\u304c\u51fa\u3084\u3059\u304f\u3001\u3053\u306e\u3042\u305f\u308a\u306f\u30d0\u30c3\u30af\u30a8\u30f3\u30c9\u306e\u6700\u9069\u5316\u3084\u5b9f\u88c5\u5dee\u306e\u5f71\u97ff\u304c\u5927\u304d\u305d\u3046\u3067\u3059\u3002CUDA\u3078\u306e\u6700\u9069\u5316\u306e\u5dee\u304c\u3001\u3053\u306e\u3042\u305f\u308a\u306b\u8868\u308c\u3066\u3044\u305d\u3046\u3067\u3059\u3002<\/p>\n\n\n\n

\u305d\u306e\u3046\u3048\u3067\u4eca\u56de\u306e\u7d50\u679c\u3092\u898b\u308b\u9650\u308a\u3001\u624b\u5143\u3067\u30ed\u30fc\u30ab\u30ebLLM\u3092\u8a66\u3059\u7528\u9014\u3067\u3042\u308c\u3070\u3001\u3084\u306f\u308aRTX\u7cfb\u306f\u304b\u306a\u308a\u6271\u3044\u3084\u3059\u3044\u9078\u629e\u80a2\u3060\u3068\u611f\u3058\u307e\u3059\u3002
\u7279\u306bpp512\u306e\u5dee\u306f\u5927\u304d\u304f\u3001E4B\u3067\u306fP106-100\u3084MI25\u3092\u304b\u306a\u308a\u5f15\u304d\u96e2\u3057\u3066\u3044\u3066\u3001\u30d7\u30ed\u30f3\u30d7\u30c8\u6295\u5165\u6642\u306e\u5feb\u9069\u3055\u306f\u982d\u3072\u3068\u3064\u629c\u3051\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u751f\u6210\u901f\u5ea6\u3060\u3051\u3092\u898b\u308b\u3068\u4ed6\u306b\u3082\u5065\u95d8\u3057\u3066\u3044\u308b\u69cb\u6210\u306f\u3042\u308a\u307e\u3059\u304c\u3001\u5b9f\u969b\u306b\u89e6\u3063\u3066\u3044\u3066\u52b9\u3044\u3066\u304f\u308b\u306e\u306f\u6700\u521d\u306e\u5fdc\u7b54\u307e\u3067\u306e\u8efd\u3055\u3060\u3063\u305f\u308a\u3059\u308b\u306e\u3067\u3001\u305d\u306e\u610f\u5473\u3067\u3082RTX3060\u30af\u30e9\u30b9\u306f\u30d0\u30e9\u30f3\u30b9\u304c\u826f\u3055\u305d\u3046\u3067\u3059\u3002
\u6d88\u8cbb\u96fb\u529b\u3092100W\u306b\u6291\u3048\u3066\u3082\u5341\u5206\u901f\u304f\u3001\u5e38\u7528\u74b0\u5883\u3068\u3057\u3066\u3082\u304b\u306a\u308a\u73fe\u5b9f\u7684\u306a\u30e9\u30a4\u30f3\u306b\u53ce\u307e\u3063\u3066\u3044\u307e\u3059\u3002

Radeon Instinct MI25\u306f\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u81ea\u4f53\u306e\u30dd\u30c6\u30f3\u30b7\u30e3\u30eb\u306f\u611f\u3058\u308b\u3082\u306e\u306e\u3001\u3053\u3046\u3057\u305f\u30b9\u30d1\u30b3\u30f3\u7528\u9014\u306eGPU\u3092\u30ed\u30fc\u30ab\u30eb\u3067\u632f\u308a\u56de\u3059\u306e\u306f\u306a\u304b\u306a\u304b\u697d\u3057\u3044\u3068\u3053\u308d\u3067\u3059\u3002\u305f\u3060\u3001\u5b9f\u969b\u306e\u6027\u80fd\u9762\u3067\u306f\u3084\u3084\u30ed\u30de\u30f3\u67a0\u306b\u7559\u307e\u3063\u3066\u3057\u307e\u3063\u3066\u3044\u308b\u5370\u8c61\u3082\u3042\u308a\u307e\u3059\u3002
\u307e\u305f\u3001\u4ee5\u4e0b\u8a18\u4e8b\u306e\u3088\u3046\u306b\u51b7\u5374\u306b\u306f\u975e\u5e38\u306b\u82e6\u52b4\u3059\u308b\u306e\u3067\u3001\u5e38\u7528\u74b0\u5883\u3068\u3057\u3066\u4f7f\u3046\u306b\u306f\u6c11\u751f\u306eGPU\u3092\u304a\u3059\u3059\u3081\u3057\u307e\u3059\u3002
https:\/\/blog.gurees.net\/?p=708<\/a><\/p>\n\n\n\n

\u3068\u306f\u3044\u3048\u3001\u30d9\u30f3\u30c1\u306e\u6570\u5024\u3060\u3051\u3067\u306f\u898b\u3048\u306a\u3044\u90e8\u5206\u3082\u591a\u3044\u306e\u3067\u3001\u5b9f\u969b\u306e\u30e6\u30fc\u30b9\u30b1\u30fc\u30b9\u3067\u3069\u3053\u307e\u3067\u4f7f\u3048\u308b\u304b\u306f\u5f15\u304d\u7d9a\u304d\u78ba\u8a8d\u3057\u3066\u3044\u304d\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

Gemma 4\u304c\u30ea\u30ea\u30fc\u30b9\u3055\u308c\u3066\u4ee5\u964d\u3001\u30ed\u30fc\u30ab\u30ebLLM\u754c\u9688\u304c\u307e\u305f\u5c11\u3057\u6d3b\u6c17\u3065\u3044\u3066\u304d\u305f\u5370\u8c61\u304c\u3042\u308a\u307e\u3059\u3002\u7279\u306b\u4eca\u56de\u306e\u30e2\u30c7\u30eb\u306f\u3001\u30b7\u30b9\u30c6\u30e0\u30d7\u30ed\u30f3\u30d7\u30c8\u3078\u306e\u8ffd\u5f93\u6027\u304c\u304b\u306a\u308a\u826f\u304f\u3001\u30ed\u30fc\u30eb\u30d7\u30ec\u30a4\u3084\u6307\u793a\u306e\u4e00\u8cab\u6027\u304c\u5d29\u308c\u306b\u304f\u3044\u70b9\u304c\u76ee\u7acb\u3061\u307e\u3059\u3002 \u3055\u3089\u306b\u3001 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[19],"tags":[],"class_list":["post-761","post","type-post","status-publish","format-standard","hentry","category-linux"],"_links":{"self":[{"href":"https:\/\/blog.gurees.net\/index.php?rest_route=\/wp\/v2\/posts\/761","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.gurees.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.gurees.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.gurees.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.gurees.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=761"}],"version-history":[{"count":13,"href":"https:\/\/blog.gurees.net\/index.php?rest_route=\/wp\/v2\/posts\/761\/revisions"}],"predecessor-version":[{"id":766,"href":"https:\/\/blog.gurees.net\/index.php?rest_route=\/wp\/v2\/posts\/761\/revisions\/766"}],"wp:attachment":[{"href":"https:\/\/blog.gurees.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=761"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.gurees.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=761"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.gurees.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=761"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}