{"id":29,"date":"2026-06-11T09:29:53","date_gmt":"2026-06-11T01:29:53","guid":{"rendered":"http:\/\/tantan.ink\/?p=29"},"modified":"2026-06-11T09:29:53","modified_gmt":"2026-06-11T01:29:53","slug":"diffusiongemma","status":"publish","type":"post","link":"https:\/\/www.tantan.ink\/index.php\/2026\/06\/11\/diffusiongemma\/","title":{"rendered":"DiffusionGemma \u53d1\u5e03\uff1aGoogle \u7528\u6269\u6563\u6a21\u578b\u628a\u6587\u672c\u751f\u6210\u52a0\u901f\u4e86 4 \u500d"},"content":{"rendered":"

\u6628\u5929\uff086 \u6708 10 \u65e5\uff09\uff0cGoogle \u5728\u535a\u6587\u4e2d\u6b63\u5f0f\u53d1\u5e03\u4e86 DiffusionGemma<\/strong>\uff0c\u4e00\u4e2a\u57fa\u4e8e\u6587\u672c\u6269\u6563<\/strong>\uff08Text Diffusion\uff09\u6280\u672f\u7684\u5b9e\u9a8c\u6027\u5f00\u6e90\u6a21\u578b\u3002Apache 2.0 \u8bb8\u53ef\uff0c26B MoE\uff08\u5b9e\u9645\u6fc0\u6d3b\u4ec5 3.8B\uff09\uff0c\u4e13\u7528 GPU \u4e0a\u901f\u5ea6\u53ef\u8fbe\u4f20\u7edf\u6a21\u578b\u7684 4 \u500d\u3002<\/p>\n

\u8fd9\u53ef\u80fd\u662f\u4eca\u5e74\u6700\u503c\u5f97\u5173\u6ce8\u7684 “\u6362\u4e2a\u601d\u8def” \u7684\u6a21\u578b\u4e4b\u4e00\u3002<\/p>\n

\u4f20\u7edf LLM \u50cf\u6253\u5b57\u673a\uff0cDiffusionGemma \u50cf\u5370\u5237\u673a<\/h2>\n

\u7528\u8fc7\u672c\u5730\u6a21\u578b\u7684\u4eba\u90fd\u77e5\u9053\u90a3\u79cd\u4f53\u9a8c——7B \u7684\u6a21\u578b\u5728\u5c0f GPU \u4e0a\u7b97\u5f97\u633a\u5feb\uff0c\u4f46\u6bcf\u4e00\u6b21\u63a8\u7406\u90fd\u662f\u4e00\u6b21 “\u4ece\u5de6\u5230\u53f3\u3001\u9010\u4e2a\u4ee4\u724c” \u7684\u4e32\u884c\u64cd\u4f5c\u3002\u4f60\u53ef\u4ee5\u60f3\u8c61\u4e00\u4e2a\u6253\u5b57\u5458\uff0c\u6253\u5b8c\u4e00\u4e2a\u5b57\u624d\u77e5\u9053\u4e0b\u4e00\u4e2a\u5b57\u8be5\u6253\u4ec0\u4e48\u3002<\/p>\n

DiffusionGemma \u5f7b\u5e95\u98a0\u8986\u4e86\u8fd9\u4e2a\u8303\u5f0f\u3002\u5b83\u4e0d\u662f\u81ea\u56de\u5f52<\/strong>\u7684\uff1a<\/p>\n

    \n
  1. \u5148\u51c6\u5907\u4e00\u4e2a 256 token \u7684 “\u7a7a\u767d\u753b\u5e03”\uff08\u968f\u673a\u5360\u4f4d\u7b26\uff09<\/li>\n
  2. \u5728\u591a\u6b21\u8fed\u4ee3\u4e2d\u9010\u6b65\u9501\u5b9a\u6b63\u786e token\uff0c\u7528\u5df2\u9501\u5b9a\u7684\u5185\u5bb9\u505a\u4e0a\u4e0b\u6587\u7ebf\u7d22\u53bb\u63a8\u65ad\u5176\u4f59\u90e8\u5206<\/li>\n
  3. \u6700\u7ec8\u6536\u655b\u4e3a\u8fde\u8d2f\u6587\u672c<\/li>\n<\/ol>\n

    \u8fd9\u8ddf Stable Diffusion \u751f\u6210\u56fe\u7247\u7684\u8fc7\u7a0b\u672c\u8d28\u662f\u4e00\u56de\u4e8b——\u4ece\u566a\u70b9\u4e00\u6b65\u6b65\u6e05\u6670\u3002<\/p>\n

    \"\u81ea\u56de\u5f52
    \u81ea\u56de\u5f52\uff08\u6253\u5b57\u673a\uff09vs \u6587\u672c\u6269\u6563\uff08\u5370\u5237\u673a\uff09——\u4e24\u79cd\u751f\u6210\u65b9\u5f0f\u7684\u672c\u8d28\u533a\u522b\uff08\u624b\u7ed8 SVG\uff09<\/em><\/p>\n

    \u6027\u80fd\u6570\u636e\uff1a\u786e\u5b9e\u5feb<\/h2>\n

    Google \u5b98\u65b9\u7ed9\u51fa\u4e86\u5177\u4f53\u6570\u5b57\uff1a<\/p>\n

      \n
    • NVIDIA H100<\/strong>\uff1a1000+ tokens\/s<\/li>\n
    • RTX 5090<\/strong>\uff1a700+ tokens\/s<\/li>\n
    • \u76f8\u6bd4\u540c\u89c4\u6a21\u81ea\u56de\u5f52\u6a21\u578b\uff0c\u53ef\u8fbe\u5230\u7ea6 4 \u500d\u52a0\u901f<\/strong><\/li>\n<\/ul>\n

      \u56e0\u4e3a\u6fc0\u6d3b\u53c2\u6570\u53ea\u6709 3.8B\uff0c\u91cf\u5316\u540e\u80fd\u63a7\u5236\u5728 18GB VRAM \u4ee5\u5185\uff0cRTX 4090\/5090 \u7ea7\u522b\u5c31\u80fd\u8dd1\u3002<\/p>\n

      \"DiffusionGemma
      DiffusionGemma \u4e0e\u81ea\u56de\u5f52\u6a21\u578b\u5728\u76f8\u540c GPU \u4e0a\u7684 tokens\/s \u5bf9\u6bd4\uff08\u6570\u636e\u6765\u6e90\uff1aGoogle Keyword Blog\uff09<\/em><\/p>\n

      Simon Willison \u5b9e\u6d4b\u7528 NVIDIA NIM API \u8dd1\u4e86\u4e00\u904d\uff0c\u751f\u6210 2409 tokens \u53ea\u7528\u4e86 4.4 \u79d2\uff0c\u76f8\u5f53\u4e8e ~550 tokens\/s\u3002\u4ed6\u751f\u6210\u7684\u662f\u4e00\u6bb5\u5e26\u683c\u5f0f\u7684 SVG \u6e32\u67d3\u8f93\u51fa——\u8fd9\u79cd\u9700\u8981\u8de8 token \u7406\u89e3\u7ed3\u6784\u7684\u4efb\u52a1\u6b63\u662f DiffusionGemma \u7684\u5f3a\u9879\u3002<\/p>\n

      \u4e3a\u4ec0\u4e48\u5b83\u9002\u5408\u672c\u5730\u90e8\u7f72\uff1f<\/h2>\n

      \u6709\u610f\u601d\u7684\u662f\uff0c\u6269\u6563\u6a21\u578b\u7684\u52a0\u901f\u4f18\u52bf\u5728\u672c\u5730\u8fd0\u884c<\/strong>\u65f6\u53cd\u800c\u6700\u660e\u663e\u3002\u539f\u56e0\u5728\u4e8e\u81ea\u56de\u5f52\u6a21\u578b\u8981\u9760\u6279\u5904\u7406\uff08batch\uff09\u6765\u586b\u6ee1 GPU \u7b97\u529b——\u670d\u52a1\u5668\u7aef\u540c\u65f6\u5904\u7406\u4e0a\u5343\u4e2a\u8bf7\u6c42\uff0c\u6548\u7387\u5f88\u9ad8\u3002\u4f46\u672c\u5730\u53ea\u6709\u4f60\u4e00\u4e2a\u4eba\u7528\uff0cGPU \u5927\u90e8\u5206\u65f6\u95f4\u90fd\u5728\u7b49\u4e0b\u4e00\u4e2a token\u3002DiffusionGemma \u4e00\u6b21\u6027\u585e\u7ed9 GPU 256 token \u7684\u5e76\u884c\u8ba1\u7b97\u4efb\u52a1\uff0c\u628a\u786c\u4ef6\u538b\u6ee1\u3002<\/p>\n

      \u4e0d\u8fc7\u6ce8\u610f\u4e00\u4e2a\u5173\u952e\u9650\u5236\uff1aApple Silicon \u4e0a\u52a0\u901f\u4e0d\u660e\u663e<\/strong>\u3002\u56e0\u4e3a M \u7cfb\u5217\u82af\u7247\u662f\u7edf\u4e00\u5185\u5b58\u67b6\u6784\uff0c\u63a8\u7406\u65f6\u53d7\u5185\u5b58\u5e26\u5bbd\u9650\u5236\uff0c\u4e0d\u662f\u8ba1\u7b97\u74f6\u9888——\u6269\u6563\u6a21\u578b\u4f9d\u8d56\u7684\u662f\u9ad8\u8ba1\u7b97\u5bc6\u96c6\u5ea6\uff0c\u5728 Mac \u4e0a\u52a0\u901f\u6548\u679c\u4e0d\u5982\u4e13\u7528 GPU\u3002<\/p>\n

      \u5b83\u80fd\u505a\u4ec0\u4e48\uff1f\u5b9e\u7528\u573a\u666f<\/h2>\n

      DiffusionGemma \u7684\u5b9e\u9a8c\u6027\u8d28\u610f\u5473\u7740\u5b83\u4e0d\u9002\u5408\u66ff\u4ee3 Gemma 4 \u505a\u751f\u4ea7\u7ea7\u6587\u672c\u751f\u6210\u3002\u4f46\u5b83\u5728\u51e0\u4e2a\u7279\u5b9a\u573a\u666f\u4e0a\u6709\u5929\u7136\u4f18\u52bf\uff1a<\/p>\n

        \n
      • \u5185\u8054\u7f16\u8f91 \/ \u4ee3\u7801\u586b\u5145<\/strong>\uff1a\u56e0\u4e3a\u53cc\u5411\u6ce8\u610f\u529b\uff0c\u6bcf\u4e2a token \u80fd\u770b\u5230\u4e0a\u4e0b\u6587\u7684\u4e24\u4fa7\uff0c\u9002\u5408\u586b\u5145\u4ee3\u7801\u4e2d\u95f4\u6bb5<\/li>\n
      • \u975e\u81ea\u56de\u5f52\u7ed3\u6784\u751f\u6210<\/strong>\uff1aUnsloth \u56e2\u961f\u7528\u5fae\u8c03\u540e\u7684 DiffusionGemma \u73a9\u8d77\u4e86\u6570\u72ec——\u81ea\u56de\u5f52\u6a21\u578b\u5929\u751f\u641e\u4e0d\u5b9a\u8fd9\u4e2a\uff0c\u56e0\u4e3a\u540e\u9762\u7684\u6570\u5b57\u4f9d\u8d56\u524d\u9762\u7684\uff0c\u4f46\u4f60\u6ca1\u6cd5\u540c\u65f6\u8003\u8651\u524d\u540e<\/li>\n
      • \u5b9e\u65f6\u4ea4\u4e92\u5f0f\u5e94\u7528<\/strong>\uff1a\u4f4e\u5ef6\u8fdf\u8f93\u51fa\u9002\u5408\u9700\u8981 “\u8fb9\u60f3\u8fb9\u5199” \u7684\u573a\u666f<\/li>\n<\/ul>\n

        \"\u5fae\u8c03\u540e\u7684
        Unsloth \u5fae\u8c03 DiffusionGemma \u89e3\u51b3\u6570\u72ec\u95ee\u9898——\u81ea\u56de\u5f52\u6a21\u578b\u641e\u4e0d\u5b9a\u7684\u4efb\u52a1\uff0c\u6269\u6563\u6a21\u578b\u5929\u7136\u64c5\u957f\uff08\u6765\u6e90\uff1aGoogle Blog\uff09<\/em><\/p>\n

        \u751f\u6001\u652f\u6301\uff1a\u8be5\u6709\u7684\u90fd\u6709<\/h2>\n
          \n
        • \u63a8\u7406\u6846\u67b6<\/strong>\uff1a\u5df2\u652f\u6301 vLLM\uff08Red Hat \u5408\u4f5c\u96c6\u6210\uff09\u3001HuggingFace Transformers\u3001MLX\uff08Apple\uff09<\/li>\n
        • \u5fae\u8c03\u5de5\u5177<\/strong>\uff1aUnsloth\u3001NVIDIA NeMo\u3001Google \u81ea\u5bb6\u7684 Hackable Diffusion\uff08JAX\uff09<\/li>\n
        • \u91cf\u5316<\/strong>\uff1a\u652f\u6301 NVFP4\uff084-bit \u6d6e\u70b9\uff09\uff0c\u7cbe\u5ea6\u635f\u5931\u6781\u5c0f<\/li>\n
        • \u4e91\u7aef<\/strong>\uff1aGoogle Model Garden\u3001NVIDIA NIM \u5747\u53ef\u76f4\u63a5\u8bd5\u7528<\/li>\n
        • \u672a\u6765<\/strong>\uff1allama.cpp \u5b98\u65b9\u652f\u6301\u5373\u5c06\u5230\u6765<\/li>\n<\/ul>\n

          \u9700\u8981\u7559\u610f\u7684<\/h2>\n
            \n
          • \u8f93\u51fa\u8d28\u91cf\u4e0d\u5982 Gemma 4<\/strong>——\u8fd9\u662f\u53d6\u820d\uff0c\u5feb\u5fc5\u7136\u6709\u727a\u7272<\/li>\n
          • \u5b9e\u9a8c\u72b6\u6001<\/strong>——Google \u660e\u786e\u6807\u6ce8\u4e86 experimental\uff0c\u4e0d\u9002\u5408\u505a\u751f\u4ea7\u5173\u952e\u4f9d\u8d56<\/li>\n
          • \u672c\u5730\u786c\u4ef6\u95e8\u69db<\/strong>\uff1a\u867d\u8bf4\u91cf\u5316\u540e 18GB\uff0c\u4f46 RTX 4060 \u7ea7\u522b\u53ef\u80fd\u8dd1\u4e0d\u4e86<\/li>\n
          • \u4e91\u670d\u52a1\u573a\u666f\u4f18\u52bf\u9012\u51cf<\/strong>\uff1a\u9ad8\u5e76\u53d1\u6279\u5904\u7406\u65f6\uff0c\u4f20\u7edf\u81ea\u56de\u5f52\u6a21\u578b\u4e5f\u80fd\u9971\u548c\u7b97\u529b\uff0c\u6269\u6563\u6a21\u578b\u7684\u52a0\u901f\u4f1a\u6253\u6298\u6263<\/li>\n<\/ul>\n

            \u603b\u7ed3<\/h2>\n

            DiffusionGemma \u4e0d\u662f\u53c8\u4e00\u4e2a\u66f4\u5927\u66f4\u5f3a\u7684\u6a21\u578b——\u5b83\u662f\u4e00\u4e2a\u7528\u4e0d\u540c\u601d\u8def\u89e3\u51b3\u95ee\u9898\u7684\u6a21\u578b<\/strong>\u3002\u6587\u672c\u6269\u6563\u8fd9\u6761\u8def Google \u4ece Gemini Diffusion \u5c31\u5f00\u59cb\u6478\u7d22\u4e86\uff0c\u73b0\u5728\u628a\u5b83\u5f00\u6e90\u6210 Gemma \u5bb6\u65cf\u7684\u4e00\u5458\uff0c\u610f\u5473\u7740\u8fd9\u4e2a\u65b9\u5411\u6b63\u5728\u4ece\u5b9e\u9a8c\u8d70\u5411\u53ef\u7528\u3002<\/p>\n

            \u5982\u679c\u4f60\u6709 RTX 5090\/4090\uff0c\u6216\u8005\u60f3\u5728 NVIDIA NIM \u4e0a\u4f53\u9a8c\u4e00\u4e0b\uff0c\u73b0\u5728\u5c31\u53ef\u4ee5\u53bb HuggingFace \u4e0b\u8f7d\u6743\u91cd\u8dd1\u8d77\u6765\u3002llama.cpp \u652f\u6301\u4e5f\u5feb\u5230\u4e86——\u5230\u65f6\u5019\u672c\u5730\u90e8\u7f72\u4f1a\u66f4\u65b9\u4fbf\u3002<\/p>\n

            \n

            \u6765\u6e90: Google Keyword Blog, Simon Willison’s Weblog, HuggingFace \u6a21\u578b\u9875<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"

            Google \u5f00\u6e90 DiffusionGemma\uff1a26B MoE\uff08\u6fc0\u6d3b 3.8B\uff09\u6587\u672c\u6269\u6563\u6a21\u578b\uff0cApache 2.0 \u8bb8\u53ef\u3002H100 1000+ tokens\/s\uff0c\u6bd4\u4f20\u7edf\u6a21\u578b\u5feb 4 \u500d\u3002\u89e3\u8bfb\u6280\u672f\u539f\u7406\u3001\u6027\u80fd\u6570\u636e\u548c\u672c\u5730\u90e8\u7f72\u53ef\u884c\u6027\u3002<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"class_list":["post-29","post","type-post","status-publish","format-standard","hentry"],"_links":{"self":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/29","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/comments?post=29"}],"version-history":[{"count":0,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/29\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media?parent=29"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/categories?post=29"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/tags?post=29"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}