{
  "version": "https://jsonfeed.org/version/1", 
  "title": "Local LLM", 
  "description": "\u8fd9\u4e2a\u8282\u70b9\u8ba8\u8bba\u5728\u672c\u5730\u7535\u8111\u6216\u8005\u5c40\u57df\u7f51\u91cc\u8fd0\u884c LLM\uff08\u5927\u8bed\u8a00\u6a21\u578b\uff09\u7684\u6280\u672f\u7ec6\u8282", 
  "home_page_url": "https://www.v2ex.com/go/localllm", 
  "feed_url": "https://www.v2ex.com/feed/localllm.json", 
  "icon": "https://cdn.v2ex.com/navatar/c8ed/21db/722_large.png?m=1751736797", 
  "favicon": "https://cdn.v2ex.com/navatar/c8ed/21db/722_normal.png?m=1751736797", 
  "items": [
    {
      "author": {
        "url": "https://www.v2ex.com/member/invdan", 
        "name": "invdan", 
        "avatar": "https://cdn.v2ex.com/avatar/5251/8a4a/310818_large.png?m=1782096664"
      }, 
      "url": "https://www.v2ex.com/t/1221902", 
      "title": "\u5f00\u6e90\u4e86\u4e00\u4e2a LLM \u63a8\u7406\u670d\u52a1\u76d1\u63a7\u9762\u677f", 
      "id": "https://www.v2ex.com/t/1221902", 
      "date_published": "2026-06-22T02:51:37+00:00", 
      "content_html": "<p>\u5f00\u6e90\u4e86\u4e00\u4e2a LLM \u63a8\u7406\u670d\u52a1\u76d1\u63a7\u9762\u677f\n<strong>\u9879\u76ee\u5730\u5740</strong>\uff1a <a href=\"https://github.com/coolwolfqs/llm-inference-monitor\" rel=\"nofollow\">https://github.com/coolwolfqs/llm-inference-monitor</a></p>\n<hr/>\n<h4>\u4e3a\u4ec0\u4e48\u505a\u8fd9\u4e2a</h4>\n<p>\u6700\u8fd1\u5728\u7528 llama.cpp \u8dd1\u63a8\u7406\u670d\u52a1\uff0c\u4e00\u76f4\u7f3a\u4e00\u4e2a\u597d\u7528\u7684\u76d1\u63a7\u9762\u677f\u3002</p>\n<p>\u7f51\u4e0a\u65b9\u6848\u65e0\u975e\u4e24\u6761\u8def\uff1a</p>\n<ol>\n<li><strong>Prometheus + Grafana</strong> \u2192 \u592a\u91cd\u4e86\uff0c\u4e3a\u4e86\u770b\u4e2a GPU \u6e29\u5ea6\u642d\u4e00\u5957\u76d1\u63a7\u4f53\u7cfb</li>\n<li><strong>nvidia-smi \u5237\u5c4f</strong> \u2192 \u539f\u59cb\uff0c\u4f46\u5c31\u770b\u4e2a GPU \uff0cCPU/\u5185\u5b58/\u63a8\u7406\u6307\u6807\u5168\u6ca1\u6709</li>\n</ol>\n<h2>\u4e8e\u662f\u81ea\u5df1\u6413\u4e86\u4e00\u4e2a\u9762\u677f\uff0c\u73b0\u5728\u6574\u7406\u6210\u5f00\u6e90\u9879\u76ee\u653e\u51fa\u6765\u4e86\u3002\u4e0d\u4f1a\u7f16\u7a0b\uff0c\u5168\u7a0b\u5c31\u7531 hermes \u5f85\u5f00\u53d1\uff0c\u4e0d\u6210\u719f\u4e4b\u5904\u5404\u4f4d\u770b\u5b98\u591a\u5305\u6db5\u3002</h2>\n<h4>\u957f\u4ec0\u4e48\u6837</h4>\n<p>\u4e00\u4e2a\u9875\u9762\u641e\u5b9a\u6240\u6709\u76d1\u63a7\u4fe1\u606f\uff0c\u5206\u6210\u51e0\u4e2a\u533a\u57df\uff1a</p>\n<p><strong>\u670d\u52a1\u6982\u89c8\u533a</strong></p>\n<ul>\n<li>\u5f53\u524d\u8fd0\u884c\u7684\u6a21\u578b\u3001\u4e0a\u4e0b\u6587\u957f\u5ea6\u3001\u91cf\u5316\u7cbe\u5ea6</li>\n<li>\u5f15\u64ce\u7248\u672c\u53f7\uff08 llama.cpp / vllm \uff09</li>\n<li>\u5065\u5eb7\u8bc4\u5206\uff08\u786c\u4ef6\u5206 + \u7cfb\u7edf\u5206 + \u63a8\u7406\u5206\uff09</li>\n</ul>\n<p><strong>GPU \u533a</strong></p>\n<ul>\n<li>\u5229\u7528\u7387 / \u663e\u5b58 / \u6e29\u5ea6 / \u529f\u8017 \u5b9e\u65f6\u66f2\u7ebf\u56fe</li>\n<li>\u6bcf\u5f20\u5361\u7684\u8be6\u7ec6\u4fe1\u606f\uff08\u9891\u7387\u3001PCIe \u94fe\u8def\u3001\u7f16\u7801\u5668\u8d1f\u8f7d\uff09</li>\n<li><strong>\u5e26 GPU \u8fdb\u7a0b\u5217\u8868</strong>\uff08\u770b\u4e00\u773c\u5c31\u77e5\u9053\u8c01\u5728\u5403\u663e\u5b58\uff09</li>\n</ul>\n<p><strong>\u7cfb\u7edf\u533a</strong></p>\n<ul>\n<li>CPU \u6bcf\u6838\u5229\u7528\u7387\u70ed\u529b\u56fe</li>\n<li>\u5185\u5b58 / Swap / \u7f13\u5b58</li>\n<li>\u78c1\u76d8\u8bfb\u5199\u901f\u5ea6 + \u5206\u533a\u4f7f\u7528\u7387</li>\n<li>\u7f51\u7edc\u5b9e\u65f6\u541e\u5410\u91cf</li>\n</ul>\n<p><strong>\u63a8\u7406\u533a</strong></p>\n<ul>\n<li>TPS \u5b9e\u65f6\u5fc3\u7535\u56fe</li>\n<li>KV Cache \u5360\u7528 + \u5269\u4f59\u53ef\u7528 Token \u4f30\u7b97</li>\n<li>TTFT / TPOT / KV \u547d\u4e2d\u7387 / MTP \u6295\u673a\u89e3\u7801\u52a0\u901f\u6bd4</li>\n<li>IP \u7ea7 Token \u6d88\u8017\u7edf\u8ba1</li>\n</ul>\n<hr/>\n<h4>\u6280\u672f\u6808</h4>\n<pre><code>\u540e\u7aef\uff1aPython FastAPI + psutil + nvidia-smi\n\u524d\u7aef\uff1a\u7eaf HTML + CSS + JS \uff08\u65e0\u6846\u67b6\uff0c\u65e0\u9700\u6784\u5efa\uff09\n\u56fe\u8868\uff1aCanvas \u539f\u751f\u7ed8\u5236\uff08\u8d1d\u585e\u5c14\u66f2\u7ebf\uff0c\u9632\u6296\u91cd\u7ed8\uff09\n\u5b9e\u65f6\uff1aSSE \u63a8\u9001\uff08 2 \u79d2\u95f4\u9694\uff09 + HTTP \u8f6e\u8be2\uff08 30 \u79d2\u515c\u5e95\uff09\n\u90e8\u7f72\uff1apip install -r requirements.txt \u5c31\u884c\n</code></pre>\n<p>\u6574\u4e2a\u9879\u76ee 30 \u591a\u4e2a\u6587\u4ef6\uff0c\u524d\u7aef\u96f6\u4f9d\u8d56\uff0c\u540e\u7aef\u53ea\u4f9d\u8d56 FastAPI \u3001psutil \u3001aiohttp \u4e09\u4e2a\u5e93\u3002</p>\n<hr/>\n<h4>\u5feb\u901f\u4f53\u9a8c</h4>\n<pre><code class=\"language-bash\">git clone GitHub - coolwolfqs/llm-inference-monitor: Real-time monitoring dashboard for LLM inference services\ncd llm-inference-monitor\npip install -r requirements.txt\npython -m backend.server\n</code></pre>\n<p>\u6253\u5f00 http://localhost:8081 \u5c31\u80fd\u770b\u5230\u9762\u677f\u4e86\u3002</p>\n<p>\u5982\u679c\u9700\u8981\u91c7\u96c6\u63a8\u7406\u6307\u6807\uff0c\u65c1\u8fb9\u8dd1\u4e00\u4e2a llama.cpp server \uff08\u9ed8\u8ba4 8080 \u7aef\u53e3\uff09\u5c31\u884c\uff0c\u81ea\u52a8\u5bf9\u63a5\u3002</p>\n<hr/>\n<h4>\u9879\u76ee\u5730\u5740</h4>\n<p><strong><a href=\"https://github.com/coolwolfqs/llm-inference-monitor\" rel=\"nofollow\">https://github.com/coolwolfqs/llm-inference-monitor</a></strong></p>\n<p>\u6b22\u8fce Star \u3001Fork \u3001PR \uff0c\u89c9\u5f97\u6709\u7528\u7684\u8bdd\u4e5f\u6b22\u8fce\u8f6c\u53d1\u3002</p>\n<hr/>\n<p><strong>\u8865\u5145\u8bf4\u660e</strong>\uff1a\u9879\u76ee\u4ece\u751f\u4ea7\u73af\u5883\u7684\u5185\u90e8\u9762\u677f\u6574\u7406\u800c\u6765\uff0c\u6838\u5fc3\u903b\u8f91\u548c UI \u5e03\u5c40\u90fd\u4fdd\u7559\u4e86\u539f\u6837\uff0c\u53ea\u662f\u628a\u540e\u7aef\u4ece\u5355\u4f53\u6539\u6210\u4e86\u6a21\u5757\u5316\u91c7\u96c6\u5668\u67b6\u6784\uff0c\u65b9\u4fbf\u5927\u5bb6\u6309\u9700\u589e\u5220\u76d1\u63a7\u6307\u6807\u3002\u4e2d\u82f1\u6587\u53cc\u8bed\u6587\u6863\u90fd\u6709\u3002</p>\n<p>\u6709\u4ec0\u4e48\u95ee\u9898\u6216\u8005\u5efa\u8bae\u53ef\u4ee5\u76f4\u63a5\u56de\u5e16\uff0c\u4e5f\u53ef\u4ee5 GitHub \u63d0 Issue \u3002</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/jiezou", 
        "name": "jiezou", 
        "avatar": "https://cdn.v2ex.com/avatar/4cea/d8d6/720683_large.png?m=1750819087"
      }, 
      "url": "https://www.v2ex.com/t/1221894", 
      "title": "\u5927\u6a21\u578b\u5c0f\u767d\u63a8\u8350\u4e00\u4e0b\u672c\u5730\u6a21\u578b", 
      "id": "https://www.v2ex.com/t/1221894", 
      "date_published": "2026-06-22T02:31:39+00:00", 
      "content_html": "<a target=\"_blank\" href=\"https://i.imgur.com/lsS0QHT.jpeg\" rel=\"nofollow noopener\" target=\"_blank\"><img src=\"https://i.imgur.com/lsS0QHT.jpeg\" class=\"embedded_image\" rel=\"noreferrer\"></a><br />\u6709\u53f0\u95f2\u7f6e\u7684\u5c0f\u4e3b\u673a\uff0c\u6362\u4e2a 2080ti \u9b54\u6539\u663e\u5361\uff0c\u53ef\u80fd\u504f\u5411\u4e8e\u77e5\u8bc6\u5e93\u7684\u7528\u9014,\u80fd\u8dd1\u54ea\u4e9b\u672c\u5730\u6a21\u578b\u5462\uff1f"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/hihihihihi", 
        "name": "hihihihihi", 
        "avatar": "https://cdn.v2ex.com/avatar/72dd/4352/178974_large.png?m=1769391179"
      }, 
      "url": "https://www.v2ex.com/t/1221519", 
      "date_modified": "2026-06-19T08:15:22+00:00", 
      "content_html": "<p>\u524d\u51e0\u5468\u7528\u4e86\u51e0\u5929 Claude-fable-5 \u6a21\u578b\uff0c\u786e\u5b9e\u5f88\u660e\u663e\u7684\u611f\u89c9\u51c6\u786e\u7387\u975e\u5e38\u9ad8\uff0c\u7406\u89e3\u80fd\u529b\u4e5f\u975e\u5e38\u9ad8\uff0c\u57fa\u672c\u4e00\u904d\u8fc7\u3002\u5728\u8fd9\u4e2a\u4e4b\u524d\u6211\u5927\u90e8\u5206\u4f7f\u7528 Opus4.8 \u4ee5\u53ca Sonnet4.6, \u7406\u8bba\u4e0a\u6211\u7528 sonnet 4.6 \u66f4\u591a\u3002</p>\n<p>\u8fd9\u6bb5\u65f6\u95f4\u770b GLM \u8fd9\u4e48\u706b\uff0c\u6211\u4e5f\u51d1\u70ed\u95f9\u53bb\u62a2\u4e86\u4e0b\uff0c\u4e00\u76f4\u6ca1\u62a2\u5230\uff0c\u5e72\u8106\u8d2d\u4e70\u4e86\u56fd\u5916\u7248\u672c\u7684 <a href=\"http://z.ai\" rel=\"nofollow\">z.ai</a> \u7684\u5957\u9910\u3002\u8fd9\u51e0\u5929\u91cd\u5ea6\u4f7f\u7528\u4e86\u4e00\u4e0b\uff0c\u8bf4\u8bf4\u6211\u7684\u611f\u53d7\uff1a</p>\n<p>\u524d\u63d0\uff1a\u6211\u90fd\u662f\u7528\u7684 claude code cli</p>\n<ol>\n<li>\n<p>GLM5.2 \u6709\u70b9\u8bdd\u75e8\uff0c\u6709\u7684\u65f6\u5019\u8bf4\u4e00\u5806\u8bdd\uff0c\u751a\u81f3\u5927\u6bb5\u91cd\u590d\uff0c\u4e0d\u591f\u8a00\u7b80\u610f\u8d45</p>\n</li>\n<li>\n<ol>\n<li>GLM5.2 \u5bf9\u6574\u4e2a\u9879\u76ee\u7684\u628a\u63a7\u4e0d\u5982 claude code \uff0c\u8981 GLM \u505a\u4e00\u4e2a\u529f\u80fd\uff0c\u4ed6\u6709\u65f6\u4e0d\u5148\u53bb\u770b\u7a0b\u5e8f\u662f\u4e0d\u662f\u6709\u4ec0\u4e48\u5df2\u7ecf\u505a\u4e86\u7684\uff0c\u6216\u8005\u662f\u5426\u5f71\u54cd\u522b\u7684\u5730\u65b9\uff0c\u5c31\u662f\u611f\u89c9\u6574\u4e2a\u8003\u8651\u4e0d\u5468\u5230\uff0c\u8981\u6211\u6765\u6307\u6b63\u3002 \u6211\u660e\u767d\u5f88\u591a\u65f6\u5019\u9700\u8981\u63cf\u8ff0\u66f4\u6e05\u6670\u9700\u6c42\uff0c\u4f46\u662f\u6709\u7684\u9700\u6c42\u5e94\u8be5\u662f\u663e\u800c\u6613\u89c1\u7684\u3002 \u8fd9\u70b9\u6211\u611f\u89c9\u5168\u5c40\u4e0a\uff0cfable &gt; opus &gt; sonnet &gt; glm</li>\n</ol>\n</li>\n<li>\n<p>\u6162\uff0c\u5361\uff0c\u7ecf\u5e38\u4e00\u4e2a\u5c0f\u95ee\u9898\uff0c\u8981\u641e\u597d\u51e0\u5206\u949f\uff0c\u660e\u663e\u63d0\u8d76\u4e0a claude \u6548\u7387\u8981\u66f4\u9ad8\u3002</p>\n</li>\n</ol>\n<p>\u6240\u4ee5\u603b\u7ed3\u8d77\u6765\u8bf4\uff1a\u76f8\u5bf9\u76ee\u524d\u6240\u8c13\u53ef\u7528\u6a21\u578b\u7b2c\u4e00\u6765\u8bf4\uff0c\u6211\u89c9\u5f97 GLM5.2 \u8fc7\u8a89\u4e86\uff0c\u7406\u89e3\u80fd\u529b\u6b20\u7f3a\uff0c\u6574\u4f53\u628a\u63a7\u80fd\u529b\u4e0d\u591f\uff0c\u6548\u7387\u4e0d\u591f\u9ad8\u3002\u603b\u7684\u6765\u8bf4\u4e5f\u662f\u56fd\u4ea7\u6a21\u578b\u91cc\u9762\u4e00\u68af\u961f\u7684\uff0c\u4f46\u662f\u548c claude \u786e\u5b9e\u8fd8\u6709\u534a\u5e74\u5230\u4e00\u5e74\u7684\u5dee\u8ddd\u3002</p>\n<p>PS\uff1a\u4e2a\u4eba\u610f\u89c1\uff0c\u5f88\u4e3b\u89c2\uff0c\u4ec5\u4f9b\u53c2\u8003\u3002</p>\n", 
      "date_published": "2026-06-19T05:42:06+00:00", 
      "title": "GLM5.2 \u4e2a\u4eba\u611f\u89c9\u6709\u70b9\u88ab\u5439\u5927\u4e86", 
      "id": "https://www.v2ex.com/t/1221519"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/frankyzf", 
        "name": "frankyzf", 
        "avatar": "https://cdn.v2ex.com/avatar/5a39/5799/45580_large.png?m=1682000562"
      }, 
      "url": "https://www.v2ex.com/t/1221496", 
      "title": "\u6709\u652f\u6301 6000 Ada \u4f7f\u7528 deepseek v4 flash \u63a8\u7406 \u7684\u6846\u67b6\u5417", 
      "id": "https://www.v2ex.com/t/1221496", 
      "date_published": "2026-06-19T02:51:50+00:00", 
      "content_html": "<p>\u663e\u5b58\u662f\u591f\u7684\uff08\u591a\u5361\uff09\uff0c\u4f46\u67b6\u6784\u6709\u9650\u5236\uff08 SM_89 \uff09</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/mountainl", 
        "name": "mountainl", 
        "avatar": "https://cdn.v2ex.com/gravatar/00bd2c459f1285d4caca2033c21eab77?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1220978", 
      "title": "\u5206\u4eab\u4e2a\u81ea\u5df1\u5728\u7528\u7684\u73a9\u5177", 
      "id": "https://www.v2ex.com/t/1220978", 
      "date_published": "2026-06-17T01:54:55+00:00", 
      "content_html": "\u524d\u6bb5\u65f6\u95f4 qwen3.5 \u53d1\u5e03\u7684\u65f6\u5019\uff0c\u8bd5\u7740\u7528 4070 \u8dd1 9b \u7248\u672c\uff0c\u53d1\u73b0\u914d\u5408 openclaw \u73a9\u90fd\u73a9\u4e0d\u8d77\u6765\uff0c\u800c\u4e14\u4e0a\u4e0b\u6587\u53ea\u80fd\u5f00\u5230\u5927\u6982 32k \u5de6\u53f3\u3002\u6b63\u597d\u8fd9\u6bb5\u65f6\u95f4\u6ca1\u4ec0\u4e48\u597d\u6298\u817e\u7684\u4e86\uff08 NAS \u548c\u8f6f\u8def\u7531\u5df2\u7ecf\u7a33\u5b9a\u8fd0\u884c\u4e2d\uff09\uff0c\u6240\u4ee5\u4e70\u4e86\u4e24\u5757 3060 12g \u548c x99 \u7684\u5927\u677f\u548c E5 3673V3 \uff0c\u53e6\u5916\u914d\u4e86\u4e2a 1200w \u7684\u7535\u6e90\uff0c\u5185\u5b58\u7528\u4e3b\u529b\u673a\u62c6\u4e0b\u6765\u7684 16x2 \uff08\u4e3b\u529b\u6210\u4e8c\u5976\u4e86\uff09\u3002<br />\u6b63\u5de7\u8d76\u4e0a qwen3.6 \u53d1\u5e03\uff0c\u8bd5\u7740\u8dd1\u4e86 27b \u548c 35b \u6a21\u578b\uff0c\u6700\u7ec8\u4f7f\u7528 mudler/Qwen3.6-35B-A3B-APEX-GGUF \u6a21\u578b\uff0c\u5f00 128k \u4e0a\u4e0b\u6587\uff0c\u8f93\u5165 2000tps \uff0c\u8f93\u51fa\u5728 100tps \uff0c\u5f53\u7136\u4e0a\u4e0b\u6587\u8fbe\u5230\u4e00\u5b9a\u7a0b\u5ea6\u5c31\u5f00\u59cb\u80e1\u626f\u964d\u901f\u4e86\u3002<br />\u73b0\u5728\u914d\u5408 hermes agent \uff0c\u611f\u89c9\u53ef\u73a9\u6027\u633a\u9ad8\u7684\uff0c\u4f5c\u4e3a\u4ee3\u7801\u5c0f\u767d\uff0c\u53ef\u4ee5\u5e2e\u6211\u5199\u4e00\u4e9b\u5c0f\u7684\u811a\u672c<br />\u6298\u817e\u5b8c\u7d22\u7136\u65e0\u5473\u8fd8\u80fd\u51fa\u6389\u56de\u70b9\u8840\uff0c\u76f8\u5f53\u4e8e\u82b1\u4e2a\u5343\u628a\u5757\u94b1\u8ba9\u81ea\u5df1\u53c8\u723d\u73a9\u4e86\u4e00\u6bb5\u65f6\u95f4\u3002"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/davidyin", 
        "name": "davidyin", 
        "avatar": "https://cdn.v2ex.com/avatar/6dcb/35fd/43242_large.png?m=1768429019"
      }, 
      "url": "https://www.v2ex.com/t/1220434", 
      "title": "\u914d\u7f6e kiro \u7684\u95ee\u9898", 
      "id": "https://www.v2ex.com/t/1220434", 
      "date_published": "2026-06-15T01:39:27+00:00", 
      "content_html": "\u63a5\u4e0a\u56de <a target=\"_blank\" href=\"https://v2ex.com/t/1211566#reply82\" rel=\"nofollow noopener\">https://v2ex.com/t/1211566#reply82</a><br /><br />\u5728 PC A \u4e0a\uff0c\u88c5\u4e86 Ubuntu 24.04,\u7528\u7684\u662f rx6800xt 16G \u663e\u5361\uff0collama \u88c5\u4e0a\u4e86\uff0copen-webui \u4e5f\u88c5\u4e0a\u4e86\u3002<br /><br />\u4ece PC B \uff0c\u6211\u7684\u684c\u9762\u7535\u8111 Windows \u4e0a\u8bbf\u95ee\u5462\uff0c\u53ef\u4ee5\u5728\u6d4f\u89c8\u5668\u91cc\u9762\u6253\u5f00 open-webui \uff0c\u8bbf\u95ee\u90fd\u6b63\u5e38\u3002<br /><br />\u73b0\u5728\u56f0\u6270\u6211\u7684\u662f\u5982\u4f55\u914d\u7f6e kiro \uff0c\u4f7f\u5176\u80fd\u4f7f\u7528 PC A \u4e0a\u7684 ollama \uff08 qwen2.5-coder:7b \uff09\uff0c\u4f5c\u4e3a agent \uff0c\u8f85\u52a9\u7f16\u7a0b\u3002\u73b0\u5728\u8fd9\u91cc\u603b\u662f\u4e0d\u884c\u3002<br /><br />\u5411\u505a\u6210\u529f\u7684\u670b\u53cb\u8bf7\u6559"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/sjmcefc2", 
        "name": "sjmcefc2", 
        "avatar": "https://cdn.v2ex.com/gravatar/ed45fe578f1092dcabc2eaaf904f7374?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1219998", 
      "date_modified": "2026-06-13T00:47:42+00:00", 
      "content_html": "<p>macbook pro \u8dd1\u672c\u5730\u6a21\u578b\uff0c64g \u5185\u5b58\u591f\u7528\u5417\uff1f\n64g \u53ef\u4ee5\u8dd1\u54ea\u4e9b\u6a21\u578b\u5462\uff1f\n\u7b49 9 \u6708\u65b0\u54c1\u8fd8\u662f\u73b0\u5728\u5462\uff1f\n\u4e0d\u61c2 mac \u7684\u5546\u54c1\u554a</p>\n", 
      "date_published": "2026-06-12T09:45:55+00:00", 
      "title": "\u4e70 macbook pro \u7b14\u8bb0\u672c\uff0c\u8dd1\u672c\u5730\u6a21\u578b\uff0c\u600e\u4e48\u914d\u7f6e\u6027\u4ef7\u6bd4\u6bd4\u8f83\u9ad8\uff1f", 
      "id": "https://www.v2ex.com/t/1219998"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/sentinelK", 
        "name": "sentinelK", 
        "avatar": "https://cdn.v2ex.com/avatar/8d13/4c44/631792_large.png?m=1781227585"
      }, 
      "url": "https://www.v2ex.com/t/1219800", 
      "title": "lama.cpp \u76ee\u524d\u6709\u91cd\u5927\u6027\u80fd bug\uff1a checkpoint \u7684\u5de1\u56de\u903b\u8f91\u5bf9\u4e8e\u6df7\u5408\u6a21\u578b\uff08\u6bd4\u5982 qwen3.6-27B\uff09\u65e0\u6548\uff0c\u4ece\u800c\u5bfc\u81f4\u5927\u6982\u7387\u6bcf\u6b21\u5bf9\u8bdd\u90fd\u8981 prefill \u5168\u6587\uff0c\u4e25\u91cd\u62d6\u6162\u901f\u5ea6", 
      "id": "https://www.v2ex.com/t/1219800", 
      "date_published": "2026-06-12T01:27:27+00:00", 
      "content_html": "<p>\u5728\u6628\u5929\u7814\u7a76 qwen3.6-27B \u7684\u4f18\u5316\u65f6\uff0c\u770b\u5230\u4e86\u8fd9\u4e2a\u95ee\u9898\uff1a<a href=\"https://github.com/ggml-org/llama.cpp/issues/22384\" rel=\"nofollow\">server: fix context checkpoint restore for hybrid/recurrent models (DeltaNet/Mamba)</a></p>\n<p>\u5927\u6982\u610f\u601d\u5c31\u662f\uff0c\u56e0\u4e3a llama.cpp \u7684\u7f13\u5b58\u5de1\u56de\u903b\u8f91\u6709\u95ee\u9898\uff0c\u5bfc\u81f4\u4f60 n \u6b21\u8c03\u7528\u5927\u6a21\u578b\uff08 n&gt;1 \uff09\u65f6\uff0c\u5927\u6982\u7387 llama.cpp \u627e\u4e0d\u5230\u4e4b\u524d\u7684\u5bf9\u8bdd\uff0c\u4f1a\u4ece\u5934\u518d\u6b21 prefill \u4f60\u7684\u5bf9\u8bdd\u5168\u6587\u3002</p>\n<h3><strong>\u7ffb\u8bd1\u6210\u5927\u767d\u8bdd\u8bb2\uff0c\u5c31\u662f\u4f60\u5bf9\u4e00\u4e2a\u4eba\uff0c\u6bcf\u591a\u8bf4\u4e00\u53e5\u8bdd\uff0c\u5c31\u8981\u4ece\u7b2c\u4e00\u53e5\u5f00\u59cb\u91cd\u590d\u4e00\u904d\u3002</strong></h3>\n<p>\u66f4\u4e3a\u60b2\u60e8\u7684\u662f\uff1a\n\u5728 5 \u6708\u4efd\uff0cllama.cpp \u5236\u4f5c\u7ec4\u5f15\u5165\u4e86\u53e6\u5916\u4e00\u4e2a checkpoint \u903b\u8f91\uff0c\u4f7f\u5f97\u7f13\u5b58\u5de1\u56de\u6027\u80fd\u518d\u6b21\u4e0b\u964d\uff1a<a href=\"https://github.com/ggml-org/llama.cpp/commit/e98cb51\" rel=\"nofollow\">Commit e98cb51\n</a></p>\n<p><strong>\u7ecf\u8fc7\u6b64\u5e16\u4e2d\u5927\u795e\u5b9e\u6d4b\uff0cNVIDIA RTX PRO 6000 Blackwell \u5728\u8fd0\u884c qwen3.6-27B Q8 \u65f6\uff0c\u4e0a\u4e0b\u6587 50K \u7684\u957f\u5ea6\u4e0b\uff0c\u6bcf\u6b21\u8bf7\u6c42 LLM \u90fd\u4f1a\u6d6a\u8d39 40 \u79d2\uff1a</strong></p>\n<pre><code>3 consecutive full re-processings logged:\n\n\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n\u2502 Turn \u2502 Tokens reprocessed \u2502 Time \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 Task 2795 \u2502 67,608 \u2502 38.4s \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 Task 3241 \u2502 71,211 \u2502 41.0s \u2502\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n\u2502 Task 3401 \u2502 71,105 \u2502 41.4s \u2502\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n\nRoot cause visible in logs: The new prompt is ~19k tokens, but all checkpoints sit at positions 39k\u201371k (from previous longer requests). Every checkpoint\nis checked against 19340 and rejected because they're all beyond the new prompt length. Result: 0 usable checkpoints \u2192 full reprocess from BOS.\n</code></pre>\n<h3>\u7ed3\u8bba\u662f\uff0c\u76ee\u524d\u7684 llama.cpp+qwen3.6-27B \u8fd9\u4e2a\u7ec4\u5408\uff0c\u5728 Agent \u5de5\u5177\u8fd9\u4e2a\u573a\u666f\u4e0b\uff0c\u6027\u80fd\u4e0d\u53ef\u7528\u3002</h3>\n<p>\u76ee\u524d\u6b64 issues \u8fd8\u662f open \u72b6\u6001\uff0c\u5f85\u4fee\u590d\u3002</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/mingtdlb", 
        "name": "mingtdlb", 
        "avatar": "https://cdn.v2ex.com/avatar/1067/fd49/525301_large.png?m=1742795473"
      }, 
      "url": "https://www.v2ex.com/t/1219733", 
      "title": "GPU \u8dd1 LLM \u4e5f\u4f1a\u8d85\u9891\u5417\uff1f", 
      "id": "https://www.v2ex.com/t/1219733", 
      "date_published": "2026-06-11T13:02:51+00:00", 
      "content_html": "<p>\u4f01\u4e1a\u751f\u4ea7\u4e2d\u4e5f\u624b\u52a8\u53bb\u8d85\u9891\u5417\uff1f\u6211\u4ee5\u4e3a\u8ddf CPU \u4e00\u6837\uff0c\u62ff\u6765\u5c31\u76f4\u63a5\u7528\u5462</p>\n<p><img alt=\"image.png\" class=\"embedded_image\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"https://wp-cdn.4ce.cn/v2/OMtwBUv.png\"/></p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/Livid", 
        "name": "Livid", 
        "avatar": "https://cdn.v2ex.com/avatar/c4ca/4238/1_large.png?m=1781025867"
      }, 
      "url": "https://www.v2ex.com/t/1219488", 
      "date_modified": "2026-06-10T19:56:36+00:00", 
      "content_html": "<p><a href=\"https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/\" rel=\"nofollow\">https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/</a></p>\n<p>\u5728\u751f\u6210\u6587\u672c\u65f6\uff0c\u7406\u8bba\u4e0a\u53ef\u4ee5\u6bd4\u73b0\u5728\u7684\u7248\u672c\u5feb 4 \u500d\u3002</p>\n<p>\u672c\u5730\u8fd0\u884c\u6b65\u9aa4\uff1a</p>\n<p><a href=\"https://unsloth.ai/docs/models/diffusiongemma\" rel=\"nofollow\">https://unsloth.ai/docs/models/diffusiongemma</a></p>\n<p>\u76ee\u524d V2EX Chat \u7528\u7684\u6a21\u578b\u5c31\u662f gemma4:26b \u3002</p>\n<p><a href=\"https://edge.v2ex.com/chat\" rel=\"nofollow\">https://edge.v2ex.com/chat</a></p>\n", 
      "date_published": "2026-06-10T18:52:48+00:00", 
      "title": "DiffusionGemma", 
      "id": "https://www.v2ex.com/t/1219488"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/yuping913", 
        "name": "yuping913", 
        "avatar": "https://cdn.v2ex.com/gravatar/68ce659309e68ba2f67dc3cd5cd0df91?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1219170", 
      "date_modified": "2026-06-09T11:55:26+00:00", 
      "content_html": "\u663e\u5361\u53ea\u662f 3080 \u663e\u5b58 10G \uff0c\u4e4b\u524d\u8dd1 qwen3.5 9b mtp \u53ea\u6709 75token/s,\u90fd\u662f q4 \uff0c\u4eca\u5929\u8bd5\u4e86\u4e00\u4e0b Gemma4 12b \u901f\u5ea6 85~105token/s,\u73b0\u5728 MTP \u6280\u672f\u90a3\u4e48\u725b\u5417\uff1f\u6d4b\u4e86\u51e0\u4e2a\u95ee\u9898\u611f\u89c9\u8d28\u91cf\u8fd8\u6bd4 qwen3.5 9b \u597d\u90a3\u4e48\u4e00\u4e22\u4e22\u3002\u6709\u6ca1\u6709\u5927\u795e\u89e3\u60d1\uff1f<br /><br />llama-server.exe ^<br />      --model \"emma-4-12B-it-qat-q4_0-unquantized-heretic-Q4_0.gguf\" ^<br />      --mmproj \"mmproj-gemma-4-12b-it-qat-q4_0.gguf\" ^<br />      --model-draft \"gemma-4-12b-qat-it-assistant-Q4_0_Q4emb.gguf\" ^<br />      --spec-type draft-mtp --spec-draft-n-max 3  ^<br />      --spec-draft-type-k q4_0 --spec-draft-type-v q4_0 ^<br />      --n-gpu-layers-draft 999 ^<br />      --cache-type-k q4_0 ^<br />      --cache-type-v q4_0 ^<br />      --n-gpu-layers 999 ^<br />      --no-mmap ^<br />      --cache-prompt ^<br />      --mlock ^<br />      --kv-unified ^<br />      --parallel 1 ^<br />      -fa on ^<br />      --fit off ^<br />      --ctx-size 100000 --n-predict 10000 ^<br />      --host 0.0.0.0 --port 11432", 
      "date_published": "2026-06-09T11:53:28+00:00", 
      "title": "Gemma4 12b \u5c45\u7136\u6bd4 Qwen3.5 9b \u8fd8\u5feb\uff0c\u610f\u6599\u4e0d\u5230", 
      "id": "https://www.v2ex.com/t/1219170"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/ericterminal", 
        "name": "ericterminal", 
        "avatar": "https://cdn.v2ex.com/avatar/8d76/04e4/780505_large.png?m=1770439160"
      }, 
      "url": "https://www.v2ex.com/t/1219137", 
      "date_modified": "2026-06-09T09:38:49+00:00", 
      "content_html": "<p>\u539f\u8c05\u6211\u8fd9\u4e2a\u6807\u9898\u611f\u89c9\u6709\u70b9\u9a97\u4eba\u8fdb\u6765\u7684\u611f\u89c9\u4f46\u662f\u771f\u505a\u5230\u4e86</p>\n<p>\u67d0\u5929\u5199\u4ee3\u7801\u7684\u65f6\u5019\u6211\u7a81\u7136\u7075\u5149\u4e00\u73b0\uff0cApple Watch \u53ef\u4ee5\u8dd1 C/C++\uff0cllama.cpp \u5c31\u662f C++\u5199\u7684\uff0c\u90a3\u4e48\u80fd\u4e0d\u80fd\u8ba9 Apple Watch \u8dd1 llama.cpp \u5462\uff1f</p>\n<p>\u7136\u540e\u6211\u82b1\u4e86\u51e0\u5929\u52aa\u529b\u628a llama.cpp \u901a\u8fc7\u4f1e\u5934\u6587\u4ef6\u6865\u63a5\u8fdb\u4e86\u652f\u6301 Apple Watch \u7684 Swift \u7a0b\u5e8f</p>\n<p>\u6211\u521a\u624d\u628a Qwen3.5-0.8B-Q4_K_M.gguf \u585e\u8fdb\u4e86\u6211\u7684 Apple Watch S8 \u91cc\u9762</p>\n<h2>\u80fd\u8dd1\u54e6\u9f41\u9f41\u9f41\u9f41\u54e6\u9f41\u9f41\u9f41\u9f41\u2764\ufe0f\u2764\ufe0f\u2764\ufe0f\u2764\ufe0f\uff01\uff01</h2>\n<p>\u8fd9\u9897 t8301 \u633a\u8010\u64cd\u7684\uff0c\u867d\u7136\u901f\u5ea6\u6709\u70b9\u611f\u4eba\uff0c\u624d 0.27token/s \uff0c\u7eaf CPU \u7b97\u7684\uff0c\u5cf0\u503c\u80fd\u529b\u5e94\u8be5\u6709 iPhone6s \u7684\u516b\u6210\u6c34\u5e73</p>\n<p>\u4f46\u662f\u5982\u679c\u771f\u4e0a\u6700\u65b0\u7684 iPhone \u7684\u8bdd\u4f30\u8ba1\u53ef\u4ee5\u8dd1\u5230\u4e0a\u767e token/s \uff0c\u6bd5\u7adf\u6709 Metal</p>\n<p>\u4e0d\u8981\u95ee\u6709\u5565\u610f\u4e49\uff0c\u4e4b\u524d\u7ed9 iPhone \u5237 MIUI \u6ca1\u610f\u4e49\u4e0d\u4e5f\u6709\u4eba\u5e72\u4e86\u561b hhhhh</p>\n<p>\u6211\u8fd8\u60f3\u53d1 B \u7ad9\u6216\u8005\u6cb9\u7ba1\uff0c\u4f46\u662f\u8fd9\u4e2a\u901f\u5ea6\uff0c\u600e\u4e48\u597d\u8ba9\u4eba\u5bb6\u4e00\u773c\u770b\u5230\u529f\u80fd\u5462\u54c8\u54c8\u54c8\n(\u9065\u60f3\u5f53\u5e74\uff0ciPhone \u5f00\u673a\u51fa\u73b0\u7684\u90a3\u4e2a MI \u56fe\u6807)</p>\n<p>iOS \u548c watchOS \u90fd\u53ef\u4ee5\u7528\uff0cGitHub \u4ed3\u5e93\u662f\n<a href=\"https://github.com/Eric-Terminal/ETOS-LLM-Studio\" rel=\"nofollow\">https://github.com/Eric-Terminal/ETOS-LLM-Studio</a></p>\n", 
      "date_published": "2026-06-09T09:36:47+00:00", 
      "title": "\u4ec0\u4e48\uff1f Apple Watch \u4e5f\u80fd\u672c\u5730\u8dd1 Qwen \u4e86\uff1f", 
      "id": "https://www.v2ex.com/t/1219137"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/zzutmebwd", 
        "name": "zzutmebwd", 
        "avatar": "https://cdn.v2ex.com/avatar/a6d2/40df/62082_large.png?m=1782215192"
      }, 
      "url": "https://www.v2ex.com/t/1219024", 
      "title": "\u5173\u4e8e\u4f4e\u7b97\u529b gpu \u63a8\u7406\u65f6 prefill \u5728\u603b\u65f6\u957f\u4e2d\u7684\u5360\u6bd4\u95ee\u9898", 
      "id": "https://www.v2ex.com/t/1219024", 
      "date_published": "2026-06-09T04:18:31+00:00", 
      "content_html": "<p>\u770b\u5230\u5f88\u591a\u4eba\u5bf9 llm \u63a8\u7406\u901f\u5ea6\u7684\u63cf\u8ff0\u90fd\u662f decode \u4e3b\u5bfc/\u5e26\u5bbd\u63a7\u5236/prefill \u5ffd\u7565\u4e0d\u8ba1\uff0c\u6211\u60f3\u8981\u63d0\u9192\u7684\u662f\uff0c\u8fd9\u53ea\u5bf9\u9ad8\u7b97\u529b gpu/\u4ee3\u7801\u7b49\u5bc6\u96c6\u63a8\u7406\u6765\u8bf4\u662f\u5ba2\u89c2\u7684\uff0c\u5982 pro6000/5090 \u8fd9\u7c7b\uff0c\u672c\u5730 agent \u573a\u666f\u5e76\u4e0d\u662f\u8fd9\u6837\u3002</p>\n<p>\u9996\u5148\u660e\u786e\u51e0\u4e2a\u95ee\u9898\uff1a\n1 \u3001\u672a\u547d\u4e2d\u7f13\u5b58\u7684\u8f93\u5165\u91cf\uff1a\u8f93\u51fa\u91cf\u662f\u591a\u5c11\uff1f\u957f\u8f93\u51fa\u7684\u5bc6\u96c6\u63a8\u7406\u5f80\u5f80\u8f93\u51fa\u5927\u4e8e\u8f93\u5165\uff08\u672a\u547d\u4e2d\u7f13\u5b58\u90e8\u5206\uff09\uff0c\u751a\u81f3\u80fd\u8fbe\u5230 2:1 \u3002\u5de5\u5177\u5bc6\u96c6\u7684 agent \u573a\u666f\uff0c\u6839\u636e\u6211\u7684 hermes agent \u7684\u6570\u636e\uff0c\u6700\u8fd1\u4e09\u5929\u7684\u6570\u636e\u662f\u65b0\u8f93\u5165\u91cf / \u8f93\u51fa\u91cf = 4,882,795 / 377,561 \u2248 12.9 : 1,\u4e3b\u8981\u4efb\u52a1\u662f\u4fe1\u606f\u68c0\u7d22/\u6c47\u603b/\u6587\u4ef6\u5904\u7406/\u667a\u80fd\u5bb6\u5c45\u3002\n2 \u3001\u672c\u5730 agent \u66f4\u591a\u7684\u5de5\u4f5c\u5728\u54ea\u4e2a\u573a\u666f\uff1f\u6211\u8ba4\u4e3a\u4e3b\u6d41\u573a\u666f\u662f 12.9:1 \u8fd9\u79cd\uff0c\u6307\u671b\u672c\u5730 ai \u8dd1\u5bc6\u96c6\u63a8\u7406+\u7f16\u7801\u4efb\u52a1\u4e0d\u592a\u73b0\u5b9e\u554a\u3002\n3 \u3001\u4e0d\u540c\u786c\u4ef6\u7684 prefill \u901f\u5ea6\u548c decode \u901f\u5ea6\uff1f\u4ee5\u8fd1\u671f\u6700\u706b\u7684 qwen3.6 27b \u4e3a\u4f8b\uff08 8bit \u5f00 mtp \u53c2\u8003\u503c\uff09\uff0c5090 prefill 3000tps \uff0cdecode 70tps \uff0cm3 ultra prefill 300tps \uff0cdecode 30tps \u3002\n4 \u3001\u6b64\u65f6\uff0c5090 prefill 1628s \uff0cdecode 5394s \uff0c\u786e\u5b9e\u662f decode/\u5e26\u5bbd\u4e3b\u5bfc\uff1b m3 ultra prefill 16276s \uff0cdecode 12585s, prefill \u5360\u6bd4 56%\u3002\n5 \u3001\u5bf9\u4e8e\u672c\u5730\u90e8\u7f72\u5e38\u89c1\u7684 4bit \uff0cprefill \u65f6\u95f4\u5360\u6bd4\u66f4\u9ad8\u3002</p>\n<p>\u7efc\u4e0a\u6240\u8ff0\uff0c\u5bf9\u4e8e\u4f4e\u7b97\u529b/\u5927\u663e\u5b58\u8bbe\u5907\uff0cprefill \u6240\u7528\u65f6\u957f\u662f\u76f8\u5f53\u663e\u8457\u7684\uff0c\u5728\u5de5\u5177\u8c03\u7528\u5bc6\u96c6\u578b agent \u4e2d\u751a\u81f3\u5360\u6709\u4e3b\u5bfc\u5730\u4f4d\u3002</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/mingtdlb", 
        "name": "mingtdlb", 
        "avatar": "https://cdn.v2ex.com/avatar/1067/fd49/525301_large.png?m=1742795473"
      }, 
      "url": "https://www.v2ex.com/t/1218934", 
      "title": "\u73b0\u5728\u5927\u6a21\u578b\u4e3b\u6d41\u90fd\u7528\u54ea\u4e9b nVidia GPU\uff1f", 
      "id": "https://www.v2ex.com/t/1218934", 
      "date_published": "2026-06-09T01:24:08+00:00", 
      "content_html": "<p>\u4e0d\u9650\u4e8e\u53c2\u6570\u5927\u5c0f</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/TGOcc", 
        "name": "TGOcc", 
        "avatar": "https://cdn.v2ex.com/avatar/10f3/3f11/817343_large.png?m=1780891234"
      }, 
      "url": "https://www.v2ex.com/t/1218726", 
      "date_modified": "2026-06-08T05:39:04+00:00", 
      "content_html": "<p>\u5148\u8bf4\u7ed3\u8bba\uff0c\u80fd\u8dd1\uff0c\u4f46\u6ca1\u529e\u6cd5\u957f\u671f\u8dd1\uff0c\u4e3b\u8981\u95ee\u9898\u662f\u6563\u70ed\uff0c\u5916\u6302\u98ce\u6247\u652f\u67b6\u4e5f\u4e0d\u592a\u80fd\u89e3\u51b3\u95ee\u9898\uff0c\u9ad8\u5f3a\u5ea6\u8dd1\u6e29\u5ea6\u4e0a\u5347\u5feb\uff0c\u6301\u7eed\u9ad8\u6e29\u673a\u5668\u4f1a\u964d\u9891\u3002\u5982\u679c\u8003\u8651\u4fbf\u643a+\u751f\u4ea7\u529b\uff0c\u63a8\u8350\u4e0a mac book pro \u5427\u3002</p>\n<p>\u88c5\u4e86\u4e24\u4e2a\u5e73\u53f0\uff0collama \u8ddf olmx \uff0c\u6d4b\u8bd5\u4e0b\u6765\uff0colmx \u5e73\u53f0\u4f1a\u66f4\u5feb\u4e9b\uff0c\u8003\u8651\u5230\u673a\u5668 32G \u7684\u5185\u5b58\uff0c\u80fd\u8dd1\u7684\u6a21\u578b\u5927\u5c0f\u4e0d\u8981\u8d85 22GB</p>\n<p>\u9644\u4e0a\u90e8\u5206\u4e3b\u6d41\u6a21\u578b\u4e0b\u8f7d\u5bb9\u91cf\u5927\u5c0f\u53ca olmx \u5e73\u53f0\u6d4b\u8bd5\u7ed3\u679c\u7ed9\u5927\u5bb6\u505a\u53c2\u8003</p>\n<p>Qwen3.5-4B-MLX-4bit      2.85GB</p>\n<p>gemma-4-26b-a4b-it-4bit  14.57GB</p>\n<p>Qwen3.6-35B-A3B-4bit     15.13GB</p>\n<p>GLM-4.7-Flash-4bit       15.71GB</p>\n<p>gpt-oss-20b-MXFP4-Q8     11.27GB</p>\n<pre>oMLX - LLM inference, optimized for your Mac\n\nBenchmark Model: Qwen3.5-4B-MLX-4bit\n================================================================================\nSingle Request Results\n--------------------------------------------------------------------------------\nTest             TTFT(ms)    TPOT(ms)        pp TPS        tg TPS    E2E(s)    Throughput    Peak Mem\npp1024/tg128       1001.6       22.74  1022.4 tok/s    44.3 tok/s     3.889   296.2 tok/s     3.29 GB\npp4096/tg128       3540.9       23.76  1156.8 tok/s    42.4 tok/s     6.558   644.1 tok/s     3.90 GB\n\nContinuous Batching\npp1024 / tg128\n--------------------------------------------------------------------------------\nBatch         tg TPS    Speedup          pp TPS    pp TPS/req    TTFT(ms)      E2E(s)\n1x        44.3 tok/s      1.00x    1022.4 tok/s  1022.4 tok/s      1001.6       3.889\n2x        88.3 tok/s      1.99x     407.6 tok/s   203.8 tok/s      3040.1       7.924\n4x       175.1 tok/s      3.95x     322.7 tok/s    80.7 tok/s      6833.9      15.617\n\n\nBenchmark Model: gemma-4-26b-a4b-it-4bit\n================================================================================\nSingle Request Results\n--------------------------------------------------------------------------------\nTest             TTFT(ms)    TPOT(ms)        pp TPS        tg TPS    E2E(s)    Throughput    Peak Mem\npp1024/tg128       1500.5       24.21   682.4 tok/s    41.6 tok/s     4.575   251.8 tok/s    14.23 GB\npp4096/tg128       4863.4       25.14   842.2 tok/s    40.1 tok/s     8.056   524.3 tok/s    14.91 GB\n\nContinuous Batching\npp1024 / tg128\n--------------------------------------------------------------------------------\nBatch         tg TPS    Speedup          pp TPS    pp TPS/req    TTFT(ms)      E2E(s)\n1x        41.6 tok/s      1.00x     682.4 tok/s   682.4 tok/s      1500.5       4.575\n2x        82.5 tok/s      1.98x     361.6 tok/s   180.8 tok/s      3495.8       8.767\n4x       166.1 tok/s      3.99x     283.4 tok/s    70.8 tok/s      7840.6      17.536\n\n\nBenchmark Model: Qwen3.6-35B-A3B-4bit\n================================================================================\nSingle Request Results\n--------------------------------------------------------------------------------\nTest             TTFT(ms)    TPOT(ms)        pp TPS        tg TPS    E2E(s)    Throughput    Peak Mem\npp1024/tg128       1676.1       17.20   610.9 tok/s    58.6 tok/s     3.860   298.4 tok/s    18.80 GB\npp4096/tg128       5046.3       17.93   811.7 tok/s    56.2 tok/s     7.323   576.8 tok/s    19.24 GB\n\nContinuous Batching\npp1024 / tg128\n--------------------------------------------------------------------------------\nBatch         tg TPS    Speedup          pp TPS    pp TPS/req    TTFT(ms)      E2E(s)\n1x        58.6 tok/s      1.00x     610.9 tok/s   610.9 tok/s      1676.1       3.860\n2x       116.2 tok/s      1.98x     435.5 tok/s   217.8 tok/s      2973.7       6.907\n4x       230.7 tok/s      3.94x     352.0 tok/s    88.0 tok/s      6445.2      13.855\n\n\nBenchmark Model: GLM-4.7-Flash-4bit\n================================================================================\nSingle Request Results\n--------------------------------------------------------------------------------\nTest             TTFT(ms)    TPOT(ms)        pp TPS        tg TPS    E2E(s)    Throughput    Peak Mem\npp1024/tg128       1985.0       21.78   515.9 tok/s    46.3 tok/s     4.752   242.4 tok/s    16.27 GB\npp4096/tg128       6839.2       27.31   598.9 tok/s    36.9 tok/s    10.307   409.8 tok/s    17.34 GB\n\nContinuous Batching\npp1024 / tg128\n--------------------------------------------------------------------------------\nBatch         tg TPS    Speedup          pp TPS    pp TPS/req    TTFT(ms)      E2E(s)\n1x        46.3 tok/s      1.00x     515.9 tok/s   515.9 tok/s      1985.0       4.752\n2x        91.5 tok/s      1.98x     362.7 tok/s   181.3 tok/s      3549.9       8.445\n4x       174.9 tok/s      3.78x     321.2 tok/s    80.3 tok/s      6393.9      15.679\n\n\nBenchmark Model: gpt-oss-20b-MXFP4-Q8\n================================================================================\nSingle Request Results\n--------------------------------------------------------------------------------\nTest             TTFT(ms)    TPOT(ms)        pp TPS        tg TPS    E2E(s)    Throughput    Peak Mem\npp1024/tg128       1687.6       24.70   606.8 tok/s    40.8 tok/s     4.824   238.8 tok/s    11.67 GB\npp4096/tg128       4088.8       26.44  1001.8 tok/s    38.1 tok/s     7.446   567.3 tok/s    11.75 GB\n\nContinuous Batching\npp1024 / tg128\n--------------------------------------------------------------------------------\nBatch         tg TPS    Speedup          pp TPS    pp TPS/req    TTFT(ms)      E2E(s)\n1x        40.8 tok/s      1.00x     606.8 tok/s   606.8 tok/s      1687.6       4.824\n2x        82.1 tok/s      2.01x     359.0 tok/s   179.5 tok/s      3489.1       8.822\n4x       159.5 tok/s      3.91x     293.2 tok/s    73.3 tok/s      7335.0      17.180\n</pre>", 
      "date_published": "2026-06-08T04:12:47+00:00", 
      "title": "Mac book air M5 32G+1TB \u80fd\u8dd1\u672c\u5730\u5927\u6a21\u578b\uff1f", 
      "id": "https://www.v2ex.com/t/1218726"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/Flagship9945", 
        "name": "Flagship9945", 
        "avatar": "https://cdn.v2ex.com/gravatar/bbc2c775cf3f0f3a7decf56693ca676f?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1218631", 
      "date_modified": "2026-06-08T01:40:28+00:00", 
      "content_html": "<ul>\n<li>\u4ece\u90e8\u7f72\u3001\u5382\u5546\u652f\u6301\u7b49\u89d2\u5ea6\u6765\u8bf4</li>\n<li>200w \u4ee5\u5185\u663e\u5361\u9884\u7b97</li>\n</ul>\n", 
      "date_published": "2026-06-08T00:49:02+00:00", 
      "title": "\u9700\u8981\u8d2d\u4e70\u56fd\u4ea7\u663e\u5361\u672c\u5730\u90e8\u7f72\u5927\u6a21\u578b\uff0c\u54ea\u5bb6\u7684\u6bd4\u8f83\u597d", 
      "id": "https://www.v2ex.com/t/1218631"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/kakalulin", 
        "name": "kakalulin", 
        "avatar": "https://cdn.v2ex.com/avatar/7a22/35e1/345671_large.png?m=1665393617"
      }, 
      "url": "https://www.v2ex.com/t/1218163", 
      "date_modified": "2026-06-05T06:40:02+00:00", 
      "content_html": "<p>\u4e0b\u534a\u5e74\u6253\u7b97\u5165\u4e2a mac mini \uff0c\u7528\u6765\u8dd1\u672c\u5730\u6a21\u578b+hermes \u3002</p>\n<p>\u5927\u6982\u9700\u8981\u4ec0\u4e48\u914d\u7f6e\uff1f\n\uff08\u6a21\u578b-\u5bf9\u5e94-\u914d\u7f6e\uff09</p>\n<p>\u53e6\uff1a\u5927\u5bb6\u89c9\u5f97\u5e74\u5e95\uff0cmac mini \u4e8c\u624b\u4ef7\u683c\u80fd\u4e0b\u6765\u5417\uff1f</p>\n", 
      "date_published": "2026-06-05T05:20:14+00:00", 
      "title": "mac mini \u8dd1\u672c\u5730\u6a21\u578b\uff0c\u9700\u8981\u4ec0\u4e48\u914d\u7f6e\uff1f", 
      "id": "https://www.v2ex.com/t/1218163"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/CatCode", 
        "name": "CatCode", 
        "avatar": "https://cdn.v2ex.com/gravatar/7dfa8c7d43ca8f5bb37248a2009fa040?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1218059", 
      "title": "Gemma4 12B \u5982\u4f55\u8dd1\u5728 16G \u663e\u5b58\u4e0a\uff1f", 
      "id": "https://www.v2ex.com/t/1218059", 
      "date_published": "2026-06-05T00:45:01+00:00", 
      "content_html": "<p>Google \u53d1\u5e03\u4e86 Gemma 4 \u7684\u4e00\u4e2a\u65b0\u6a21\u578b\uff0c12B \u53c2\u6570\uff0c\u770b\u4ecb\u7ecd\u4e0d\u662f MoE \u3002<br/>\n<a href=\"https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12b/\" rel=\"nofollow\">https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12b/</a></p>\n<p>\u770b HF \u548c Kaggle \u4e0a\u90fd\u662f BF16 \u6570\u636e\u7c7b\u578b\uff0c\u6743\u91cd\u6587\u4ef6\u5927\u5c0f 23.9GB \u5de6\u53f3\u3002<br/>\n<a href=\"https://huggingface.co/google/gemma-4-12B-it/tree/main\" rel=\"nofollow\">https://huggingface.co/google/gemma-4-12B-it/tree/main</a><br/>\n<a href=\"https://www.kaggle.com/models/google/gemma-4/transformers/gemma-4-12b-it\" rel=\"nofollow\">https://www.kaggle.com/models/google/gemma-4/transformers/gemma-4-12b-it</a></p>\n<p>Google \u5728\u535a\u5ba2\u91cc\u4e13\u95e8\u5f3a\u8c03\u4e86 Laptop ready: Small enough to run locally with just 16GB of VRAM or unified memory.</p>\n<p>\u8fd9\u662f\u600e\u4e48\u505a\u5230\u80fd\u5728 16G \u663e\u5b58\u4e0a\u8dd1\u7684\uff1f<br/>\n\u8fd8\u662f\u8bf4 BF16 \u7684\u4e0d\u80fd\u8dd1\uff0c\u8981 FP8 \u91cf\u5316\u7684\u624d\u884c\uff1f\u4f46\u8fd9\u79cd\u91cf\u5316\u4e4b\u540e\u80fd\u5728 16G \u5361\u4e0a\u8dd1\u7684\u6a21\u578b\u5f88\u591a\u4e86\uff0c\u8fd8\u6709\u5f88\u591a\u53c2\u6570\u91cf\u66f4\u5927\u7684\u6a21\u578b\u3002</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/followadc", 
        "name": "followadc", 
        "avatar": "https://cdn.v2ex.com/avatar/befe/b8e8/717730_large.png?m=1741270230"
      }, 
      "url": "https://www.v2ex.com/t/1217501", 
      "date_modified": "2026-06-03T02:53:26+00:00", 
      "content_html": "\u6700\u8fd1\u60f3\u5728\u672c\u5730\u90e8\u5c5e\u4e2a qwenpaw \u7528\u7528\u3002\u8bbe\u5907\u662f mac m4 64g \u3002\u60f3\u77e5\u9053\u8fd9\u4e2a\u80fd\u90e8\u7f72\u54ea\u4e2a\u672c\u5730\u5927\u6a21\u578b \u4e0d\u592a\u61c2 \u7eaf\u8bf7\u6559", 
      "date_published": "2026-06-03T01:53:57+00:00", 
      "title": "mac 64g \u80fd\u90e8\u7f72\u54ea\u4e2a\u672c\u5730\u5927\u6a21\u578b", 
      "id": "https://www.v2ex.com/t/1217501"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/zhengfan2016", 
        "name": "zhengfan2016", 
        "avatar": "https://cdn.v2ex.com/gravatar/df526f138d10cac8c95b274c720a6f55?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1216752", 
      "date_modified": "2026-05-31T02:42:40+00:00", 
      "content_html": "<p>\u5982\u9898\uff0cwsl \u914d rocm \u4e0b\uff0csglang \u6ca1\u8dd1\u8d77\u6765\uff0cvllm \u8dd1\u8d77\u6765\u4e86\uff0c\u4f46\u662f\u52a8\u4e0d\u52a8\u7206\u663e\u5b58\uff0c\u53ea\u6709\u8dd1\u4e2a 2b \u7684\u6a21\u578b\u624d\u6bd4\u8f83\u7a33\u5b9a\uff0c\u800c\u4e14\u63a8\u7406\u9996\u5b57\u901f\u5ea6\u4f53\u611f\u611f\u89c9\u6bd4\u7eaf\u7528 transformer \u8fd8\u6162\u3002</p>\n<p>transformer \u6211\u8bd5\u4e86\u53ef\u4ee5\u6210\u529f\u8dd1\u4e2a 9b \u7684 gptq \u6a21\u578b(vllm \u8fd9\u4e2a\u6a21\u578b\u8dd1\u4e0d\u6210\u529f\u62a5\u9519 qwen3.5 \u4ec0\u4e48 config \u6709\u95ee\u9898\uff0cclaudecode \u4fee\u4e0d\u4e86)\uff0c\u662f\u6211\u4e0d\u4f1a\u7528 vllm \u8fd8\u662f\u6d88\u8d39\u7ea7\u663e\u5361\u5c31\u662f\u4e0d\u9002\u5408\u7528\u8fd9\u7c7b\u63a8\u7406\u6846\u67b6\uff1f</p>\n", 
      "date_published": "2026-05-31T02:41:52+00:00", 
      "title": "\u6d88\u8d39\u7ea7\u663e\u5361(16G A \u5361)\u662f\u4e0d\u662f\u4e0d\u9002\u5408\u8fd0\u884c vllm \u548c sglang\uff0c\u597d\u50cf\u4f7f\u7528 transformer \u63a8\u7406\u90fd\u6bd4\u8fd9\u4e24\u4e2a\u6846\u67b6\u5feb\uff0c\u5e76\u4e14\u5360\u7528\u663e\u5b58\u4f4e", 
      "id": "https://www.v2ex.com/t/1216752"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/SteveRogers", 
        "name": "SteveRogers", 
        "avatar": "https://cdn.v2ex.com/gravatar/83774984a69850febb48ab0d6bedb840?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1215946", 
      "date_modified": "2026-05-28T00:40:49+00:00", 
      "content_html": "\u8fd1\u671f\u770b\u4e86\u4e0b\uff0c\u56e0\u4e3a\u4eba\u5728\u6df1\u5733\uff0c\u53bb hk \u975e\u5e38\u65b9\u4fbf\uff0c\u52a0\u4e0a\u6709\u6559\u80b2\u6e20\u9053\uff0c\u60f3\u9009\u4e00\u53f0 mac studio \u4f5c\u4e3a\u4e2a\u4eba\u672c\u5730\u8dd1\u9f99\u867e\u4e13\u7528\uff0c\u4e5f\u4e0d\u51c6\u5907\u5168\u90e8\u672c\u5730\uff0c\u90a3\u4e2a\u5185\u5b58\u6700\u4f4e\u90fd\u8981 128G \uff0c\u6027\u4ef7\u6bd4\u592a\u4f4e\uff0c\u60f3\u7740\u7aef\u4e91\u7ed3\u5408\u7740\u6765\u3002<br /><br />\u6a21\u578b\u4f30\u8ba1\u662f\uff1a<br />1.gemm4 31b<br />2.qwen3 32b<br /><br />\u60f3\u95ee\u95ee\u6709\u672c\u5730\u8dd1\u6a21\u578b\u7684\u5417\uff0c\u8dd1\u7684\u600e\u4e48\u6837\uff1f<br /><br />\u521d\u6b65\u8ba1\u5212\u786c\u4ef6\u4e3a<br /><br />M4Max 64G 1Tb \u8fd9\u4e2a\u7248\u672c\uff0c22749 \u6298\u5408\u4e0b\u6765\u4f30\u8ba1\u5728 19000 \u5de6\u53f3", 
      "date_published": "2026-05-27T08:18:02+00:00", 
      "title": "\u672c\u5730\u5927\u6a21\u578b\u6700\u4f73 Mac \u914d\u7f6e\u9009\u62e9", 
      "id": "https://www.v2ex.com/t/1215946"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/tootfsg", 
        "name": "tootfsg", 
        "avatar": "https://cdn.v2ex.com/gravatar/7ae7755b83c1755a4e5c091e9b63d167?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1213838", 
      "date_modified": "2026-05-19T08:07:00+00:00", 
      "content_html": "\u524d\u7f6e\u6761\u4ef6\uff1a5070ti 16g \uff0cllama.cpp \uff0c\u5168\u8dd1\u5728\u663e\u5b58\u3002<br /><br />1. \u8dd1 gemma4 26b a4b iq4_xs \u91cf\u5316\uff08 MoE \u7ed3\u6784\uff09<br /><br />    \u901f\u5ea6\u5927\u6982\u662f 120t/s-150t/s \uff0c\u9996 token \u548c\u540e\u7eed\u8f93\u51fa\u90fd\u5f88\u5feb<br /><br />2. \u8dd1 devstral small2 24b q4_k_m \u91cf\u5316 \uff08\u7a20\u5bc6\u7ed3\u6784\uff09<br /><br />    \u901f\u5ea6\u5927\u6982\u662f 8t/s-10t/s \uff0c\u9996 token \u53ef\u80fd\u5f88\u6162\uff0c\u6574\u4f53\u8f93\u51fa\u90fd\u6162\u5f97\u591a\u3002<br /><br /><br /><br />\u601d\u8003\uff1a<br /><br />\u73b0\u5728\u7684\u6a21\u578b\u6709\u4e24\u79cd\u7ed3\u6784\uff1a\u7a20\u5bc6\uff08 Dense \uff09\u548c MoE \uff08\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff09\u3002<br /><br />\u4ee5\u4e0a\u8ff0\u4e24\u79cd\u6a21\u578b\u4e3e\u4f8b<br /><br />    \u7a20\u5bc6\u6a21\u578b\u662f\u6240\u6709\u5c42\uff08 dev \u8fd9\u4e2a\u6709 40 \u5c42\uff09\u90fd\u53c2\u4e0e\u8ba1\u7b97\uff0c\u6d88\u8017 24b \u7684\u5b8c\u6574\u7b97\u529b\uff0c\u4e5f\u5c31\u662f\u5355 token 2x24b=48gflops \uff08\u4e0d\u7b97\u91cf\u5316\uff09\uff0c\u7b97\u529b\u6d88\u8017\u5927\uff0c\u63a8\u7406\u6210\u672c\u9ad8\u3002<br /><br />    moe \u662f\u603b\u5171 26b \u53c2\u6570\uff0c\u6bcf\u6b21\u63a8\u7406\u53ea\u6fc0\u6d3b 4b<br /><br />\u53c2\u6570\uff0c\u53ea\u6d88\u8017\u6fc0\u6d3b\u53c2\u6570 4b \u7684\u7b97\u529b\uff0c\u5355 token \u7b97\u529b\u6d88\u8017 2x4=8gflops \uff0c\u7b97\u529b\u6d88\u8017\u5c0f\u5f88\u591a\uff0c\u4f46\u6709 26b \u7684\u53c2\u6570\uff08\u77e5\u8bc6\uff09\u3002gemma \u8fd9\u4e2a\u6709 128 \u4e2a\u4e13\u5bb6\uff0c\u6bcf\u6b21\u6fc0\u6d3b 8 \u4e2a\u4e13\u5bb6\u548c 1 \u4e2a\u5171\u4eab\u4e13\u5bb6\uff08\u6240\u6709 token \u5fc5\u987b\u9996\u5148\u7ecf\u8fc7\u5171\u4eab\u4e13\u5bb6\uff09\uff0cmoe \u6a21\u578b\u662f\u901a\u8fc7\u52a8\u6001\u8def\u7531\u5224\u65ad\u9009\u62e9\u4e13\u5bb6\u7684\u3002<br /><br /><br /><br />\u53ef\u4ee5\u770b\u51fa\u7b97\u529b\u9700\u6c42\u5dee\u5f02\u5de8\u5927\u3002<br /><br /><br /><br />\u5e38\u89c1\u7684\u51e0\u4e2a\u9876\u7ea7\u5f00\u6e90\u6a21\u578b<br /><br />glm5.1 \u53c2\u6570 754b \u6fc0\u6d3b 40b<br /><br />deepseek-v4 pro \u53c2\u6570 1.6t \u6fc0\u6d3b 49b<br /><br />v4 flash \u53c2\u6570 284b \u6fc0\u6d3b 13b<br /><br />minimax2.5 \u53c2\u6570 229b \u6fc0\u6d3b 10b<br /><br /><br /><br />moe \u6a21\u578b\u867d\u7136\u6bcf\u6b21\u6fc0\u6d3b\u7684\u53c2\u6570\u5c11\uff0c\u4f46\u5fc5\u987b\u628a\u5b8c\u6574\u53c2\u6570\u90fd\u5168\u91cf\u52a0\u8f7d\u5230\u663e\u5b58\u4e2d\u3002\u4e5f\u5c31\u662f\u8bf4\u7b97\u529b\u6d88\u8017\u5927\u5927\u51cf\u5c11\uff0c\u4f46\u663e\u5b58\u9700\u6c42\u6ca1\u53d8\u3002<br /><br /><br /><br />\u53ef\u4ee5\u5927\u6982\u63a8\u6d4b\uff0c\u9876\u7ea7\u5927\u6a21\u578b\u4ee5\u540e\u53ef\u80fd\u53ea\u6709 moe \u7ed3\u6784\u4e86\uff0c\u53c2\u6570\u5c0f\u7684\u53ef\u80fd\u6709\u7a20\u5bc6\u67b6\u6784\uff0c\u56e0\u4e3a\u7b97\u529b\u6210\u672c\u8fd8\u5c1a\u53ef\u63a5\u53d7\uff0c\u53c2\u6570\u91cf\u5f88\u5927\u7684\u7a20\u5bc6\u7ed3\u6784\uff0c\u6050\u6015\u7b97\u529b\u6210\u672c\u9ad8\u5230\u5382\u5546\u4e5f\u96be\u4ee5\u5546\u7528\u5427\u3002<br /><br /><br /><br />\u672c\u5730\u90e8\u7f72\uff0c\u6211\u770b\u6765\u63a8\u7406\u901f\u5ea6\u6709 40-50token/s \uff0c\u57fa\u672c\u53ef\u4ee5\u81ea\u7528\u4e86\uff0c\u8fd9\u662f\u4e00\u4e2a\u53ca\u683c\u7ebf\u3002<br /><br /><br /><br />\u6211\u770b\u6765\u6709\u4e24\u79cd\u6bd4\u8f83\u597d\u7684\u672c\u5730\u90e8\u7f72\u65b9\u6848<br /><br /><br /><br />1. \u4e70 nv \u5de5\u4f5c\u7ad9\u663e\u5361\uff0cpro6000 96g \u54b8\u9c7c 6w \u591a\uff0cpro6000d 84g \uff08\u663e\u5b58\u6ca1 ecc \uff0c\u6574\u4f53\u6bd4 6000 \u7565\u5dee\uff09\u54b8\u9c7c 4w \uff0cpro5000 84g \u8fd9\u79cd\u3002<br /><br />2. \u7528\u540c\u7b49\u4ef7\u94b1\u7a0d\u5fae\u4f4e\u70b9\uff0c\u7b49 m5 pro \u7684 mac mini/studio \u53d1\u5e03\u540e\u8d2d\u4e70\u3002<br /><br /><br /><br />\u6539\u663e\u5b58\uff0c\u77ff\u5361\uff0c\u4e8c\u624b\u7684\u5f88\u4e45\u7684\u4e13\u4e1a\u5361\u7b49\u5c31\u4e0d\u8ba8\u8bba\u4e86\uff0c\u4e0d\u61c2\u8fd9\u90e8\u5206\u3002<br /><br /><br /><br />mac \u8dd1\u63a8\u7406\uff0colmx \u5b98\u7f51\u6211\u770b\u4e86\u6a21\u578b\u63a8\u7406\u901f\u5ea6\u6392\u884c\u699c\uff0c\u8fd8\u662f\u5dee\u4e86\u70b9\uff0c\u4e0d\u77e5\u9053 4w \u4ef7\u94b1\u7684 m5 pro \u7684 mac mini/studio \u4f1a\u4e0d\u4f1a\u660e\u663e\u63d0\u9ad8\u3002<br /><br /><br /><br />\u8fd8\u6709\u5c31\u662f\u6bd4\u5982\u53cc 5070ti \u8dd1\u6a21\u578b\u63a8\u7406\uff0c\u4e0d\u77e5\u9053\u901f\u5ea6\u600e\u4e48\u6837\uff0c\u4ef7\u94b1\u76f8\u5bf9\u4e0d\u8d35\u3002\u6211\u7528\u7684\u662f ddr4 pcie 4.0 \u7684\u4e3b\u677f\uff0c\u53cc\u663e\u5361\u8981 pcie \u62c6\u5206 8x8 \uff0cpcie5.0 \u80af\u5b9a\u66f4\u597d\uff0c\u6211\u5f97\u6362\u4e3b\u677f\u6362\u5185\u5b58\uff0c\u6210\u672c\u592a\u9ad8\uff0c\u6ca1\u6cd5\u6d4b\u8bd5\uff0c\u5982\u679c\u5185\u5b58\u6ca1\u8fd9\u4e48\u8d35\uff0c\u5c31\u6362\u4e3b\u677f\u4e70\u5185\u5b58\u641e\u4e2a 5060ti 16g \u6765\u6d4b\u8bd5\u4e86\uff0c\u8fd9\u4e2a\u53ef\u80fd\u4e5f\u662f\u4e00\u79cd\u65b9\u6848\u5427\u3002", 
      "date_published": "2026-05-19T07:51:13+00:00", 
      "title": "\u5173\u4e8e 5070ti \u6a21\u578b\u63a8\u7406\u7684\u901f\u5ea6\u548c\u672c\u5730\u90e8\u7f72\u601d\u8003", 
      "id": "https://www.v2ex.com/t/1213838"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/faketemp", 
        "name": "faketemp", 
        "avatar": "https://cdn.v2ex.com/gravatar/4c4e26185d17f23be03e0c8097d10f3b?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1213462", 
      "date_modified": "2026-05-18T06:35:38+00:00", 
      "content_html": "<p>\u60f3\u7528 ollama \u6216 llm studio \u7b49\u5de5\u5177\u4e0b\u8f7d\u4e2a Hugging Face \u4e0a\u7684\u6a21\u578b\uff0c\u5728 windows7 \u4e2d\u79bb\u7ebf\u8fd0\u884c\u5904\u7406\u4e00\u4e9b\u65e5\u5e38\u5c0f\u9700\u6c42\uff0c\u53d1\u73b0\u8fd9\u4e9b\u5de5\u5177\u6700\u4f4e\u90fd\u662f\u4ec5\u652f\u6301 win10/win11 \uff0c\u6709\u6ca1\u6709\u5927\u4f6c\u7814\u7a76\u8fc7\u6709\u6ca1\u6709\u7c7b\u4f3c\u7684\u80fd\u591f\u52a0\u8f7d\u4f7f\u7528\u79bb\u7ebf\u6a21\u578b\u7684\u517c\u5bb9 Win7 \u7684\u5de5\u5177\u63a8\u8350</p>\n", 
      "date_published": "2026-05-18T03:43:06+00:00", 
      "title": "\u6709\u6ca1\u6709\u80fd\u591f\u517c\u5bb9 Win7 \u7684\u79bb\u7ebf\u6a21\u578b\u5de5\u5177", 
      "id": "https://www.v2ex.com/t/1213462"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/davidyin", 
        "name": "davidyin", 
        "avatar": "https://cdn.v2ex.com/avatar/6dcb/35fd/43242_large.png?m=1768429019"
      }, 
      "url": "https://www.v2ex.com/t/1211566", 
      "title": "\u60f3\u6298\u817e\u4e00\u4e2a AI \u4e3b\u673a\uff0c\u8bf7\u884c\u5bb6\u51fa\u624b", 
      "id": "https://www.v2ex.com/t/1211566", 
      "date_published": "2026-05-09T17:02:50+00:00", 
      "content_html": "\u6253\u7b97\u81ea\u7ec4\u4e00 AI \u4e3b\u673a\uff0c\u7528\u4e8e\u672c\u5730 llm \u3002 \u53ef\u7528\u4e8e kiro IDE \u7684\uff0cgitlab duo \u3002<br /><br /><br />\u53ef\u884c\u6027\u6709\u591a\u5927\uff0c\u80fd\u5426\u4ee3\u66ff\u8ba2\u9605\u7684\u90a3\u4e9b ai \u670d\u52a1\uff1f<br /><br />\u914d\u7f6e\u6709\u6ca1\u6709\u63a8\u8350\u7684\uff0c\u5404 AI \u884c\u5bb6\u8bf7\u51fa\u624b\u76f8\u52a9\u3002"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/KaiWuBOSS", 
        "name": "KaiWuBOSS", 
        "avatar": "https://cdn.v2ex.com/gravatar/243db3a31aa62a02d726471a3fd1782e?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1211391", 
      "title": "\u9524\u5b50\u627e\u9489\u5b50\u7684\u9879\u76ee\u5206\u4eab\uff1a\u5047\u60f3\u4f01\u4e1a\u672c\u5730\u90e8\u7f72\u540e\u4e0d\u7528\u4eba\u5de5\u6d17\u5e93\u63a5\u5165 llm \u7684\u4e2d\u95f4\u5c42\u3002", 
      "id": "https://www.v2ex.com/t/1211391", 
      "date_published": "2026-05-09T03:15:17+00:00", 
      "content_html": "<h1>\u9524\u5b50\u627e\u9489\u5b50\u7684\u9879\u76ee\u5206\u4eab\uff1a\u5047\u60f3\u4f01\u4e1a\u672c\u5730\u90e8\u7f72\u540e\u4e0d\u7528\u4eba\u5de5\u6d17\u5e93\u63a5\u5165 LLM \u7684\u4e2d\u95f4\u5c42</h1>\n<h2>\u6211\u95ee AI \uff0c\u4f01\u4e1a\u6570\u5b57\u5316\u5dee\u4ec0\u4e48\uff1f</h2>\n<h3>\u4ed6\u8bf4\u6700\u96be\u7684\u662f\u6570\u636e\u6e05\u6d17\uff0c\u5e93\u592a\u591a\uff0c\u6570\u636e\u5f55\u5165\u4e0d\u89c4\u8303\uff0c\u5b57\u6bb5\u547d\u540d\u4e71\u3002ai \u8981\u9760\u731c\u3002</h3>\n<p>\u6240\u4ee5\u82b1\u4e86\u4e24\u5468\u5199\u4e86\u4e2a\u4e2d\u95f4\u5c42\uff0c\u60f3\u89e3\u51b3\"\u4f01\u4e1a\u591a\u4e2a\u6570\u636e\u5e93\u63a5 LLM \u65f6\u5b57\u6bb5\u4e71\u3001\u6743\u9650\u4e71\u3001\u53e3\u5f84\u4e71\"\u7684\u95ee\u9898\u3002\u5199\u4e86 7000 \u884c Python \u3001134 \u4e2a\u6d4b\u8bd5\u30013 \u4efd\u67b6\u6784 spec \u3002\u7136\u540e\u610f\u8bc6\u5230\uff1a\u6211\u6ca1\u6709\u7528\u6237\uff0c\u6ca1\u6709\u771f\u5b9e\u573a\u666f\u9a8c\u8bc1\uff0c\u53ef\u80fd\u4ece\u5934\u5230\u5c3e\u5728\u89e3\u51b3\u4e00\u4e2a\u6211\u60f3\u8c61\u51fa\u6765\u7684\u95ee\u9898\u3002</p>\n<p>\u53d1\u51fa\u6765\u7ed9\u5927\u5bb6\u770b\u770b\uff0c\u4e5f\u8bb8\u6709\u4eba\u771f\u9047\u5230\u8fc7\u8fd9\u4e2a\u75db\u70b9\uff0c\u4e5f\u8bb8\u5927\u5bb6\u5e2e\u6211\u786e\u8ba4\u8fd9\u5c31\u662f\u4e2a\u9524\u5b50\u627e\u9489\u5b50\u3002</p>\n<hr/>\n<h2>\u60f3\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898</h2>\n<p>\u4f01\u4e1a\u5185\u90e8\u901a\u5e38\u6709\u597d\u51e0\u4e2a\u6570\u636e\u5e93\uff1a\u9500\u552e\u7528 MySQL \u3001\u8d22\u52a1\u7528 PostgreSQL \u3001HR \u7528 SQL Server \u3002\u73b0\u5728\u8001\u677f\u8bf4\u8981\u63a5 LLM \u8ba9\u4e1a\u52a1\u4eba\u5458\u81ea\u7136\u8bed\u8a00\u67e5\u6570\u636e\u3002</p>\n<p>\u76f4\u63a5\u63a5\u4f1a\u9047\u5230\u8fd9\u4e9b\u95ee\u9898\uff1a</p>\n<table>\n<thead>\n<tr>\n<th>\u95ee\u9898</th>\n<th>\u4e3e\u4f8b</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>\u5b57\u6bb5\u540d\u65e0\u610f\u4e49</td>\n<td><code>aa</code>\u5b57\u6bb5\u662f\u5355\u4ef7\uff0c<code>hj</code>\u662f\u5408\u8ba1\uff0cLLM \u731c\u4e0d\u51fa\u6765</td>\n</tr>\n<tr>\n<td>\u540c\u540d\u4e0d\u540c\u4e49</td>\n<td>\u9500\u552e\u5e93\u7684\"\u91d1\u989d\"\u662f\u56de\u6b3e\uff0c\u8d22\u52a1\u5e93\u7684\"\u91d1\u989d\"\u662f\u5f00\u7968</td>\n</tr>\n<tr>\n<td>\u6743\u9650\u5931\u63a7</td>\n<td>\u9500\u552e\u5458\u80fd\u67e5\u5230\u6210\u672c\u548c\u5229\u6da6\u7387</td>\n</tr>\n<tr>\n<td>\u6ca1\u6709 SQL \u5ba1\u67e5</td>\n<td>LLM \u751f\u6210\u7684 SQL \u53ef\u80fd DROP TABLE</td>\n</tr>\n<tr>\n<td>\u654f\u611f\u6570\u636e\u88f8\u5954</td>\n<td>\u624b\u673a\u53f7\u8eab\u4efd\u8bc1\u660e\u6587\u8fd4\u56de</td>\n</tr>\n</tbody></table><p>\u6211\u7684\u60f3\u6cd5\u662f\u5728\u6570\u636e\u5e93\u548c LLM \u4e4b\u95f4\u52a0\u4e00\u5c42\uff0c\u628a\u8fd9\u4e9b\u810f\u6d3b\u81ea\u52a8\u5316\uff1a</p>\n<pre><code>\u4f01\u4e1a\u6570\u636e\u5e93\u7fa4\uff08 MySQL/PG/SQLite/Oracle/\u8fbe\u68a6\uff09\n        \u2193\n\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n\u2502         KaiwuBridge             \u2502\n\u2502  \u81ea\u52a8\u7406\u89e3\u5b57\u6bb5\u542b\u4e49\uff08\u4e0d\u7528\u4eba\u5de5\u6807\u6ce8\uff09  \u2502\n\u2502  \u6743\u9650\u63a7\u5236 + SQL \u5ba1\u67e5 + \u6570\u636e\u8131\u654f   \u2502\n\u2502  \u8de8\u5e93\u5b57\u6bb5\u81ea\u52a8\u5bf9\u9f50               \u2502\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n        \u2193\n\u4efb\u610f LLM \uff08\u672c\u5730 Ollama / DeepSeek / GPT \uff09\n</code></pre>\n<p>\u6838\u5fc3\u5356\u70b9\u662f<strong>\u4e0d\u7528\u4eba\u5de5\u6d17\u5e93</strong>\u2014\u2014\u4f20\u7edf\u505a\u6cd5\u662f DBA \u82b1\u51e0\u5468\u7ed9\u6bcf\u4e2a\u5b57\u6bb5\u5199\u6ce8\u91ca\u3001\u5efa\u6570\u636e\u5b57\u5178\uff0c\u6211\u60f3\u7528 LLM+\u7edf\u8ba1\u65b9\u6cd5\u81ea\u52a8\u641e\u5b9a\u3002</p>\n<hr/>\n<h2>\u5b9e\u73b0\u4e86\u4ec0\u4e48</h2>\n<h3>1. \u81ea\u52a8\u7406\u89e3\u5b57\u6bb5\u542b\u4e49\uff08\u56fe\u4f20\u64ad\u65b9\u6848\uff09</h3>\n<p>\u4e0d\u662f\u7b80\u5355\u8ba9 LLM \u770b\u5b57\u6bb5\u540d\u731c\u542b\u4e49\uff0c\u800c\u662f\uff1a</p>\n<ol>\n<li><strong>\u6570\u636e\u753b\u50cf</strong>\uff1a\u7edf\u8ba1\u6bcf\u4e2a\u5b57\u6bb5\u7684\u5206\u5e03\u3001\u7a7a\u503c\u7387\u3001\u552f\u4e00\u503c\u6bd4\u4f8b</li>\n<li><strong>\u4ee3\u6570\u5173\u7cfb\u68c0\u6d4b</strong>\uff1a\u81ea\u52a8\u53d1\u73b0 <code>\u5355\u4ef7 \u00d7 \u6570\u91cf \u2248 \u5408\u8ba1</code> \u8fd9\u79cd\u5173\u7cfb</li>\n<li><strong>\u5efa\u56fe</strong>\uff1a\u628a\u5b57\u6bb5\u3001\u5916\u952e\u3001\u4ee3\u6570\u5173\u7cfb\u5efa\u6210\u4e00\u5f20\u4f9d\u8d56\u56fe</li>\n<li><strong>\u56fe\u4f20\u64ad</strong>\uff1aLLM \u5728\u56fe\u4e0a\u8fed\u4ee3 3-5 \u8f6e\uff0c\u6bcf\u8f6e\u770b\u90bb\u5c45\u5b57\u6bb5\u7684\u63cf\u8ff0\u6765\u4fee\u6b63\u81ea\u5df1\u7684\u7406\u89e3</li>\n</ol>\n<p>\u8fd9\u6837\u5373\u4f7f\u5b57\u6bb5\u540d\u662f<code>aa</code>\uff0c\u7cfb\u7edf\u4e5f\u80fd\u901a\u8fc7\"aa \u00d7 \u6574\u6570\u5b57\u6bb5 \u2248 hj\"\u63a8\u65ad\u51fa aa \u662f\u5355\u4ef7\u3002</p>\n<p>\u7075\u611f\u6765\u81ea 2026 \u5e74 3 \u6708\u7684 DBAutoDoc \u8bba\u6587\uff0c\u6838\u5fc3\u601d\u60f3\u662f schema \u7406\u89e3\u672c\u8d28\u4e0a\u662f\u56fe\u7ed3\u6784\u95ee\u9898\u3002</p>\n<h3>2. \u4e03\u5c42\u5b89\u5168\u9632\u7ebf</h3>\n<pre><code>\u7269\u7406\u5c42\uff08\u53ea\u8bfb\u8d26\u53f7\uff09\u2192 SQL \u767d\u540d\u5355\uff08\u53ea\u5141\u8bb8 SELECT \uff09\u2192 \u6ce8\u91ca\u7ed5\u8fc7\u9632\u62a4 \u2192\n\u5b57\u6bb5\u7ea7\u6743\u9650\uff08 LLM \u770b\u4e0d\u5230=\u67e5\u4e0d\u5230\uff09\u2192 \u884c\u7ea7\u8fc7\u6ee4 RLAC \uff08\u534e\u4e1c\u5458\u5de5\u53ea\u770b\u534e\u4e1c\u6570\u636e\uff09\u2192\n\u6570\u636e\u8131\u654f\uff08\u624b\u673a\u53f7\u81ea\u52a8\u6253\u7801\uff09\u2192 \u52a8\u6001\u8131\u654f\uff08\u6309\u89d2\u8272\u8fd4\u56de\u4e0d\u540c\u7cbe\u5ea6\uff09\n</code></pre>\n<h3>3. \u89e3\u8026\u67b6\u6784\uff08\u4e09\u4e2a\u63a5\u53e3\uff09</h3>\n<pre><code>GET  /v1/context  \u2014 Agent \u83b7\u53d6 schema+\u6743\u9650+\u6620\u5c04+\u6b67\u4e49\u4fe1\u53f7\nPOST /v1/execute  \u2014 Agent \u63d0\u4ea4 SQL \uff0c\u4e2d\u95f4\u5c42\u8d1f\u8d23\u5b89\u5168\u68c0\u67e5+\u6267\u884c+\u8131\u654f\nPOST /v1/chat/completions \u2014 OpenAI \u517c\u5bb9\u63a5\u53e3\uff08\u517c\u5bb9\u5c42\uff09\n</code></pre>\n<p>Agent \u5c42\u548c\u6570\u636e\u5c42\u5f7b\u5e95\u5206\u79bb\u3002Agent \u53ea\u7ba1\u751f\u6210 SQL \uff0c\u4e2d\u95f4\u5c42\u53ea\u7ba1\u5b89\u5168\u6267\u884c\u3002</p>\n<h3>4. \u8de8\u5e93\u5b57\u6bb5\u81ea\u52a8\u5bf9\u9f50</h3>\n<ul>\n<li>bge-m3 embedding + Wasserstein \u5206\u5e03\u8ddd\u79bb</li>\n<li>\u4e3b\u52a8\u5b66\u4e60\uff1a\u4f18\u5148\u63a8\u9001\u7f6e\u4fe1\u5ea6 0.6-0.8 \u7684\u6a21\u7cca\u6848\u4f8b\u7ed9\u4eba\u5ba1\u6838\uff08\u4fe1\u606f\u4ef7\u503c\u6700\u9ad8\uff09</li>\n<li>\u7528\u6237\u786e\u8ba4/\u62d2\u7edd\u540e\u81ea\u52a8\u63d0\u53d6\u89c4\u5219\uff0c\u4e0d\u662f\u8c03\u9608\u503c</li>\n</ul>\n<h3>5. \u544a\u8b66\u8fc7\u6ee4</h3>\n<p>\u540c\u4e00\u4e2a\u9519\u8bef\u77ed\u65f6\u95f4\u5185\u53cd\u590d\u51fa\u73b0\u4e14\u4ece\u672a\u6210\u529f \u2192 \u81ea\u52a8\u538b\u5236\uff0c\u4e0d\u6253\u6270\u7528\u6237\u3002\u7ba1\u7406\u5458\u53ef\u4ee5\u770b\u5230\"\u50f5\u5c38\u89c4\u5219\"\u5217\u8868\u3002</p>\n<h3>6. Schema Linking \uff08 LLM \u8def\u7531\uff09</h3>\n<p>\u4f01\u4e1a\u53ef\u80fd\u6709\u51e0\u5341\u5f20\u8868\u3001\u51e0\u767e\u4e2a\u5b57\u6bb5\uff0c\u4e0d\u53ef\u80fd\u5168\u585e\u7ed9 LLM \u3002\u9700\u8981\u6839\u636e\u7528\u6237\u95ee\u9898\u7cbe\u51c6\u5b9a\u4f4d\u5230\u76f8\u5173\u7684 2-3 \u5f20\u8868\u3002</p>\n<p>\u505a\u6cd5\u53c2\u8003\u4e86 SchemaGraphSQL \uff08 ACL ARR 2025 \uff09\uff1a</p>\n<ol>\n<li><strong>\u5efa\u56fe</strong>\uff1a\u628a\u6240\u6709\u8868\u4f5c\u4e3a\u8282\u70b9\uff0c\u5916\u952e\u5173\u7cfb+\u8de8\u5e93\u6620\u5c04\u4f5c\u4e3a\u8fb9</li>\n<li><strong>LLM \u5b9e\u4f53\u63d0\u53d6</strong>\uff1a\u4e00\u6b21\u8c03\u7528\u4ece\u95ee\u9898\u4e2d\u63d0\u53d6\u5173\u952e\u5b9e\u4f53\uff0c\u6620\u5c04\u5230\u76f8\u5173\u8868</li>\n<li><strong>BFS \u6269\u5c55</strong>\uff1a\u5728\u56fe\u4e0a\u4ece\u76f8\u5173\u8868\u51fa\u53d1\u8d70 2 \u8df3\uff0c\u628a JOIN \u9700\u8981\u7684\u5173\u8054\u8868\u4e5f\u5e26\u4e0a</li>\n<li><strong>\u7cbe\u9009\u5b50\u96c6</strong>\uff1a\u6700\u591a\u7ed9 LLM \u770b 5 \u5f20\u8868\u7684 schema \uff0c\u800c\u4e0d\u662f\u5168\u91cf\u51e0\u5341\u5f20</li>\n</ol>\n<p>\u8fd9\u6837 LLM \u751f\u6210 SQL \u65f6\u53ea\u770b\u5230\u7cbe\u9009\u7684\u3001\u548c\u95ee\u9898\u76f8\u5173\u7684\u8868\uff0c\u4e0d\u4f1a\u88ab\u65e0\u5173\u8868\u5e72\u6270\uff0c\u751f\u6210\u51c6\u786e\u7387\u663e\u8457\u63d0\u5347\u3002</p>\n<p>\u96f6\u6837\u672c\u3001\u4e0d\u9700\u8981 embedding \u6a21\u578b\u3001\u4e0d\u9700\u8981\u8bad\u7ec3\u3002\u4e00\u6b21 LLM \u8c03\u7528\u641e\u5b9a\u8def\u7531\u3002</p>\n<hr/>\n<h2>\u529f\u80fd\u5168\u666f\uff08\u7ecf\u8fc7\u51e0\u6b21\u8fed\u4ee3\u540e\u7684\u5f53\u524d\u72b6\u6001\uff09</h2>\n<p>\u4ece\u6700\u521d\u53ea\u6709\"\u8fde\u6570\u636e\u5e93+\u8c03 LLM\"\uff0c\u5230\u73b0\u5728\u585e\u4e86\u4e00\u5806\u529f\u80fd\u3002\u7528\u4e00\u5f20\u8868\u8bf4\u6e05\u695a\u6bcf\u4e2a\u6a21\u5757\u5e72\u4ec0\u4e48\uff1a</p>\n<table>\n<thead>\n<tr>\n<th>\u529f\u80fd\u6a21\u5757</th>\n<th>\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898</th>\n<th>\u4ec0\u4e48\u573a\u666f\u7528</th>\n<th>\u539f\u7406/\u6280\u672f</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>\u6570\u636e\u753b\u50cf</strong> (<a href=\"http://profiler.py\" rel=\"nofollow\">profiler.py</a>)</td>\n<td>\u5b57\u6bb5\u540d\u65e0\u610f\u4e49\u65f6\u65e0\u6cd5\u7406\u89e3\u6570\u636e</td>\n<td>scan \u65f6\u81ea\u52a8\u8fd0\u884c\uff0c\u7ed9\u6bcf\u4e2a\u5b57\u6bb5\u5efa\u7edf\u8ba1\u6863\u6848</td>\n<td>\u7a7a\u503c\u7387/\u552f\u4e00\u503c\u6bd4\u4f8b/\u6570\u503c\u5206\u5e03/\u9ad8\u9891\u503c\u91c7\u6837</td>\n</tr>\n<tr>\n<td><strong>\u4ee3\u6570\u5173\u7cfb\u68c0\u6d4b</strong> (<a href=\"http://profiler.py\" rel=\"nofollow\">profiler.py</a>)</td>\n<td><code>aa\u00d7bb\u2248cc</code>\u8fd9\u79cd\u9690\u542b\u4e1a\u52a1\u5173\u7cfb\u4eba\u770b\u4e0d\u51fa\u6765</td>\n<td>\u540c\u8868\u5185\u6570\u503c\u5b57\u6bb5\u4e09\u5143\u7ec4\u679a\u4e3e</td>\n<td>numpy \u5411\u91cf\u5316\u8ba1\u7b97\uff0c5%\u8bef\u5dee\u5bb9\u5fcd\u5ea6</td>\n</tr>\n<tr>\n<td><strong>\u56fe\u4f20\u64ad\u5f15\u64ce</strong> (<a href=\"http://graph_propagation.py\" rel=\"nofollow\">graph_propagation.py</a>)</td>\n<td>\u5355\u770b\u4e00\u4e2a\u5b57\u6bb5\u731c\u4e0d\u51fa\u542b\u4e49\uff0c\u9700\u8981\u4e0a\u4e0b\u6587</td>\n<td>scan --semantic \u65f6\u66ff\u4ee3\u9010\u5b57\u6bb5 LLM \u751f\u6210</td>\n<td>\u5efa\u4f9d\u8d56\u56fe\u2192LLM \u8fed\u4ee3 3-5 \u8f6e\u2192\u90bb\u5c45\u63cf\u8ff0\u4f5c\u4e3a context \u7cbe\u5316</td>\n</tr>\n<tr>\n<td><strong>Schema Linking \u8def\u7531</strong> (<a href=\"http://schema_graph.py\" rel=\"nofollow\">schema_graph.py</a>)</td>\n<td>\u51e0\u5341\u5f20\u8868\u4e0d\u80fd\u5168\u585e\u7ed9 LLM</td>\n<td>\u6bcf\u6b21\u7528\u6237\u63d0\u95ee\u65f6\u81ea\u52a8\u89e6\u53d1</td>\n<td>\u5916\u952e\u56fe+LLM \u5b9e\u4f53\u63d0\u53d6+BFS 2 \u8df3\u6269\u5c55\uff0c\u7cbe\u9009\u22645 \u5f20\u8868</td>\n</tr>\n<tr>\n<td><strong>\u8de8\u5e93\u8bed\u4e49\u5339\u914d</strong> (<a href=\"http://matching.py\" rel=\"nofollow\">matching.py</a>)</td>\n<td>\u4e0d\u540c\u5e93\u7684\"\u91d1\u989d\"\u53ef\u80fd\u662f\u4e0d\u540c\u6982\u5ff5</td>\n<td>scan \u540e\u81ea\u52a8\u4e24\u4e24\u5339\u914d\uff0c\u751f\u6210 pending \u6620\u5c04</td>\n<td>bge-m3 embedding + Wasserstein \u5206\u5e03\u8ddd\u79bb</td>\n</tr>\n<tr>\n<td><strong>\u4e3b\u52a8\u5b66\u4e60</strong> (<a href=\"http://matching.py\" rel=\"nofollow\">matching.py</a> RuleExtractor)</td>\n<td>\u4eba\u5de5\u5ba1\u6838\u6548\u7387\u4f4e\uff0c\u4e0d\u77e5\u9053\u5148\u5ba1\u54ea\u4e2a</td>\n<td>\u7ba1\u7406\u754c\u9762\u5c55\u793a\u5f85\u5ba1\u6838\u6620\u5c04\u65f6\u6392\u5e8f</td>\n<td>\u4f18\u5148\u63a8\u9001\u7f6e\u4fe1\u5ea6 0.6-0.8 \u7684\u6848\u4f8b\uff08\u4fe1\u606f\u4ef7\u503c\u6700\u9ad8\uff09</td>\n</tr>\n<tr>\n<td><strong>SQL \u767d\u540d\u5355\u5ba1\u67e5</strong> (<a href=\"http://security.py\" rel=\"nofollow\">security.py</a>)</td>\n<td>LLM \u53ef\u80fd\u751f\u6210 DROP TABLE</td>\n<td>\u6bcf\u6b21\u6267\u884c SQL \u524d\u5f3a\u5236\u68c0\u67e5</td>\n<td>sqlparse \u8bed\u6cd5\u6811\u5206\u6790\uff0c\u53ea\u653e\u884c SELECT/WITH</td>\n</tr>\n<tr>\n<td><strong>\u5b57\u6bb5\u7ea7\u6743\u9650</strong> (<a href=\"http://permissions.py\" rel=\"nofollow\">permissions.py</a>)</td>\n<td>\u9500\u552e\u5458\u4e0d\u8be5\u770b\u5230\u6210\u672c\u5b57\u6bb5</td>\n<td>schema \u53d1\u7ed9 LLM \u524d\u8fc7\u6ee4</td>\n<td>\u914d\u7f6e denied_columns \uff0c\u7269\u7406\u79fb\u9664\u5b57\u6bb5</td>\n</tr>\n<tr>\n<td><strong>\u884c\u7ea7\u8fc7\u6ee4 RLAC</strong> (<a href=\"http://executor.py\" rel=\"nofollow\">executor.py</a>)</td>\n<td>\u534e\u4e1c\u5458\u5de5\u53ea\u80fd\u770b\u534e\u4e1c\u6570\u636e</td>\n<td>SQL \u6267\u884c\u65f6 CTE \u5b50\u67e5\u8be2\u5305\u88c5\u6ce8\u5165 WHERE</td>\n<td>\u4e0d\u4f9d\u8d56 LLM\"\u81ea\u89c9\"\uff0c\u6267\u884c\u5c42\u5f3a\u5236\u6ce8\u5165</td>\n</tr>\n<tr>\n<td><strong>\u6570\u636e\u8131\u654f</strong> (<a href=\"http://security.py\" rel=\"nofollow\">security.py</a> + <a href=\"http://executor.py\" rel=\"nofollow\">executor.py</a>)</td>\n<td>\u624b\u673a\u53f7\u8eab\u4efd\u8bc1\u4e0d\u80fd\u660e\u6587\u8fd4\u56de</td>\n<td>\u7ed3\u679c\u8fd4\u56de\u524d\u81ea\u52a8\u5904\u7406</td>\n<td>\u6b63\u5219\u6253\u7801 + \u6309\u89d2\u8272\u52a8\u6001\u7cbe\u5ea6\uff08 full/partial/round \uff09</td>\n</tr>\n<tr>\n<td><strong>\u544a\u8b66\u8fc7\u6ee4</strong> (<a href=\"http://alert_filter.py\" rel=\"nofollow\">alert_filter.py</a>)</td>\n<td>\u540c\u4e00\u4e2a\u9519\u8bef\u53cd\u590d\u5f39\u51fa\u70e6\u6b7b\u4eba</td>\n<td>\u517c\u5bb9\u5c42\u6267\u884c\u5931\u8d25\u65f6\u5224\u65ad</td>\n<td>\u6ed1\u52a8\u7a97\u53e3\u9891\u7387\u7edf\u8ba1\uff0c\u22655 \u6b21\u4e14 0 \u6210\u529f\u2192\u538b\u5236</td>\n</tr>\n<tr>\n<td><strong>\u6b67\u4e49\u68c0\u6d4b</strong> (<a href=\"http://server.py\" rel=\"nofollow\">server.py</a>)</td>\n<td>\"\u9500\u552e\u989d\"\u5728\u4e24\u4e2a\u5e93\u90fd\u6709\uff0c\u7528\u54ea\u4e2a\uff1f</td>\n<td>/v1/context \u63a5\u53e3\u8fd4\u56de\u6b67\u4e49\u4fe1\u53f7</td>\n<td>\u8bed\u4e49\u540d\u7247\u5339\u914d+\u591a\u5e93\u6765\u6e90\u68c0\u6d4b\uff0c\u542b confidence</td>\n</tr>\n<tr>\n<td><strong>\u6570\u636e\u65b0\u9c9c\u5ea6</strong> (<a href=\"http://executor.py\" rel=\"nofollow\">executor.py</a>)</td>\n<td>\u67e5\u5230\u7684\u6570\u636e\u53ef\u80fd\u662f\u4e0a\u5468\u7684</td>\n<td>\u6267\u884c\u6210\u529f\u540e\u9644\u52a0\u63d0\u793a</td>\n<td>\u67e5 MAX(updated_at)\uff0c\u8d85 24 \u5c0f\u65f6\u8b66\u544a</td>\n</tr>\n<tr>\n<td><strong>\u6620\u5c04\u5bfc\u5165\u5bfc\u51fa</strong> (<a href=\"http://admin.py\" rel=\"nofollow\">admin.py</a>)</td>\n<td>DBA \u60f3\u5728 Excel \u91cc\u6279\u91cf\u7ef4\u62a4\u6620\u5c04\u5173\u7cfb</td>\n<td>\u7ba1\u7406\u540e\u53f0 CSV \u4e0a\u4f20\u4e0b\u8f7d</td>\n<td>CSV \u89e3\u6790 + LLM \u9a8c\u8bc1\u5c42\uff08\u68c0\u67e5\u660e\u663e\u9519\u8bef\uff09</td>\n</tr>\n<tr>\n<td><strong>\u6301\u7eed\u5b66\u4e60</strong> (<a href=\"http://admin.py\" rel=\"nofollow\">admin.py</a> + <a href=\"http://matching.py\" rel=\"nofollow\">matching.py</a>)</td>\n<td>\u7528\u6237\u53cd\u9988\u5e94\u8be5\u8ba9\u7cfb\u7edf\u8d8a\u6765\u8d8a\u51c6</td>\n<td>confirm/reject \u6620\u5c04\u65f6\u81ea\u52a8\u89e6\u53d1</td>\n<td>\u8d1d\u53f6\u65af\u66f4\u65b0\u9608\u503c + \u89c4\u5219\u63d0\u53d6\uff08\u4e0d\u53ea\u662f\u8c03\u53c2\uff09</td>\n</tr>\n<tr>\n<td><strong>\u89e3\u8026\u63a5\u53e3</strong> (<a href=\"http://server.py\" rel=\"nofollow\">server.py</a>)</td>\n<td>Agent \u5c42\u548c\u6570\u636e\u5c42\u8026\u5408\u5728\u4e00\u8d77\u4e0d\u597d\u6269\u5c55</td>\n<td>Agent \u81ea\u5df1\u751f\u6210 SQL \u65f6\u7528 context+execute</td>\n<td>REST \u5206\u79bb\uff1acontext \u53ea\u7ed9\u6570\u636e\uff0cexecute \u53ea\u7ba1\u6267\u884c</td>\n</tr>\n</tbody></table><p>\u4e00\u5171 22 \u4e2a Python \u6a21\u5757\uff0c7015 \u884c\u4ee3\u7801\u3002\u8bf4\u5b9e\u8bdd\u5199\u5230\u540e\u9762\u81ea\u5df1\u90fd\u89c9\u5f97\u529f\u80fd\u5806\u592a\u591a\u4e86\u3002</p>\n<hr/>\n<h2>\u6d4b\u8bd5\u548c\u7ed3\u679c</h2>\n<h3>\u4ee3\u6570\u5173\u7cfb\u68c0\u6d4b</h3>\n<p>\u7528 100 \u884c\u6a21\u62df\u8ba2\u5355\u6570\u636e\u6d4b\u8bd5\uff1a</p>\n<ul>\n<li>\u53ec\u56de\u7387\uff1a100%\uff08 2/2 \u4e2a\u6807\u6ce8\u5173\u7cfb\u5168\u90e8\u68c0\u6d4b\u5230\uff09</li>\n<li>\u8bef\u62a5\u7387\uff1a0%\uff08\u7f16\u7801\u5b57\u6bb5\u6ca1\u6709\u88ab\u8bef\u5224\u4e3a\u4ee3\u6570\u5173\u7cfb\uff09</li>\n</ul>\n<h3>\u8bed\u4e49\u5339\u914d\u57fa\u7ebf\uff08\u8bda\u5b9e\u62a5\u544a\uff09</h3>\n<p>\u7528 10 \u5bf9\u624b\u5de5\u6807\u6ce8\u7684\u8de8\u5e93\u5b57\u6bb5\u5bf9\u6d4b\u8bd5\uff1a</p>\n<ul>\n<li>**\u8d1f\u4f8b\u62d2\u7edd\u7387\uff1a100%**\uff08\u4e0d\u76f8\u5173\u5b57\u6bb5\u4e0d\u4f1a\u88ab\u8bef\u5339\u914d\uff09</li>\n<li>**\u6b63\u4f8b\u53ec\u56de\u7387\uff1a0%**\uff08\u88f8\u82f1\u6587\u5b57\u6bb5\u540d\u5728 bge-m3 \u4e0a\u8bed\u4e49\u5206\u5168\u90e8\u4f4e\u4e8e\u9608\u503c\uff09</li>\n</ul>\n<p>\u8fd9\u4e2a 0%\u662f\u9884\u671f\u7684\u2014\u2014\u8bc1\u660e\u4e86\u56fe\u4f20\u64ad\u5c42\u7684\u5fc5\u8981\u6027\u3002\u88f8\u5b57\u6bb5\u540d<code>sales_amount</code>\u548c<code>revenue</code>\u7684 embedding \u76f8\u4f3c\u5ea6\u53ea\u6709 0.67 \uff0c\u4f4e\u4e8e 0.85 \u9608\u503c\u3002\u9700\u8981\u56fe\u4f20\u64ad\u5148\u751f\u6210\u4e2d\u6587\u63cf\u8ff0\uff08\"\u6bcf\u7b14\u8ba2\u5355\u7684\u542b\u7a0e\u9500\u552e\u91d1\u989d\"\uff09\uff0c\u518d\u505a\u5339\u914d\u624d\u6709\u610f\u4e49\u3002</p>\n<p><strong>\u4f46\u6211\u8fd8\u6ca1\u6709\u5728\u771f\u5b9e\u6570\u636e\u5e93\u4e0a\u8dd1\u8fc7\u5b8c\u6574\u6d41\u6c34\u7ebf\u3002</strong></p>\n<h3>\u5b89\u5168\u6d4b\u8bd5</h3>\n<p>65 \u4e2a\u5b89\u5168\u6d4b\u8bd5\u8986\u76d6\uff1aSQL \u6ce8\u5165\uff08\u542b\u6ce8\u91ca\u7ed5\u8fc7\uff09\u3001JWT \u4f2a\u9020\u3001\u8d8a\u6743\u8bbf\u95ee\u3001\u9891\u7387\u9650\u5236\u3001\u6570\u636e\u8131\u654f\u3002\u5168\u90e8\u901a\u8fc7\u3002</p>\n<h3>\u603b\u8ba1</h3>\n<pre><code>134 passed, 0 failed, 21 warnings\n</code></pre>\n<hr/>\n<h2>\u6280\u672f\u6808</h2>\n<ul>\n<li>Python 3.12 + FastAPI + SQLAlchemy 2.0</li>\n<li>sentence-transformers (bge-m3) \u505a embedding</li>\n<li>numpy/scipy \u505a\u7edf\u8ba1\u9a8c\u8bc1</li>\n<li>SQLite \u5b58\u5143\u6570\u636e\uff08\u96f6\u90e8\u7f72\uff09</li>\n<li>\u652f\u6301 MySQL / PostgreSQL / SQLite / SQL Server / Oracle / \u8fbe\u68a6 / \u4eba\u5927\u91d1\u4ed3</li>\n</ul>\n<p>\u5168\u90e8\u4f9d\u8d56 Apache 2.0 / MIT / BSD \uff0c\u53ef\u5546\u7528\u3002</p>\n<hr/>\n<h2>\u4e3a\u4ec0\u4e48\u8bf4\u662f\u9524\u5b50\u627e\u9489\u5b50</h2>\n<p>\u5199\u5b8c\u4e4b\u540e\u51b7\u9759\u4e0b\u6765\u60f3\u4e86\u51e0\u4e2a\u95ee\u9898\uff1a</p>\n<p><strong>1. \u8c01\u662f\u7528\u6237\uff1f</strong></p>\n<p>\u6211\u5047\u60f3\u7684\u573a\u666f\u662f\"\u4e2d\u578b\u4f01\u4e1a\uff0c\u6709 3-5 \u4e2a\u4e1a\u52a1\u6570\u636e\u5e93\uff0c\u60f3\u8ba9\u4e1a\u52a1\u4eba\u5458\u81ea\u7136\u8bed\u8a00\u67e5\u6570\u636e\"\u3002\u4f46\u6211\u6ca1\u6709\u627e\u5230\u4e00\u4e2a\u5177\u4f53\u7684\u4f01\u4e1a\u8bf4\"\u6211\u9700\u8981\u8fd9\u4e2a\"\u3002</p>\n<p><strong>2. \u771f\u5b9e\u573a\u666f\u4e0b\u8fd9\u4e2a\u95ee\u9898\u5b58\u5728\u5417\uff1f</strong></p>\n<p>\u4e5f\u8bb8\u5b58\u5728\uff0c\u4f46\u89e3\u51b3\u65b9\u6848\u53ef\u80fd\u4e0d\u662f\u6211\u60f3\u7684\u8fd9\u6837\uff1a</p>\n<ul>\n<li>\u5927\u4f01\u4e1a\u6709\u6570\u636e\u4e2d\u53f0\u56e2\u961f\uff0c\u4eba\u5de5\u5efa\u6570\u636e\u5b57\u5178\u4e0d\u662f\u95ee\u9898</li>\n<li>\u5c0f\u4f01\u4e1a\u53ef\u80fd\u5c31\u4e00\u4e2a MySQL \uff0c\u4e0d\u9700\u8981\u8de8\u5e93\u5bf9\u9f50</li>\n<li>\u4e2d\u578b\u4f01\u4e1a\u53ef\u80fd\u66f4\u9700\u8981\u7684\u662f BI \u5de5\u5177\u800c\u4e0d\u662f\u81ea\u7136\u8bed\u8a00\u67e5\u8be2</li>\n</ul>\n<p><strong>3. \"\u4e0d\u7528\u4eba\u5de5\u6d17\u5e93\"\u8fd9\u4e2a\u5356\u70b9\u6210\u7acb\u5417\uff1f</strong></p>\n<p>\u56fe\u4f20\u64ad\u65b9\u6848\u7406\u8bba\u4e0a\u80fd\u81ea\u52a8\u7406\u89e3\u5b57\u6bb5\u542b\u4e49\uff0c\u4f46\uff1a</p>\n<ul>\n<li>\u9700\u8981 LLM \uff08\u672c\u5730 7B \u6a21\u578b\u591f\u4e0d\u591f\uff1f\u9700\u8981 API \u8c03\u7528\uff1f\uff09</li>\n<li>\u51c6\u786e\u7387\u672a\u5728\u771f\u5b9e\u810f\u6570\u636e\u4e0a\u9a8c\u8bc1</li>\n<li>\u4f01\u4e1a\u53ef\u80fd\u5b81\u613f\u82b1\u4e00\u5468\u4eba\u5de5\u6807\u6ce8\u4e5f\u4e0d\u613f\u610f\u4fe1\u4efb\u81ea\u52a8\u5316\u7ed3\u679c</li>\n</ul>\n<p><strong>4. \u8fc7\u5ea6\u5de5\u7a0b\u4e86\u5417\uff1f</strong></p>\n<p>7000 \u884c\u4ee3\u7801\u3001\u56fe\u4f20\u64ad\u3001\u4e3b\u52a8\u5b66\u4e60\u3001\u544a\u8b66\u8fc7\u6ee4\u3001\u52a8\u6001\u8131\u654f\u2026\u2026\u5982\u679c\u7b2c\u4e00\u4e2a\u7528\u6237\u53ea\u9700\u8981\"\u8fde MySQL + \u6743\u9650\u63a7\u5236 + \u8c03 DeepSeek\"\uff0c\u90a3 90%\u7684\u4ee3\u7801\u90fd\u662f\u63d0\u524d\u4f18\u5316\u3002</p>\n<hr/>\n<h2>\u5982\u679c\u4f60\u9047\u5230\u8fc7\u8fd9\u4e2a\u95ee\u9898</h2>\n<p>\u60f3\u542c\u542c\u5927\u5bb6\u7684\u770b\u6cd5\uff1a</p>\n<ol>\n<li>\u662f\u6211\u60f3\u7684\u8fd9\u4e48\u7b80\u5355\u4e48\u6570\u5b57\u5316\u843d\u5730?LLM + \u4f18\u5316\u5c42 \u8ba1\u5165\u6570\u636e\u5e93\uff0c\u5c31 AI \u843d\u5730\u4e48\uff1f </li>\n<li>\u771f\u5b9e\u4f01\u4e1a\u6570\u5b57\u5316\u843d\u5730\u6700\u96be\u653b\u514b\u4ec0\u4e48\uff1f</li>\n<li>\u8fd9\u4e2a\u65b9\u5411\u503c\u5f97\u7ee7\u7eed\u505a\u5417\uff1f\u8fd8\u662f\u5e94\u8be5 pivot \u6210\u66f4\u5177\u4f53\u7684\u4e1c\u897f\uff08\u6bd4\u5982\u53ea\u505a SQL \u5b89\u5168\u5ba1\u67e5\u5c42\uff09\uff1f</li>\n</ol>\n<p>\u4ee3\u7801\u5728\u672c\u5730\uff0c\u5982\u679c\u6709\u4eba\u611f\u5174\u8da3\u53ef\u4ee5\u5f00\u6e90\u3002\u4e5f\u6b22\u8fce\u76f4\u63a5\u544a\u8bc9\u6211\u8fd9\u662f\u4e2a\u4f2a\u9700\u6c42\uff0c\u7701\u5f97\u6211\u7ee7\u7eed\u5f80\u91cc\u9762\u6295\u65f6\u95f4\u3002</p>\n<hr/>\n<h2>\u53c2\u8003\u7684\u8bba\u6587\u548c\u5f00\u6e90\u9879\u76ee</h2>\n<table>\n<thead>\n<tr>\n<th>\u6765\u6e90</th>\n<th>\u7528\u5728\u54ea</th>\n<th>\u600e\u4e48\u7528\u7684</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><a href=\"https://arxiv.org/abs/2501.xxxxx\" rel=\"nofollow\">SchemaGraphSQL</a> (ACL ARR 2025)</td>\n<td>Schema Linking \u8def\u7531</td>\n<td>\u6838\u5fc3\u601d\u60f3\uff1a\u7528\u5916\u952e\u5173\u7cfb\u56fe+LLM \u5b9e\u4f53\u63d0\u53d6+BFS \u8def\u5f84\u641c\u7d22\u505a schema linking \uff0c\u96f6\u6837\u672c\u4e0d\u9700\u8981\u8bad\u7ec3\u3002\u6211\u76f4\u63a5\u5b9e\u73b0\u4e86\u8fd9\u4e2a\u65b9\u6848</td>\n</tr>\n<tr>\n<td><a href=\"https://arxiv.org/abs/2603.23050\" rel=\"nofollow\">DBAutoDoc</a> (2026.03)</td>\n<td>\u56fe\u4f20\u64ad\u5f15\u64ce</td>\n<td>\u6838\u5fc3\u601d\u60f3\uff1aschema \u7406\u89e3\u662f\u56fe\u7ed3\u6784\u95ee\u9898\uff0c\u901a\u8fc7\u4f9d\u8d56\u56fe\u8fed\u4ee3\u4f20\u64ad\u8bed\u4e49\u4fee\u6b63\u76f4\u5230\u6536\u655b\u3002\u6211\u7b80\u5316\u4e86\u5b9e\u73b0\uff0c\u6ca1\u7528\u539f\u6587\u7684 GNN \uff0c\u76f4\u63a5 LLM \u8fed\u4ee3</td>\n</tr>\n<tr>\n<td><a href=\"https://arxiv.org/abs/2504.xxxxx\" rel=\"nofollow\">LLM-FK</a> (2025)</td>\n<td>\u5916\u952e\u53d1\u73b0\u601d\u8def</td>\n<td>\u4e09 agent \u534f\u4f5c\uff08 Interpreter/Refiner/Verifier \uff09\u7684\u601d\u8def\u542f\u53d1\u4e86\u6211\u7684\u7ea6\u675f\u53d1\u73b0\u8bbe\u8ba1\uff0c\u4f46\u6211\u6ca1\u5b9e\u73b0\u591a agent \uff0c\u53ea\u7528\u4e86\u7edf\u8ba1\u65b9\u6cd5</td>\n</tr>\n<tr>\n<td><a href=\"https://github.com/delftdata/valentine\" rel=\"nofollow\">Valentine</a></td>\n<td>\u8de8\u5e93\u5339\u914d baseline</td>\n<td>schema matching \u7684\u5f00\u6e90 benchmark \uff0c\u53c2\u8003\u4e86\u5b83\u7684\u8bc4\u4f30\u65b9\u6cd5\u8bba\uff08 precision/recall on labeled pairs \uff09</td>\n</tr>\n<tr>\n<td><a href=\"https://arxiv.org/abs/2012.xxxxx\" rel=\"nofollow\">ALITE</a></td>\n<td>\u7ea6\u675f\u53d1\u73b0</td>\n<td>\u7528\u6570\u636e\u5206\u6790\u53d1\u73b0\u51fd\u6570\u4f9d\u8d56\u548c\u5305\u542b\u4f9d\u8d56\u7684\u601d\u8def\uff0c\u6211\u7b80\u5316\u6210\u4e86\u4ee3\u6570\u5173\u7cfb\u68c0\u6d4b\uff08 A\u00d7B\u2248C \uff09</td>\n</tr>\n<tr>\n<td><a href=\"https://github.com/UKPLab/sentence-transformers\" rel=\"nofollow\">sentence-transformers</a></td>\n<td>embedding \u8ba1\u7b97</td>\n<td>\u76f4\u63a5\u7528\u7684 bge-m3 \u6a21\u578b\u505a\u5b57\u6bb5\u8bed\u4e49\u5411\u91cf\u5316</td>\n</tr>\n<tr>\n<td><a href=\"https://github.com/tiangolo/fastapi\" rel=\"nofollow\">FastAPI</a></td>\n<td>Web \u6846\u67b6</td>\n<td>OpenAI \u517c\u5bb9\u63a5\u53e3</td>\n</tr>\n<tr>\n<td><a href=\"https://github.com/sqlalchemy/sqlalchemy\" rel=\"nofollow\">SQLAlchemy</a></td>\n<td>\u6570\u636e\u5e93\u8fde\u63a5</td>\n<td>\u591a\u6570\u636e\u5e93\u7edf\u4e00\u9002\u914d\u5c42</td>\n</tr>\n<tr>\n<td><a href=\"https://github.com/andialbrecht/sqlparse\" rel=\"nofollow\">sqlparse</a></td>\n<td>SQL \u5b89\u5168\u5ba1\u67e5</td>\n<td>\u8bed\u6cd5\u6811\u5206\u6790\uff0c\u767d\u540d\u5355\u9a8c\u8bc1\uff0c\u8868\u540d\u63d0\u53d6</td>\n</tr>\n</tbody></table><p>\u90e8\u5206\u8bba\u6587 ai \u641c\u7684\uff0c\uff0c\uff0c\uff0c\n\u8bf4\u5b9e\u8bdd\uff0c\u8bba\u6587\u8bfb\u4e86\u4e0d\u5c11\uff0c\u4f46\u771f\u6b63\u843d\u5730\u65f6\u5927\u5e45\u7b80\u5316\u4e86\u3002DBAutoDoc \u539f\u6587\u7528\u7684\u662f GNN \u505a\u56fe\u4f20\u64ad\uff0c\u6211\u76f4\u63a5\u7528 LLM \u8fed\u4ee3\u66ff\u4ee3\u4e86\uff08\u56e0\u4e3a\u76ee\u6807\u573a\u666f\u662f\u4f01\u4e1a\u5185\u90e8\u51e0\u5341\u5f20\u8868\uff0c\u4e0d\u662f\u51e0\u5343\u5f20\u8868\u7684\u5b66\u672f benchmark \uff0cLLM \u8fed\u4ee3 3-5 \u8f6e\u5b8c\u5168\u591f\u7528\uff09\u3002</p>\n<hr/>\n<p><em>\u6280\u672f\u7ec6\u8282\uff1aPython 3.12 / FastAPI / SQLAlchemy / bge-m3 / \u56fe\u4f20\u64ad\u67b6\u6784 / 134 \u6d4b\u8bd5\u5168\u7eff</em></p>\n<p>\u9644\u4ed3\u5e93\uff08\u4e3a\u4e86\u907f\u514d\u8bf4\u63a8\u5e7f\u4ed3\u5e93\u7684\uff0c\u6240\u4ee5\u653e\u6700\u540e\uff09\uff1a\n<a href=\"https://github.com/val1813/kwb\" rel=\"nofollow\">https://github.com/val1813/kwb</a></p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/Livid", 
        "name": "Livid", 
        "avatar": "https://cdn.v2ex.com/avatar/c4ca/4238/1_large.png?m=1781025867"
      }, 
      "url": "https://www.v2ex.com/t/1210644", 
      "date_modified": "2026-05-06T10:39:09+00:00", 
      "content_html": "<a target=\"_blank\" href=\"https://ollama.com/library/gemma4:31b-coding-mtp-bf16\" rel=\"nofollow noopener\">https://ollama.com/library/gemma4:31b-coding-mtp-bf16</a><br /><br />\u672c\u5730\u90e8\u7f72\u7684\u65f6\u5019\uff0cBest Practices \u90e8\u5206\u6709\u4e00\u4e9b\u6709\u7528\u4fe1\u606f\u3002", 
      "date_published": "2026-05-06T10:38:59+00:00", 
      "title": "gemma4:31b-coding-mtp-bf16", 
      "id": "https://www.v2ex.com/t/1210644"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/linxiaojialin", 
        "name": "linxiaojialin", 
        "avatar": "https://cdn.v2ex.com/avatar/e295/16ba/288578_large.png?m=1533403565"
      }, 
      "url": "https://www.v2ex.com/t/1210478", 
      "date_modified": "2026-05-09T05:34:03+00:00", 
      "content_html": "<p>\u60f3\u6362\u4e2a\u65b0\u7684\u53f0\u5f0f\u7535\u8111\uff0c\u5e73\u5e38\u5f00\u53d1\u7528\u7684\uff0c\u5076\u5c14\u6253\u6e38\u620f\uff0c\u60f3\u7a7a\u95f2\u65f6\u95f4\u80fd\u591f\u672c\u5730\u8bad\u7ec3 AI \uff0c\u8bf7\u95ee\u5927\u5bb6\u6709\u63a8\u8350\u7684\u914d\u7f6e\u6e05\u5355\u5417\uff1f\u6709\u6807\u8bb0\u4ef7\u683c\u5c31\u66f4\u597d\u4e86</p>\n", 
      "date_published": "2026-05-06T03:23:32+00:00", 
      "title": "\u6709\u9002\u5408\u672c\u5730\u8dd1\u8bad\u7ec3 AI \u7684\u7535\u8111\u914d\u7f6e\u5417\uff1f", 
      "id": "https://www.v2ex.com/t/1210478"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/babymonster", 
        "name": "babymonster", 
        "avatar": "https://cdn.v2ex.com/avatar/a37b/761f/743705_large.png?m=1777516327"
      }, 
      "url": "https://www.v2ex.com/t/1210410", 
      "title": "\u90fd 2026 \u5e74\u4e86\uff0c\u4e3a\u4ec0\u4e48\u8fd8\u6709\u4eba\u89c9\u5f97 AMD \u6bd4 Nvidia \u66f4\u9002\u5408\u90e8\u7f72\u672c\u5730\u5927\u6a21\u578b\uff1f", 
      "id": "https://www.v2ex.com/t/1210410", 
      "date_published": "2026-05-06T01:16:58+00:00", 
      "content_html": "\u4e94\u4e00\u8282\u5047\u65e5\u671f\u95f4\uff0c\u6709\u4e00\u4e2a\u540c\u4e8b\u60f3\u672c\u5730\u90e8\u7f72\u5927\u6a21\u578b\uff0c\u5728\u7fa4\u91cc\u8be2\u95ee\uff0c\u6211\u4eec\u90fd\u7ed9\u4ed6\u63a8\u8350\u4e86 Nvidia \u7684\u5361\uff0c\u7ed3\u679c\u4ed6\u53bb\u5237 B \u7ad9\uff0c\u9009\u62e9\u4e86 AMD AI MAX+ 395 \u3002<br /><br />\u96be\u9053\u5c0f\u767d\u771f\u7684\u5c31\u8fd9\u6837\u88ab\u5272\u97ed\u83dc\u5417\uff1f"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/zsj1029", 
        "name": "zsj1029", 
        "avatar": "https://cdn.v2ex.com/avatar/38a0/9f4d/189266_large.png?m=1777977403"
      }, 
      "url": "https://www.v2ex.com/t/1210290", 
      "title": "LiteChat \u8f7b\u91cf\u7ea7\u672c\u5730\u5927\u6a21\u578b\u804a\u5929 WebUI\uff0c\u652f\u6301 vLLM", 
      "id": "https://www.v2ex.com/t/1210290", 
      "date_published": "2026-05-05T05:26:23+00:00", 
      "content_html": "<a target=\"_blank\" href=\"https://github.com/zsj1029/LiteChat\" rel=\"nofollow noopener\">https://github.com/zsj1029/LiteChat</a><br /><br />\u4f01\u4e1a\u5185\u90e8\u573a\u666f\u9002\u7528\uff0c\u4ece llama-cpp \u7684 webui \u6252\u51fa\u6765\u7684\uff0c\u672c\u5730\u6539\u9020\u4e86\u4e0b\u652f\u6301 vllm<br /><br />\u5168\u7a0b Qwen3.6 27B (vLLM), Claude Vscode \u6539\u9020"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/SzgSw5zGyN1iy", 
        "name": "SzgSw5zGyN1iy", 
        "avatar": "https://cdn.v2ex.com/gravatar/20a36637a8fa71a4f8095b77e39308ed?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1210145", 
      "title": "DGX Spark\u3001ASUS GX10\u3001MSI EdgeXpert \u770b\u8d77\u6765\u90fd\u50cf\u662f\u4e00\u4e2a\u6bcd\u80ce\u7684\u4ea7\u54c1\uff0c\u7528\u8d77\u6765\u6709\u5dee\u522b\u5417\uff1f", 
      "id": "https://www.v2ex.com/t/1210145", 
      "date_published": "2026-05-03T15:15:04+00:00", 
      "content_html": "<p>Spark \u6700\u8d35\uff0c\u548c\u540e\u4e24\u8005\u4ef7\u94b1\u6709\u70b9\u513f\u533a\u522b\uff0c\u4e0d\u8fc7\u770b\u8d77\u6765\u90fd\u50cf\u662f\u540c\u4e00\u4e2a\u65b9\u6848\uff0c\u53ea\u662f\u516c\u7248\u548c\u5404\u5bb6\u81ea\u5df1\u724c\u5b50\u7684\u533a\u522b\u800c\u5df2\uff1f</p>\n<ul>\n<li>Spark \u81ea\u5e26\u5f00\u7bb1\u5373\u7528\u7684\u5de5\u5177\u96c6\uff0c\u4f1a\u6709\u8001\u9ec4\u5bb6\u5728 Spark \u4e0a\u624d\u80fd\u7528\u7684\u5de5\u5177\u5417\uff1f</li>\n<li>\u7528\u7ebf\u5bf9\u8054 Spark \u4e5f\u80fd\u548c\u53e6\u5916\u4e24\u4e2a\u673a\u578b\uff0c\u6269\u5c55\u4f7f\u7528 LLM \uff1f</li>\n<li>\u4e09\u8005\u6709\u4ec0\u4e48\u4ea7\u54c1\u786c\u4ef6\u4e0a\u7684\u5dee\u5f02\u533a\u522b\uff1f</li>\n</ul>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/Hermitist", 
        "name": "Hermitist", 
        "avatar": "https://cdn.v2ex.com/gravatar/da6c1e355f86d79cd2887cb34a3c864e?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1210041", 
      "date_modified": "2026-05-03T03:50:51+00:00", 
      "content_html": "<a target=\"_blank\" href=\"https://tps.bunai.cc/ranking?gpu=apple_m5_32g&amp;ic=nvlink5\" rel=\"nofollow noopener\">https://tps.bunai.cc/ranking?gpu=apple_m5_32g&amp;ic=nvlink5</a>", 
      "date_published": "2026-05-02T22:04:03+00:00", 
      "title": "\u63a8\u8350\u4e00\u4e2a GPU \u63a8\u7406\u901f\u5ea6\u8ba1\u7b97\u5668, \u53ef\u80fd\u65b9\u4fbf\u4e70\u914d\u4ef6\u81ea\u5efa\u672c\u5730\u5927\u6a21\u578b\u7684\u4eba\u7528\u4e0a", 
      "id": "https://www.v2ex.com/t/1210041"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/stefwoo", 
        "name": "stefwoo", 
        "avatar": "https://cdn.v2ex.com/gravatar/3235f0cda8845c15590a7b5de5c89f10?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1210011", 
      "title": "github \u770b\u5230\u4e00\u4e2a\u9879\u76ee\uff0c 3090 \u8dd1 27B\uff0c 129tps\uff0c\u6700\u9ad8 207tps", 
      "id": "https://www.v2ex.com/t/1210011", 
      "date_published": "2026-05-02T13:05:12+00:00", 
      "content_html": "<p><a href=\"https://github.com/Luce-Org/lucebox-hub\" rel=\"nofollow\">https://github.com/Luce-Org/lucebox-hub</a></p>\n<p>DFlash DDtree Qwen3.5 &amp; Qwen3.6 27B GGUF on RTX 3090\nFirst GGUF port of DFlash speculative decoding. Qwen3.5-27B on a single RTX 3090, Q4_K_M target + BF16 draft, DDTree budget=22.</p>\n<p>Up to 207 tok/s in the demo (207.6 tok/s DFlash vs 38.0 tok/s AR, 5.46\u00d7)\n129.5 tok/s mean on the HumanEval 10-prompt bench\n3.43\u00d7 faster than autoregressive (+15% over chain speculative decoding)\n2.8\u00d7 faster than SGLang AWQ on the same hardware\nUp to 256K context in 24 GB via TurboQuant TQ3_0 KV cache (128K Q4_0 bench: 134.78 tok/s at ctx=131072)</p>\n<p>PFlash speculative prefill on RTX 3090\nIn-process speculative prefill, C++/CUDA only. A drafter (Qwen3-0.6B BF16) loaded directly into the dflash daemon scores per-token importance over a long prompt; the heavy target (Qwen3.6-27B Q4_K_M) only prefills the spans that matter. Both models share the same ggml allocator on a single RTX 3090. No Python, no Triton, no PyTorch at runtime \u2014 just the dflash binary and four custom CUDA kernels (mean_K \u2192 score \u2192 select \u2192 sparse_fwd) plus BSA (mit-han-lab/Block-Sparse-Attention, FA-2 derived, sm_80+) for the long-context drafter forward.</p>\n<p>~10.4\u00d7 TTFT on 128K context: 24.8 s dflash daemon vs ~257 s llama.cpp (FA on, Q4_0 KV).\n10.0\u00d7 TTFT on 64K context: 13.5 s dflash vs 134.95 s llama.cpp.\nNIAH single-needle retrieved at every measured context (32K \u2192 128K), keep_ratio=0.05, DFLASH_FP_ALPHA=0.85.</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/ken2025", 
        "name": "ken2025", 
        "avatar": "https://cdn.v2ex.com/gravatar/ea474030c4b6eba9d5fee409e34a152b?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1209904", 
      "date_modified": "2026-05-01T22:53:55+00:00", 
      "content_html": "", 
      "date_published": "2026-05-01T14:31:22+00:00", 
      "title": "\u8bf7\u95ee\u5404\u4f4d\u5927\u795e\uff0c\u5728\u9694\u79bb\u73af\u5883\u4e2d\uff0c\u6709\u672c\u5730 qwen \u5927\u6a21\u578b\uff0c\u6709\u6ca1\u4ec0\u4e48\u89e3\u51b3\u65b9\u6848\uff0c\u505a\u672c\u5730\u7684\u77e5\u8bc6\u5e93\u7684\u65b9\u6848\uff0c\u7c7b\u4f3c\u8c37\u6b4c\u90a3\u4e2a notebooklm \uff0c\u4e5f\u52c9\u5f3a\u53ef\u4ee5\uff1f", 
      "id": "https://www.v2ex.com/t/1209904"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/alangz", 
        "name": "alangz", 
        "avatar": "https://cdn.v2ex.com/gravatar/5e4292e3b311398ca05de583f2fb3526?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1209674", 
      "title": "\u6709\u4e00\u53f0 16 \u5bf8 m1max 64g+1T \u6ee1 GPU \u7684 MacBook Pro \u9002\u5408\u90e8\u7f72\u54ea\u4e2a\u672c\u5730\u6a21\u578b", 
      "id": "https://www.v2ex.com/t/1209674", 
      "date_published": "2026-04-30T07:49:02+00:00", 
      "content_html": "<p>\u914d\u7f6e\u4e3a m1max 64g+1T \uff0c\u6700\u8fd1\u641e\u4e86\u4e2a\u5c0f\u5c0f\u9f99\u867e\uff0c\u6d88\u8017\u7684 token \u592a\u5feb\u4e86\uff0c\u6253\u7b97\u90e8\u7f72\u5404\u672c\u5730\u6a21\u578b\uff0c\u4e00\u6765\u4e86\u89e3\u4e86\u89e3\uff0c\u800c\u6765\u662f\u60f3\u505a\u4e0b\u7b80\u5355\u7684\u7ffb\u8bd1\u3001\u6587\u6863\u5904\u7406\u7684\u5de5\u4f5c\u3002\u80fd\u6709\u5408\u9002\u7684\u672c\u5730\u6a21\u578b\u5417\uff1f</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/babymonster", 
        "name": "babymonster", 
        "avatar": "https://cdn.v2ex.com/avatar/a37b/761f/743705_large.png?m=1777516327"
      }, 
      "url": "https://www.v2ex.com/t/1209582", 
      "title": "\u79c1\u6709\u5316\u90e8\u7f72\u5927\u6a21\u578b\u7684\u201c\u7ec8\u70b9\u201d\u662f Mac \u8fd8\u662f Nvidia\uff1f", 
      "id": "https://www.v2ex.com/t/1209582", 
      "date_published": "2026-04-30T02:34:14+00:00", 
      "content_html": "\u81ea\u5df1\u5bb6\u91cc\u7684 5070Ti \u8dd1\u6a21\u578b\u8d77\u6765\u592a\u8d39\u52b2\u4e86\uff0c\u7528\u4e86\u4e00\u4e0b\u540c\u4e8b\u7684 macbook \u9876\u914d\u7248\u8dd1\u6a21\u578b\u6bd4 5070Ti \u8981\u5f3a\u4e00\u70b9\u70b9\uff0c\u611f\u89c9\u90fd\u5dee\u4e0d\u591a\uff0c\u6240\u4ee5\u5927\u4f6c\u53ef\u4ee5\u6307\u70b9\u4e0b\u6709\u6ca1\u6709\u5fc5\u8981\u641e\u4e00\u4e2a Mac studio \u8fd8\u662f Nvidia thor \u6216\u8005 DGX Spark"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/babymonster", 
        "name": "babymonster", 
        "avatar": "https://cdn.v2ex.com/avatar/a37b/761f/743705_large.png?m=1777516327"
      }, 
      "url": "https://www.v2ex.com/t/1209353", 
      "date_modified": "2026-04-29T07:48:44+00:00", 
      "content_html": "<p>\u5404\u4f4d\u5927\u4f6c\u4eec\uff0c\u6211\u81ea\u5df1\u7535\u8111\u914d\u7f6e\u4e5f\u633a\u9ad8 9800x3d+5070ti,\u4f46\u662f\u81ea\u5df1\u73a9\u5927\u6a21\u578b\u611f\u89c9\u7b97\u529b\u4e0d\u591f\uff0c\u8f93\u51fa\u901f\u5ea6\u597d\u6162\uff0c\u6709\u6ca1\u6709\u5927\u4f6c\u63a8\u8350\u4e00\u4e0b\u4ec0\u4e48\u663e\u5361\u73a9\u5927\u6a21\u578b\u7b97\u529b\u8231\u8fd8\u6bd4\u8f83\u4e0d\u9519\u7684</p>\n", 
      "date_published": "2026-04-29T05:33:08+00:00", 
      "title": "\u6211\u81ea\u5df1\u7684\u7535\u8111\u662f 5070Ti\uff0c\u603b\u611f\u89c9\u8dd1\u4e00\u4e9b\u6a21\u578b\u7b97\u529b\u4e0d\u591f", 
      "id": "https://www.v2ex.com/t/1209353"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/KaiWuBOSS", 
        "name": "KaiWuBOSS", 
        "avatar": "https://cdn.v2ex.com/gravatar/243db3a31aa62a02d726471a3fd1782e?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1209195", 
      "title": "\u80fd\u4e00\u8d77\u7ed9\u672c\u5730\u90e8\u7f72\u7684\u5f00\u6e90\u6a21\u578b\u505a\u4e2a\u9002\u914d\u7684 coding agent \u5417\uff1f\u6211\u618b\u4e86\u53e3\u6c14", 
      "id": "https://www.v2ex.com/t/1209195", 
      "date_published": "2026-04-28T13:24:51+00:00", 
      "content_html": "<h1>\u6211\u505a\u4e86\u4e00\u4e2a\u4e13\u95e8\u4e3a\u672c\u5730\u5f00\u6e90\u6a21\u578b\u4f18\u5316\u7684 Coding Agent \uff0c\u5e0c\u671b\u66f4\u591a\u534e\u4eba\u5f00\u53d1\u8005\u4e00\u8d77\u6765\u641e</h1>\n<blockquote>\n<p>\u672c\u8d34\u53d1\u5e03\u7684\u76ee\u7684\u4e0d\u662f\u63a8\u4ea7\u54c1\uff0c\u4e0d\u662f\u70ab\u6280\uff0c\u800c\u662f\u60f3\u626c\u7709\u5410\u6c14\u2014\u2014\u548c\u534e\u4eba\u5f00\u53d1\u8005\u4e00\u8d77\uff0c\u548c\u5f00\u6e90\u6a21\u578b\u672c\u5730\u90e8\u7f72\u5f00\u53d1\u8005\u4e00\u8d77\uff0c\u505a\u4e00\u4ef6\u6211\u4eec\u81ea\u5df1\u7684\u4e8b\u3002</p>\n</blockquote>\n<hr/>\n<h2>\u4e00\u3001\u6211\u9047\u5230\u4e86\u4ec0\u4e48\u95ee\u9898</h2>\n<p>\u53bb\u5e74\u5f00\u59cb\u7528\u672c\u5730\u6a21\u578b\u505a\u7f16\u7a0b\u8f85\u52a9\u3002\u539f\u56e0\u5f88\u7b80\u5355\uff1a\u516c\u53f8\u4ee3\u7801\u4e0d\u80fd\u4f20\u5230\u6d77\u5916\u670d\u52a1\u5668\uff0cClaude Code \u548c Cursor \u8d70\u4e0d\u901a\u3002</p>\n<p>\u4f46\u66f4\u5927\u7684\u95ee\u9898\u662f\uff1a<strong>\u4e2d\u56fd\u5f00\u53d1\u8005\u6839\u672c\u6ca1\u6709\u4e00\u4e2a\u597d\u7528\u7684\u672c\u5730 coding agent \u5e73\u53f0\u3002</strong></p>\n<p>CC \u9700\u8981\u7ffb\u5899\uff0c\u8fd8\u8981\u8ba2\u9605\u3002Cursor \u540c\u6837\u3002Codex \u521a\u51fa\u6765\u4e5f\u662f\u6d77\u5916\u670d\u52a1\u3002Hermes \u8fd9\u7c7b\u5f00\u6e90\u5de5\u5177\u4e0d\u652f\u6301 Windows \u539f\u751f\u8fd0\u884c\uff0c\u8981\u88c5 WSL2 \uff0c\u529d\u9000\u4e86\u5927\u591a\u6570\u56fd\u5185\u5f00\u53d1\u8005\u3002\u6700\u540e\u5927\u5bb6\u7684\u9009\u62e9\u662f\uff1a\u8981\u4e48\u7ffb\u5899\u51d1\u5408\u7528\uff0c\u8981\u4e48\u5fcd\u7740\u4e0d\u7528\u3002</p>\n<p><strong>\u8fd9\u662f\u4e00\u4e2a\u771f\u5b9e\u5b58\u5728\u7684\u7a7a\u7f3a\uff0c\u6ca1\u6709\u4eba\u586b\u3002</strong></p>\n<p>\u672c\u5730\u8dd1 qwen3:8b \uff0c\u7136\u540e\u53d1\u73b0\u95ee\u9898\u4e00\u4e2a\u63a5\u4e00\u4e2a\uff1a</p>\n<p><strong>\ud83d\udd34 \u65e0\u9650\u5faa\u73af\uff0c\u50cf\u5361\u5e26\u4e00\u6837</strong></p>\n<p>\u8fd9\u662f\u672c\u5730\u5c0f\u6a21\u578b\u6700\u8ba9\u4eba\u6293\u72c2\u7684\u95ee\u9898\u3002\u9047\u5230\u5b83\u4e0d\u4f1a\u5904\u7406\u7684\u573a\u666f\uff0c\u5b83\u4e0d\u4f1a\u8bf4\"\u6211\u4e0d\u77e5\u9053\"\uff0c\u800c\u662f\u5f00\u59cb\u91cd\u590d\u2014\u2014\u540c\u4e00\u53e5\u8bdd\u8bf4\u4e09\u904d\uff0c\u540c\u4e00\u4e2a\u9519\u8bef\u7684\u4fee\u6539\u5efa\u8bae\u5faa\u73af\u51fa\u73b0\uff0c\u540c\u4e00\u6bb5\u4ee3\u7801\u53cd\u590d\u751f\u6210\u3002\u6574\u4e2a\u4efb\u52a1\u5361\u6b7b\uff0c\u53ea\u80fd\u624b\u52a8\u5f3a\u5236\u9000\u51fa\u3002\u8fd9\u4e0d\u662f\u5076\u53d1\u73b0\u8c61\uff0c\u662f\u5c0f\u6a21\u578b\u5728\u63a8\u7406\u80fd\u529b\u4e0d\u8db3\u65f6\u7684\u5178\u578b\u5d29\u6e83\u6a21\u5f0f\u3002</p>\n<p><strong>\ud83d\udd34 \u4fee bug \u53cd\u590d\u8e29\u540c\u4e00\u4e2a\u5751</strong></p>\n<p>\u8ba9\u5b83\u4fee\u4e00\u4e2a\u51fd\u6570\uff0c\u7b2c\u4e00\u6b21\u5931\u8d25\uff0c\u7b2c\u4e8c\u6b21\u7528\u5b8c\u5168\u4e00\u6837\u7684\u65b9\u5f0f\u518d\u8bd5\uff0c\u7b2c\u4e09\u6b21\u4f9d\u7136\u3002\u4e09\u6b21\u673a\u4f1a\u5168\u6d6a\u8d39\u5728\u540c\u4e00\u4e2a\u9519\u8bef\u4e0a\uff0c\u4ec0\u4e48\u90fd\u6ca1\u63a8\u8fdb\u3002</p>\n<p><strong>\ud83d\udd34 \u6a21\u578b\u80fd\u529b\u672c\u8eab\u5c31\u5f31\u4e8e API \u6a21\u578b</strong></p>\n<p>\u8fd9\u662f\u65e0\u6cd5\u56de\u907f\u7684\u73b0\u5b9e\u30028B \u300114B \u7684\u53c2\u6570\u91cf\uff0c\u63a8\u7406\u80fd\u529b\u548c Claude Opus \u3001GPT-4 \u5dee\u8ddd\u660e\u663e\u3002\u8ba9\u4e00\u4e2a 8B \u6a21\u578b\u625b\u4e0b\u4e00\u4e2a\u590d\u6742\u4efb\u52a1\u7684\u5168\u90e8\u63a8\u7406\uff0c\u6210\u529f\u7387\u5f88\u4f4e\uff0c\u8fd9\u4e0d\u662f\u54ea\u4e2a\u5de5\u5177\u7684\u95ee\u9898\uff0c\u662f\u6a21\u578b\u672c\u8eab\u7684\u8fb9\u754c\u3002</p>\n<p><strong>\ud83d\udd34 \u627e\u4e0d\u5230\u8981\u6539\u7684\u6587\u4ef6</strong></p>\n<p>\u9879\u76ee\u5927\u4e86\u4e4b\u540e\uff0c\u6a21\u578b\u6839\u672c\u4e0d\u77e5\u9053\u8981\u6539\u54ea\u4e2a\u6587\u4ef6\u3002\u8ba9\u5b83\u627e bug \uff0c\u5b83\u8981\u4e48\u731c\u9519\uff0c\u8981\u4e48\u8bf4\"\u6211\u9700\u8981\u770b\u66f4\u591a\u4ee3\u7801\"\uff0c\u7136\u540e\u628a\u6574\u4e2a\u9879\u76ee\u585e\u8fdb context \uff0c\u7136\u540e context \u53c8\u7206\u4e86\u3002</p>\n<p><strong>\ud83d\udd34 \u5bf9\u8bdd\u51e0\u8f6e\u5c31\u5f00\u59cb\u9057\u5fd8</strong></p>\n<p>8B \u6a21\u578b context \u7a97\u53e3\u53ea\u6709 8K \uff0c\u5bf9\u8bdd\u591a\u4e86\u5c31\u6ee1\u4e86\uff0c\u6a21\u578b\u5f00\u59cb\u7ed9\u51fa\u9a74\u5507\u4e0d\u5bf9\u9a6c\u5634\u7684\u56de\u7b54\u3002</p>\n<hr/>\n<p>\u8fd9\u4e9b\u95ee\u9898\u53e0\u5728\u4e00\u8d77\uff0c\u7528\u672c\u5730\u6a21\u578b\u505a\u5f00\u53d1\u8f85\u52a9\u7684\u4f53\u9a8c\u6781\u5dee\u3002</p>\n<p>\u6240\u4ee5\u6211\u60f3\u81ea\u5df1\u505a\u4e00\u4e2a\u4ea7\u54c1\u6765\u8dd1\u3002\u6709\u4eba\u5c31\u4f1a\u8bf4\uff1a\u4e3a\u4ec0\u4e48\u4e0d\u76f4\u63a5\u7528 ollama + cc \uff1f\u8fd8\u53cb\u60c5\u6307\u5bfc\u6211\u547d\u4ee4\u3002</p>\n<p>\u54ce\u3002</p>\n<p>\u5927\u5382\u7684\u4ea7\u54c1\u53ea\u4f1a\u4e3a\u5b83\u7684\u5546\u4e1a\u6a21\u5f0f\u670d\u52a1\u3002ollama \u653e\u5f03\u4e86\u53c2\u6570\u5fae\u8c03\u6765\u6362\u53d6\u7a33\u5b9a\uff0clm \u8ba9\u5f00\u53d1\u8005\u7ea0\u7ed3\u4ec0\u4e48\u662f\u6700\u4f18\uff0cCC/Codex/Cursor \u90fd\u662f\u5356 token \uff0c<strong>\u6ca1\u6709\u4eba\u4f1a\u771f\u7684\u8ba4\u771f\u60f3\u672c\u5730\u90e8\u7f72\u7f3a\u4ec0\u4e48\uff0c\u9700\u8981\u4f18\u5316\u4ec0\u4e48\uff0c\u8bb0\u5fc6\u600e\u4e48\u4f18\u5316\uff0c\u4e0a\u4e0b\u6587\u600e\u4e48\u538b\u7f29\uff0c\u5c0f\u53c2\u6570\u600e\u4e48\u8f85\u52a9\u3002</strong></p>\n<p>\u4f46\u6211\u4eba\u5fae\u8a00\u8f7b\uff0c\u6240\u4ee5\u6211\u505a\u4e86\u4e2a MVP \u60f3\u629b\u7816\u5f15\u7389\u3002\u6211\u4eec\u53ef\u4ee5\u4e00\u8d77\u628a\u8981\u4f18\u5316\u7684\u90fd\u4f18\u5316\u4e86\uff0c\u6253\u9020\u6211\u4eec\u81ea\u5df1\u7684\u4ea7\u54c1\u3002</p>\n<p>\u6709\u4eba\u4e5f\u8bf4\uff0c\u6211\u80fd\u529b\u4e0d\u591f\u3002</p>\n<p>\u90a3\u6211\u7684\u601d\u8def\u662f\uff1a<strong>\u4e0d\u591f\u5c31\u505a\u6574\u5408\uff0c\u591f\u4e86\u5c31\u505a\u7a81\u7834\u3002</strong></p>\n<p>\u6240\u4ee5\u6211\u505a\u4e86 KWCode \uff0c\u4e0d\u662f\u4e3a\u4e86\u5546\u4e1a\u5316\uff0cMIT \u4efb\u4f55\u4eba\u90fd\u80fd\u62ff\u8d70\uff0c\u53ea\u5e0c\u671b\u54ea\u4e2a\u611f\u5174\u8da3\u7684\u5927\u795e\uff0c\u613f\u610f\u548c\u6211\u6216\u8005\u548c\u6240\u6709\u5f00\u53d1\u8005\u4e00\u8d77\u628a\u5b83\u5b9e\u73b0\u5e76\u5f00\u6e90\uff0c\u7ed9\u6240\u6709\u88ab\u672c\u5730\u90e8\u7f72\u8188\u5e94\u7684\u5b9d\u5b50\u4eec\u3002</p>\n<hr/>\n<h2>\u4e8c\u3001\u6211\u7528\u4e86\u54ea\u4e9b\u601d\u8def</h2>\n<h3>\u601d\u8def\u4e00\uff1aMoE \u67b6\u6784\u2014\u2014\u8ba9 LLM \u53ea\u505a\u5b83\u64c5\u957f\u7684\u90a3\u4e00\u6b65</h3>\n<p>\u8fd9\u662f KWCode \u6700\u6838\u5fc3\u7684\u8bbe\u8ba1\u51b3\u7b56\uff0c\u4e5f\u662f\u89e3\u51b3\u4e0a\u9762\u6240\u6709\u95ee\u9898\u7684\u6839\u672c\u601d\u8def\u3002</p>\n<p>\u4f20\u7edf coding agent \u7684\u67b6\u6784\u662f\uff1a<strong>\u4e00\u4e2a LLM \u625b\u5168\u90e8</strong>\u2014\u2014\u7406\u89e3\u9700\u6c42\u3001\u5b9a\u4f4d\u4ee3\u7801\u3001\u751f\u6210\u4fee\u6539\u3001\u9a8c\u8bc1\u7ed3\u679c\uff0c\u5168\u8ba9\u540c\u4e00\u4e2a\u6a21\u578b\u505a\u3002\u5f3a\u6a21\u578b\u80fd\u625b\uff0c\u5c0f\u6a21\u578b\u625b\u4e0d\u4f4f\uff0c\u7136\u540e\u5c31\u5f00\u59cb\u5faa\u73af\u3001\u5e7b\u89c9\u3001\u4e71\u8bf4\u3002</p>\n<p>KWCode \u7528\u7684\u662f <strong>MoE \uff08 Mixture of Experts \uff09\u67b6\u6784</strong>\uff1a\u628a\u4efb\u52a1\u5207\u788e\uff0c\u6bcf\u4e2a\u4e13\u5bb6\u53ea\u505a\u4e00\u4ef6\u4e8b\uff0cLLM \u53ea\u8d1f\u8d23 Gate \u5206\u7c7b\u548c\u5185\u5bb9\u751f\u6210\uff0c\u5176\u4ed6\u6b65\u9aa4\u80fd\u4e0d\u8c03 LLM \u5c31\u4e0d\u8c03\u3002</p>\n<pre><code>\u7528\u6237\u8f93\u5165\n  \u2514\u2500\u25ba Gate \uff08 LLM \u505a\u4e00\u6b21\u5206\u7c7b\uff0c\u5224\u65ad\u4efb\u52a1\u7c7b\u578b\uff09\n        \u2514\u2500\u25ba Locator \uff08 BM25 + \u8c03\u7528\u56fe\uff0c\u4e0d\u8c03 LLM \uff0c\u6beb\u79d2\u7ea7\u5b9a\u4f4d\u6587\u4ef6\u548c\u51fd\u6570\uff09\n              \u2514\u2500\u25ba Generator \uff08 LLM \u53ea\u5199\u9700\u8981\u4fee\u6539\u7684\u90a3\u51e0\u884c\u4ee3\u7801\uff09\n                    \u2514\u2500\u25ba Verifier \uff08\u81ea\u52a8\u8dd1\u8bed\u6cd5\u68c0\u67e5 + pytest \uff0c\u4e0d\u8c03 LLM \uff09\n                          \u2514\u2500\u25ba SearchAugmentor \uff08\u4e24\u6b21\u5931\u8d25\u540e\u81ea\u52a8\u641c\u7d22\uff09\n</code></pre>\n<p>LLM \u5728\u8fd9\u6761\u6d41\u6c34\u7ebf\u91cc\u7684\u4efb\u52a1\u88ab\u538b\u5230\u4e86\u6700\u5c0f\uff1aGate \u505a\u4e00\u6b21\u5206\u7c7b\uff0cGenerator \u751f\u6210\u51e0\u884c\u4ee3\u7801\u3002\u5b9a\u4f4d\u6587\u4ef6\u3001\u9a8c\u8bc1\u7ed3\u679c\u8fd9\u4e24\u4ef6\u6700\u8017\u63a8\u7406\u80fd\u529b\u7684\u4e8b\uff0c\u5b8c\u5168\u4e0d\u8ba9 LLM \u505a\u3002</p>\n<blockquote>\n<p>\u53c2\u8003\uff1aAgentless \u8bba\u6587\uff08 ICSE 2025 \uff09\u2014\u2014\u786e\u5b9a\u6027\u6d41\u6c34\u7ebf\u5728 SWE-bench \u4e0a\u540c\u65f6\u8fbe\u5230\u6700\u9ad8\u901a\u8fc7\u7387\u548c\u6700\u4f4e\u6210\u672c\uff0c\u4f18\u4e8e\u8ba9 LLM \u81ea\u4e3b\u51b3\u7b56\u7684\u590d\u6742 agent \u3002\u539f\u56e0\u5f88\u7b80\u5355\uff1a\u6bcf\u4e00\u6b65 scope \u6781\u5c0f\uff0c\u5c0f\u6a21\u578b\u5728\u5c0f scope \u91cc\u8868\u73b0\u7a33\u5b9a\u3002</p>\n</blockquote>\n<hr/>\n<h3>\u601d\u8def\u4e8c\uff1a\u7528\u8c03\u7528\u56fe\u5b9a\u4f4d\u4ee3\u7801\uff0c\u4e0d\u9760 LLM \u731c</h3>\n<p>\u4ee3\u7801\u5b9a\u4f4d\u662f\u5c0f\u6a21\u578b\u6700\u5bb9\u6613\u5931\u8d25\u7684\u6b65\u9aa4\uff0c\u628a\u5b83\u4ece LLM \u624b\u91cc\u62ff\u8d70\uff0c\u6362\u6210\u786e\u5b9a\u6027\u7b97\u6cd5\u3002</p>\n<p>CodeCompass \uff08 arXiv:2602.20048 \uff0c2026 \u5e74\uff09\u505a\u4e86 258 \u6b21\u5b9e\u9a8c\uff0c\u53d1\u73b0\u4e86\u4e00\u4e2a\u5173\u952e\u7ed3\u8bba\uff1a</p>\n<blockquote>\n<p>\u771f\u5b9e\u9879\u76ee\u91cc\uff0c\u5f88\u591a bug \u7684\u6839\u56e0\u6587\u4ef6\u540d\u548c\u9519\u8bef\u63cf\u8ff0\u6beb\u65e0\u5173\u8054\uff0c\u53ea\u80fd\u901a\u8fc7\u8c03\u7528\u94fe\u8ffd\u8e2a\u624d\u80fd\u627e\u5230\u3002\u5bf9\u8fd9\u7c7b\"\u9690\u85cf\u4f9d\u8d56\"\u4efb\u52a1\uff0cBM25 \u5173\u952e\u8bcd\u641c\u7d22\u51c6\u786e\u7387\u53ea\u6709 **76.2%**\uff0c\u800c\u56fe\u904d\u5386\u8fbe\u5230 **99.4%**\uff0c\u5dee\u4e86 23 \u4e2a\u767e\u5206\u70b9\u3002</p>\n</blockquote>\n<p>KWCode \u7684\u4e24\u9636\u6bb5\u68c0\u7d22\uff1a</p>\n<ol>\n<li><strong>BM25 \u5173\u952e\u8bcd\u53ec\u56de</strong>\uff08\u6beb\u79d2\u7ea7\uff0c\u4e0d\u8c03 LLM \uff09\uff1a\u4ece\u4ee3\u7801\u5e93\u6240\u6709\u51fd\u6570/\u7c7b\u4e2d\uff0c\u5feb\u901f\u53ec\u56de top-20 \u5019\u9009</li>\n<li><strong>AST \u8c03\u7528\u56fe\u5c55\u5f00</strong>\uff08\u6beb\u79d2\u7ea7\uff0c\u4e0d\u8c03 LLM \uff09\uff1a\u5bf9\u6bcf\u4e2a\u5019\u9009\u51fd\u6570\uff0c\u6cbf\u8c03\u7528\u56fe\u5411\u4e0a\u5411\u4e0b\u5404\u5c55\u5f00 2 \u8df3\uff0c\u53d1\u73b0\u9690\u85cf\u4f9d\u8d56</li>\n</ol>\n<p>\u6574\u4e2a\u8fc7\u7a0b\u4e0d\u8c03 LLM \uff0cSQLite \u6301\u4e45\u5316\u8c03\u7528\u56fe\uff0c\u91cd\u542f\u4e0d\u91cd\u5efa\u3002</p>\n<p>\u6280\u672f\u6808\uff1a<code>tree-sitter</code> + <code>rank-bm25</code> + <code>SQLite</code>\u3002\u4e0d\u9700\u8981 Neo4j \uff0c\u4e0d\u9700\u8981 embedding \u6a21\u578b\uff0c\u4e0d\u9700\u8981\u989d\u5916 Docker \u3002</p>\n<hr/>\n<h3>\u601d\u8def\u4e09\uff1a\u6253\u7834\u5faa\u73af\u2014\u2014\u5931\u8d25\u65f6\u5f3a\u5236\u6362\u7b56\u7565</h3>\n<p>\u9488\u5bf9\"\u53cd\u590d\u8e29\u540c\u4e00\u4e2a\u5751\"\u548c\"\u65e0\u9650\u5faa\u73af\"\u8fd9\u4e24\u4e2a\u95ee\u9898\uff1a</p>\n<p><strong>\u53cd\u65e0\u9650\u5faa\u73af</strong>\uff1aMAX_RETRIES \u786c\u7f16\u7801\u4e3a 3 \uff0c\u6ca1\u6709\u4efb\u4f55\u8def\u5f84\u80fd\u7ed5\u8fc7\u3002\u540c\u65f6\u68c0\u6d4b\u8fde\u7eed\u4e24\u6b21\u751f\u6210\u5b8c\u5168\u76f8\u540c\u7684 patch \uff0c\u76f4\u63a5\u8df3\u8fc7\u4e0d\u91cd\u8bd5\uff0c\u544a\u8bc9\u7528\u6237\"\u6a21\u578b\u5361\u4f4f\u4e86\uff0c\u5efa\u8bae\u7f29\u5c0f\u4efb\u52a1\u8303\u56f4\"\u3002</p>\n<p><strong>\u53cd\u91cd\u590d\u5931\u8d25</strong>\uff1a\u4e09\u6b21\u91cd\u8bd5\u5f3a\u5236\u7528\u4e09\u79cd\u4e0d\u540c\u7684\u95ee\u9898\u8868\u8ff0\uff1a</p>\n<table>\n<thead>\n<tr>\n<th>\u7b2c\u51e0\u6b21</th>\n<th>\u7b56\u7565</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>\u7b2c\u4e00\u6b21</td>\n<td>\u6b63\u5e38\u63cf\u8ff0\u9700\u6c42</td>\n</tr>\n<tr>\n<td>\u7b2c\u4e8c\u6b21</td>\n<td>\u4ece\u9519\u8bef\u4fe1\u606f\u51fa\u53d1\uff1a\"\u76f4\u63a5\u4fee\u590d\u8fd9\u4e2a\u62a5\u9519\uff0c\u4e0d\u8981\u89e3\u91ca\"</td>\n</tr>\n<tr>\n<td>\u7b2c\u4e09\u6b21</td>\n<td>\u6700\u5c0f\u5316\u4fee\u6539\uff1a\"\u53ea\u6539\u8fd9\u4e00\u4e2a\u51fd\u6570\uff0c\u5176\u4ed6\u4ee3\u7801\u4e00\u884c\u4e0d\u52a8\"</td>\n</tr>\n</tbody></table><p>\u7b2c\u4e00\u6b21\u5931\u8d25\u540e\u5148\u505a <strong>Reflection</strong>\uff1a\u8ba9 LLM \u4e00\u53e5\u8bdd\u5206\u6790\u4e0a\u6b21\u5931\u8d25\u7684\u539f\u56e0\uff0c\u7136\u540e\u628a\u8fd9\u4e2a\u5206\u6790\u6ce8\u5165\u4e0b\u6b21\u7684 prompt \u3002\u4e0d\u662f\u8ba9\u6a21\u578b\u81ea\u7531\u53d1\u6325\uff0c\u662f\u5f3a\u5236\u5b83\u5148\u8bca\u65ad\u518d\u4fee\u3002</p>\n<hr/>\n<h3>\u601d\u8def\u56db\uff1a\u4e13\u5bb6\u98de\u8f6e\uff0c\u8d8a\u7528\u8d8a\u61c2\u4f60\u7684\u9879\u76ee</h3>\n<blockquote>\n<p>\u53c2\u8003\uff1aEE-MCP \uff08 NeurIPS 2025 \uff09\u2014\u2014\u4ece\u4efb\u52a1\u6267\u884c\u8f68\u8ff9\u81ea\u52a8\u63d0\u53d6\u7ecf\u9a8c\uff0c\u9a8c\u8bc1\u53ef\u663e\u8457\u63d0\u5347\u540e\u7eed\u540c\u7c7b\u4efb\u52a1\u6210\u529f\u7387\u3002</p>\n</blockquote>\n<p>KWCode \u9884\u7f6e\u4e86 15 \u4e2a\u4e13\u5bb6\uff08 BugFix \u3001TestGen \u3001SpringBoot \u3001FastAPI \u7b49\uff09\uff0c\u6bcf\u4e2a\u4e13\u5bb6\u6709\u72ec\u7acb\u7684 system prompt \u3002</p>\n<p>\u540c\u7c7b\u4efb\u52a1\u6210\u529f 5 \u6b21\u4e4b\u540e\uff0c\u98de\u8f6e\u81ea\u52a8\u5206\u6790\u8f68\u8ff9\uff0c\u751f\u6210\u65b0\u4e13\u5bb6\uff0c\u7ecf\u8fc7<strong>\u4e09\u9053\u9a8c\u8bc1\u95e8</strong>\u540e\u6295\u4ea7\uff1a</p>\n<ul>\n<li><strong>\u56de\u6d4b\u95e8</strong>\uff1a\u65b0\u4e13\u5bb6\u6210\u529f\u7387\u5fc5\u987b \u2265 \u539f\u6d41\u6c34\u7ebf</li>\n<li><strong>AB \u6d4b\u8bd5\u95e8</strong>\uff1a10 \u6b21\u771f\u5b9e\u5bf9\u6bd4\uff0c\u63d0\u5347\u8d85\u8fc7 10% \u624d\u6295\u4ea7</li>\n<li><strong>\u751f\u547d\u5468\u671f</strong>\uff1anew \u2192 mature \u2192 declining \u2192 archived \uff0c\u81ea\u52a8\u6dd8\u6c70\u53d8\u5dee\u7684\u4e13\u5bb6</li>\n</ul>\n<p>\u4e13\u5bb6\u53ef\u4ee5\u5bfc\u51fa\u6210 <code>.kwx</code> \u6587\u4ef6\uff0c<code>kwcode expert install URL</code> \u4e00\u884c\u5b89\u88c5\u522b\u4eba\u5206\u4eab\u7684\u4e13\u5bb6\u3002</p>\n<hr/>\n<h3>\u601d\u8def\u4e94\uff1a\u6a21\u578b\u80fd\u529b\u81ea\u9002\u5e94</h3>\n<p>CC \u4e0d\u9700\u8981\u8003\u8651\u8fd9\u4e2a\uff0c\u56e0\u4e3a\u5b83\u53ea\u7528\u4e00\u4e2a\u6a21\u578b\u3002KWCode \u9700\u8981\u3002</p>\n<p>\u81ea\u52a8\u68c0\u6d4b\u5f53\u524d\u6a21\u578b\u7684\u53c2\u6570\u91cf\uff0c\u7136\u540e\u5e94\u7528\u4e0d\u540c\u7b56\u7565\uff1a</p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b\u89c4\u6a21</th>\n<th>\u81ea\u52a8\u7b56\u7565</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>&lt; 10B \uff08 qwen3:8b \uff09</td>\n<td>\u5f3a\u5236\u8ba1\u5212\u786e\u8ba4 \u00b7 \u4efb\u52a1\u8303\u56f4\u9650 2 \u4e2a\u6587\u4ef6 \u00b7 \u7b2c 1 \u6b21\u5931\u8d25\u89e6\u53d1\u641c\u7d22</td>\n</tr>\n<tr>\n<td>10-30B \uff08 qwen3:14b \uff09</td>\n<td>\u53ef\u9009\u8ba1\u5212 \u00b7 4 \u4e2a\u6587\u4ef6\u8303\u56f4 \u00b7 \u7b2c 2 \u6b21\u5931\u8d25\u89e6\u53d1\u641c\u7d22</td>\n</tr>\n<tr>\n<td>&gt; 30B \uff08 qwen3:72b \uff09</td>\n<td>\u5bbd\u677e\u7b56\u7565 \u00b7 8 \u4e2a\u6587\u4ef6 \u00b7 \u81ea\u52a8\u5904\u7406\u590d\u6742\u4efb\u52a1</td>\n</tr>\n</tbody></table><p>\u5207\u6362\u6a21\u578b\uff0c\u7b56\u7565\u81ea\u52a8\u5207\u6362\u3002</p>\n<hr/>\n<h2>\u4e09\u3001\u73b0\u5728\u505a\u4e86\u4ec0\u4e48</h2>\n<p>\u6838\u5fc3\u529f\u80fd\u8dd1\u901a\u4e86\u3002<strong>282/282 \u5355\u5143\u6d4b\u8bd5\u901a\u8fc7\uff0cE2E \u9a8c\u6536\u901a\u8fc7\u7387 87%\uff08 26/30 \uff0c4 \u4e2a\u5931\u8d25\u662f\u6a21\u578b\u80fd\u529b\u8fb9\u754c\uff0c\u4e0d\u662f\u6846\u67b6\u95ee\u9898\uff09\u3002</strong></p>\n<p><strong>\u4ee3\u7801\u80fd\u529b</strong></p>\n<ul>\n<li>BM25 + AST \u8c03\u7528\u56fe\u4e24\u9636\u6bb5\u5b9a\u4f4d\uff0cG3 \u9690\u85cf\u4f9d\u8d56\u51c6\u786e\u7387 99.4%\uff08\u8bba\u6587\u9a8c\u8bc1\uff09</li>\n<li>\u4e09\u9636\u6bb5\u91cd\u8bd5 + Reflection \uff0c\u4e0d\u91cd\u590d\u540c\u6837\u7684\u9519</li>\n<li>\u4e13\u5bb6\u98de\u8f6e\u4e09\u9053\u95e8\uff08\u8f68\u8ff9 \u2192 \u6a21\u5f0f \u2192 AB \u6d4b\u8bd5 \u2192 \u6295\u4ea7\uff09</li>\n<li>15 \u4e2a\u9884\u7f6e\u4e13\u5bb6\uff08\u901a\u7528 + SpringBoot / MyBatis / FastAPI / UniApp \u7b49\uff09</li>\n<li>Office \u6587\u6863\u751f\u6210\uff08 Excel / PPT / Word \uff0c\u6709\u6837\u5f0f\u4e0d\u662f\u767d\u5e95\uff09</li>\n</ul>\n<p><strong>\u5de5\u7a0b\u80fd\u529b</strong></p>\n<ul>\n<li><a href=\"http://KWCODE.md\" rel=\"nofollow\">KWCODE.md</a> \u9879\u76ee\u89c4\u5219\u6587\u4ef6\uff0c\u6309\u4efb\u52a1\u7c7b\u578b\u5206\u6bb5\u6ce8\u5165\uff0c\u6c38\u8fdc\u4e0d\u5fd8</li>\n<li><code>/plan</code> \u8ba1\u5212\u6a21\u5f0f + \u98ce\u9669\u8bc4\u4f30\uff08 High/Medium/Low \uff0c\u57fa\u4e8e\u5386\u53f2\u5931\u8d25\u8bb0\u5f55\uff09</li>\n<li>Checkpoint \u6587\u4ef6\u5feb\u7167\uff0c\u5931\u8d25\u4e00\u952e\u8fd8\u539f</li>\n<li>\u975e\u4ee3\u7801\u6587\u4ef6\u8bfb\u53d6\uff08 PDF / Word / Markdown \uff0cBM25 \u6bb5\u843d\u5339\u914d\u6ce8\u5165\uff09</li>\n<li>\u641c\u7d22\u589e\u5f3a\uff08 SearXNG \u81ea\u90e8\u7f72 + DDG fallback \uff0c\u56db\u7ea7\u5185\u5bb9\u63d0\u53d6\uff09</li>\n</ul>\n<p><strong>\u4f53\u9a8c</strong></p>\n<ul>\n<li>Windows cmd/PowerShell \u539f\u751f\u652f\u6301\uff0c\u4e0d\u9700\u8981 WSL2</li>\n<li>\u9996\u6b21\u5f15\u5bfc\uff08 API \u914d\u7f6e + \u8fde\u901a\u6027\u9a8c\u8bc1\uff09</li>\n<li>\u6267\u884c\u8fc7\u7a0b\u53ea\u663e\u793a spinner \uff0c\u5b8c\u6210\u540e\u8f93\u51fa\u7528\u6237\u53ef\u8bfb\u7684\u7ed3\u679c\u6458\u8981</li>\n<li>\u652f\u6301\u4efb\u4f55 OpenAI \u517c\u5bb9 API \uff08\u672c\u5730 Ollama / DeepSeek / \u7845\u57fa\u6d41\u52a8\u7b49\uff09</li>\n</ul>\n<hr/>\n<h2>\u56db\u3001\u8fd8\u5dee\u4ec0\u4e48</h2>\n<p>\u8bf4\u5b9e\u8bdd\uff0c\u6709\u4e9b\u5730\u65b9\u8fd8\u633a\u7c97\u7cd9\u7684\uff1a</p>\n<ul>\n<li>AST \u8c03\u7528\u56fe\u76ee\u524d\u53ea\u5b8c\u6574\u652f\u6301 Python \uff0c\u5176\u4ed6\u8bed\u8a00\u8c03\u7528\u56fe\u51c6\u786e\u7387\u8fd8\u6ca1\u6709\u5145\u5206\u9a8c\u8bc1</li>\n<li>\u4e13\u5bb6\u98de\u8f6e\u7684 Gate 2 \u56de\u6d4b\u903b\u8f91\u504f\u7b80\u5355\uff0c\u8fd8\u4e0d\u591f\u4e25\u683c</li>\n<li>Windows \u4e0a\u7684\u5404\u79cd\u8fb9\u754c\u60c5\u51b5\uff08 AMD \u663e\u5361\u3001\u90e8\u5206 Ollama \u7248\u672c\u517c\u5bb9\u6027\u3001\u4e2d\u6587\u8def\u5f84\uff09\u6ca1\u6709\u5145\u5206\u6d4b\u8bd5</li>\n<li>\u9489\u9489/\u98de\u4e66 webhook \u6ca1\u505a\uff0c\u624b\u673a\u53d1\u6d88\u606f\u89e6\u53d1 agent \u8fd9\u4e2a\u573a\u666f\u8bbe\u8ba1\u4e86\u4f46\u6ca1\u5b9e\u73b0</li>\n<li>\u6ca1\u6709 IDE \u63d2\u4ef6\uff0c\u76ee\u524d\u53ea\u6709 CLI</li>\n<li>Prompt Optimizer \uff08\u7528 Opus API \u81ea\u52a8\u8fed\u4ee3\u4f18\u5316\u4e13\u5bb6 prompt \uff09\u53ea\u505a\u4e86\u6846\u67b6\uff0c\u6ca1\u6709\u8dd1\u8d77\u6765</li>\n</ul>\n<hr/>\n<h2>\u4e94\u3001\u4e3a\u4ec0\u4e48\u60f3\u8ba9\u66f4\u591a\u4eba\u4e00\u8d77\u505a</h2>\n<p>\u6211\u4e00\u4e2a\u4eba\u505a\u8fd9\u4e2a\u5de5\u5177\u6709\u660e\u663e\u7684\u4e0a\u9650\uff0c\u4e0d\u662f\u6280\u672f\u4e0a\u7684\u4e0a\u9650\uff0c\u662f\u89c6\u91ce\u4e0a\u7684\u4e0a\u9650\u3002</p>\n<p>\u6211\u81ea\u5df1\u4e3b\u8981\u7528 Python \u548c FastAPI \uff0c\u6240\u4ee5\u8fd9\u65b9\u9762\u60f3\u5f97\u7ec6\u3002\u4f46\u6211\u4e0d\u77e5\u9053\u6bcf\u5929\u5199 Spring Boot \u7684\u4eba\u6700\u75db\u7684\u70b9\u5728\u54ea\uff0c\u4e0d\u77e5\u9053\u641e Rust \u7684\u4eba\u5728\u672c\u5730\u6a21\u578b\u4e0a\u9047\u5230\u4ec0\u4e48\u95ee\u9898\uff0c\u4e0d\u77e5\u9053\u505a\u5c0f\u7a0b\u5e8f\u7684\u4eba\u9700\u8981\u4ec0\u4e48\u3002</p>\n<p>\u66f4\u91cd\u8981\u7684\u662f\uff0c<strong>\u8fd9\u4ef6\u4e8b\u4e0d\u5e94\u8be5\u53ea\u662f\u4e00\u4e2a\u4eba\u7684\u5de5\u5177\uff0c\u5e94\u8be5\u662f\u4e2d\u56fd\u5f00\u53d1\u8005\u793e\u533a\u7684\u5de5\u5177\u3002</strong></p>\n<p>CC \u662f Anthropic \u7684\uff0cCursor \u662f\u7f8e\u56fd\u516c\u53f8\u7684\uff0cHermes \u662f\u5916\u56fd\u793e\u533a\u505a\u7684\u3002\u6211\u4eec\u7528\u7684\u5de5\u5177\uff0c\u6211\u4eec\u7684\u4f7f\u7528\u4e60\u60ef\u3001\u6280\u672f\u6808\u504f\u597d\u3001\u672c\u5730\u5316\u9700\u6c42\uff0c\u4ece\u6765\u90fd\u662f\u522b\u4eba\u987a\u624b\u52a0\u8fdb\u53bb\u7684\u529f\u80fd\uff0c\u4e0d\u662f\u7b2c\u4e00\u4f18\u5148\u7ea7\u3002</p>\n<p>\u6211\u60f3\u505a\u7684\u662f\u53cd\u8fc7\u6765\u2014\u2014<strong>\u628a\u4e2d\u56fd\u5f00\u53d1\u8005\u7684\u9700\u6c42\u653e\u5728\u7b2c\u4e00\u4f4d\uff0c\u628a\u672c\u5730\u5f00\u6e90\u6a21\u578b\u7684\u9002\u914d\u653e\u5728\u7b2c\u4e00\u4f4d\uff0c\u7136\u540e\u628a\u8fd9\u4e2a\u5de5\u5177\u505a\u5230\u80fd\u548c\u5927\u5382\u4ea7\u54c1\u63b0\u624b\u8155\u3002</strong></p>\n<p>\u8fd9\u4ef6\u4e8b\u4e00\u4e2a\u4eba\u505a\u4e0d\u5230\uff0c\u4f46\u5f00\u6e90\u793e\u533a\u53ef\u4ee5\u3002</p>\n<p>Linux \u6253\u8d25\u4e86 Unix \uff0c\u4e0d\u662f\u56e0\u4e3a\u67d0\u4e00\u4e2a\u5929\u624d\uff0c\u800c\u662f\u5168\u7403\u5f00\u53d1\u8005\u5171\u540c\u7ef4\u62a4\u4e86\u51e0\u5341\u5e74\u3002VSCode \u80fd\u8d85\u8fc7\u90a3\u4e48\u591a\u5546\u4e1a IDE \uff0c\u4e5f\u662f\u56e0\u4e3a\u80cc\u540e\u6709\u5e9e\u5927\u7684\u63d2\u4ef6\u548c\u8d21\u732e\u751f\u6001\u3002</p>\n<p>KWCode \u4e0d\u9700\u8981\u4f60\u6709\u591a\u9ad8\u7684\u6c34\u5e73\uff0c\u53ea\u9700\u8981\u4f60\u5728\u7528\u672c\u5730\u6a21\u578b\u505a\u5f00\u53d1\uff0c\u7136\u540e\u628a\u4f60\u9047\u5230\u7684\u95ee\u9898\u3001\u4f60\u7684\u89e3\u6cd5\u3001\u4f60\u7684\u6539\u8fdb\u8d21\u732e\u8fdb\u6765\u3002<strong>\u591a\u4e00\u4e2a\u4eba\uff0c\u5c31\u591a\u4e00\u4e2a\u4f7f\u7528\u573a\u666f\u88ab\u7167\u987e\u5230\uff0c\u591a\u4e00\u4e2a\u5751\u88ab\u586b\u6389\u3002</strong></p>\n<blockquote>\n<p>Fork \u8fd9\u4e2a\u9879\u76ee\uff0c\u6539\u8fdb\u4f60\u6700\u75db\u7684\u90a3\u4e2a\u70b9\uff0c\u63d0 PR \uff0c\u6211\u4eec\u4e92\u76f8\u501f\u529b\uff0c\u4e00\u8d77\u628a\u5b83\u505a\u597d\u3002</p>\n</blockquote>\n<p>\u95ed\u6e90\u5927\u5382\u6709\u94b1\u6709\u4eba\u6709\u7b97\u529b\uff0c\u6211\u4eec\u6709\u4ec0\u4e48\uff1f\u6211\u4eec\u6709\u771f\u5b9e\u7684\u4f7f\u7528\u573a\u666f\uff0c\u6709\u5bf9\u672c\u5730\u90e8\u7f72\u7684\u771f\u5b9e\u9700\u6c42\uff0c\u6709\u4e0d\u4f9d\u8d56\u6d77\u5916\u670d\u52a1\u7684\u52a8\u529b\u3002<strong>\u8fd9\u5df2\u7ecf\u8db3\u591f\u4e86\u3002</strong></p>\n<hr/>\n<h2>\u516d\u3001\u600e\u4e48\u53c2\u4e0e</h2>\n<p><strong>\u9879\u76ee\u5730\u5740</strong>\uff1a<a href=\"https://github.com/val1813/kwcode\" rel=\"nofollow\">github.com/val1813/kwcode</a></p>\n<pre><code class=\"language-bash\"># Fork \u9879\u76ee\uff0c\u514b\u9686\u5230\u672c\u5730\ngit clone https://github.com/your-fork/kwcode.git\ncd kwcode\n\n# \u5b89\u88c5\u5f00\u53d1\u7248\npip install -e \".[dev]\"\n\n# \u8fd0\u884c\u6d4b\u8bd5\u786e\u8ba4\u73af\u5883\u6b63\u5e38\npython -m pytest kaiwu/tests/ -v\n\n# \u627e\u4e00\u4e2a\u4f60\u6700\u60f3\u6539\u7684\u5730\u65b9\uff0c\u5f00\u59cb\u52a8\u624b\ngit checkout -b fix/your-improvement\n</code></pre>\n<p>\u6539\u4ec0\u4e48\u90fd\u53ef\u4ee5\uff1a</p>\n<ul>\n<li>\u4f60\u6bcf\u5929\u7528 Go \u5199\u4ee3\u7801\uff0c\u89c9\u5f97 Go \u7684 AST \u8c03\u7528\u56fe\u652f\u6301\u4e0d\u591f\u597d\uff0c\u5c31\u53bb\u6539\u5b83</li>\n<li>\u4f60\u5728\u7528 Qwen3 \u53d1\u73b0\u67d0\u4e2a\u573a\u666f\u603b\u662f\u89e6\u53d1\u65e0\u9650\u5faa\u73af\uff0c\u5c31\u53bb\u4fee\u5b83</li>\n<li>\u4f60\u6709\u66f4\u597d\u7684 context \u538b\u7f29\u7b97\u6cd5\uff0c\u5c31\u66ff\u6362\u6389\u73b0\u6709\u7684</li>\n<li>\u4f60\u53d1\u73b0 README \u5199\u9519\u4e86\uff0c\u6539\u4e00\u4e2a\u5b57\u4e5f\u7b97</li>\n</ul>\n<p><strong>Issues</strong> \u91cc\u5217\u4e86\u5df2\u77e5\u95ee\u9898\u548c\u89c4\u5212\u4e2d\u7684\u529f\u80fd\uff0c\u53ef\u4ee5\u4ece\u90a3\u91cc\u627e\u65b9\u5411\u3002<strong>Discussions</strong> \u91cc\u53ef\u4ee5\u804a\u6280\u672f\u601d\u8def\uff0c\u804a\u67d0\u4e2a\u65b9\u5411\u503c\u4e0d\u503c\u5f97\u505a\u3002</p>\n<p>\u6ca1\u6709\u4ec0\u4e48\u8d21\u732e\u592a\u5c0f\u3002</p>\n<hr/>\n<h2>\u4e03\u3001\u6700\u540e\u8bf4\u4e00\u53e5</h2>\n<p>\u6211\u4e0d\u77e5\u9053 KWCode \u80fd\u4e0d\u80fd\u771f\u7684\u8d85\u8d8a CC \u6216\u8005 Hermes \u3002</p>\n<p>\u4f46\u6211\u77e5\u9053\uff0c\u5982\u679c\u4e2d\u56fd\u5f00\u53d1\u8005\u4e00\u76f4\u7528\u522b\u4eba\u505a\u7684\u5de5\u5177\uff0c\u4e00\u76f4\u628a\u81ea\u5df1\u7684\u9700\u6c42\u5f53\u4f5c\"\u6b21\u8981\u529f\u80fd\"\u7b49\u522b\u4eba\u6765\u5b9e\u73b0\uff0c\u8fd9\u4ef6\u4e8b\u6c38\u8fdc\u4e0d\u4f1a\u6709\u7b54\u6848\u3002</p>\n<p><strong>\u6709\u4e9b\u4e1c\u897f\uff0c\u53ea\u6709\u81ea\u5df1\u505a\u624d\u77e5\u9053\u80fd\u4e0d\u80fd\u505a\u5230\u3002</strong></p>\n<p>\u9879\u76ee\u662f MIT \u5f00\u6e90\u7684\uff0c\u4f60\u8d21\u732e\u7684\u4ee3\u7801\u6c38\u8fdc\u662f\u4f60\u7684\u3002\u5982\u679c KWCode \u6700\u540e\u505a\u6210\u4e86\uff0c\u8fd9\u4ef6\u4e8b\u662f\u6240\u6709\u53c2\u4e0e\u7684\u4eba\u4e00\u8d77\u505a\u6210\u7684\u3002</p>\n<hr/>\n<p><strong>\u9879\u76ee\u5730\u5740</strong>\uff1a<a href=\"https://github.com/val1813/kwcode\" rel=\"nofollow\">github.com/val1813/kwcode</a></p>\n<p><em>\u5929\u5de5\u5f00\u7269 \u00b7 KWCode \u00b7 \u4e2d\u56fd\u5f00\u53d1\u8005\u81ea\u5df1\u7684\u672c\u5730 Coding Agent</em></p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/Livid", 
        "name": "Livid", 
        "avatar": "https://cdn.v2ex.com/avatar/c4ca/4238/1_large.png?m=1781025867"
      }, 
      "url": "https://www.v2ex.com/t/1208974", 
      "title": "\u7528 antirez \u7684 llama.cpp fork \u628a DeepSeek v4 Flash \u5728\u672c\u5730\u8dd1\u8d77\u6765\u4e86", 
      "id": "https://www.v2ex.com/t/1208974", 
      "date_published": "2026-04-27T17:53:59+00:00", 
      "content_html": "<p><img alt=\"\" class=\"embedded_image\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"https://i.v2ex.co/86xoYoTs.png\"/></p>\n<p><a href=\"https://github.com/antirez/llama.cpp-deepseek-v4-flash\" rel=\"nofollow\">https://github.com/antirez/llama.cpp-deepseek-v4-flash</a></p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/KaiWuBOSS", 
        "name": "KaiWuBOSS", 
        "avatar": "https://cdn.v2ex.com/gravatar/243db3a31aa62a02d726471a3fd1782e?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1208904", 
      "title": "\u5168\u7403\u672c\u5730\u90e8\u7f72\u5f00\u53d1\u8005\u4eec\u4e00\u8d77\uff0c\u6253\u9020\u4e00\u4e2a\u771f\u6b63\u5c5e\u4e8e\u5f00\u6e90\u793e\u533a\u7684 Coding Agent \u4e86", 
      "id": "https://www.v2ex.com/t/1208904", 
      "date_published": "2026-04-27T10:19:27+00:00", 
      "content_html": "<h1>\u540c\u5fd7\u4eec\uff0c\u662f\u65f6\u5019\u6253\u9020\u4e00\u628a\u771f\u6b63\u5c5e\u4e8e\u5f00\u6e90\u793e\u533a\u7684 Coding Agent \u4e86\uff01</h1>\n<h2>\u73b0\u72b6\uff1a\u6211\u4eec\u88ab\u5927\u5382\u201c\u5582\u5c4e\u201d\uff0c\u8fd8\u8981\u81ea\u5df1\u64e6\u5c41\u80a1</h2>\n<p>\u6211\u5199\u8fc7 <a href=\"https://github.com/val1813/kaiwu\" rel=\"nofollow\">kaiwu</a>\uff08\u4e00\u4e2a\u672c\u5730\u6a21\u578b\u90e8\u7f72\u5668\uff09\uff0c\u7ed3\u679c\u53d1\u73b0\u2014\u2014<strong>\u7528 Local LLM \u505a\u5f00\u53d1\u7684\u670b\u53cb\uff0c\u591a\u5f97\u8d85\u51fa\u60f3\u8c61</strong>\u3002</p>\n<p>\u5927\u5bb6\u4e0d\u65ad\u63d0\u9700\u6c42\uff1a\u4e0a\u4e0b\u6587\u538b\u7f29\u3001Think \u6a21\u5f0f\u5f00\u5173\u3001\u8054\u7f51\u641c\u7d22\u3001\u5de5\u5177\u8c03\u7528\u2026\u2026</p>\n<p>\u53ef\u8fd9\u4e9b<strong>\u6839\u672c\u4e0d\u662f Ollama \u6216 LM Studio \u7684\u4e8b</strong>\uff01<br/>\n\u5b83\u4eec\u53ea\u8d1f\u8d23\u628a\u6a21\u578b\u8dd1\u8d77\u6765\uff0c\u81f3\u4e8e\u201c\u600e\u4e48\u8ba9\u6a21\u578b\u53d8\u806a\u660e\u201d\u2014\u2014\u90a3\u662f Cursor \u3001Codex \u3001Hermes \u7684\u4e8b\u3002</p>\n<p>\u4f46\u5927\u5382\u4eec\u5728\u5e72\u561b\uff1f</p>\n<ul>\n<li>Cursor \u56f4\u7740\u81ea\u5bb6\u6a21\u578b\u8f6c  </li>\n<li>Codex \u9760\u5356 token \u8d5a\u94b1  </li>\n<li>Hermes \u867d\u5f00\u6e90\uff0c\u5374\u4e0d\u652f\u6301 Windows \u539f\u751f\uff08\u903c\u4f60\u88c5 WSL2 \uff0c\u529d\u9000\u4e00\u534a\u4eba\uff09</li>\n</ul>\n<p><strong>\u5b83\u4eec\u4e0d\u4f1a\u82b1\u7cbe\u529b\u4f18\u5316\u672c\u5730\u5c0f\u6a21\u578b\u3002</strong><br/>\n\u56e0\u4e3a\u672c\u5730\u8dd1\u5f97\u723d\uff0c\u8c01\u8fd8\u4e70\u5b83\u4eec\u7684 API \uff1f</p>\n<p>\u66f4\u522b\u63d0\u90a3\u5835\u5899\u4e86\u2014\u2014<br/>\n\u56fd\u5185\u7f51\u7edc\u65f6\u65ad\u65f6\u7eed\uff0c\u4efb\u52a1\u8dd1\u5230\u4e00\u534a\u65ad\u8fde\uff0c\u4f53\u9a8c\u50cf\u5403\u82cd\u8747\u3002<br/>\n\u60f3\u7528 Claude \uff1f\u5f97\u627e\u4e2d\u8f6c\u3001\u4e70\u6ce8\u6c34\u8d26\u53f7\u3001\u88ab\u6536\u5272\u3001\u8fd8\u88ab\u9119\u89c6\u3002</p>\n<p>\u4f46\u5899\u80fd\u62e6\u4f4f\u8d44\u672c\uff0c\u62e6\u4e0d\u4f4f\u4eba\u6c11\u3002<br/>\n<strong>\u56fd\u9645\u5171\u4ea7\u4e3b\u4e49\u7cbe\u795e\uff0c\u5c31\u4f53\u73b0\u5728\u4e00\u884c\u884c\u5f00\u6e90\u4ee3\u7801\u91cc\u3002</strong></p>\n<hr/>\n<h2>\u75db\u70b9\uff1a\u6211\u4eec\u6bcf\u5929\u90fd\u88ab\u8fd9\u516d\u628a\u5200\u6345</h2>\n<h3>1. \u4e0a\u4e0b\u6587\u592a\u77ed\uff0c\u538b\u7f29\u5c31\u201c\u5931\u5fc6\u201d</h3>\n<ul>\n<li>Opus \u7684 1M \u7a97\u53e3\u7528\u8fc7\u5c31\u56de\u4e0d\u53bb\u4e86\uff0c\u6c38\u8fdc\u4e0d\u7528 compact \u3002  </li>\n<li>\u5c0f\u6a21\u578b\u5728 8G/16G \u663e\u5b58\u4e0a\u53ea\u80fd\u8dd1\u5341\u51e0 K \uff0c\u7a0d\u5fae\u5927\u70b9\u7684\u4efb\u52a1\u76f4\u63a5\u70b8\u3002  </li>\n<li>Hermes \u538b\u7f29\u51e0\u6b21\u5c31\u53d8\u50bb\u5b50\u2014\u2014\u5fd8\u4e86\u81ea\u5df1\u4e24\u8f6e\u524d\u8bf4\u8fc7\u4ec0\u4e48\u3002</li>\n</ul>\n<h3>2. \u7f51\u7edc\u50cf\u4e00\u5835\u5899\uff0c\u5899\u5185\u5916\u90fd\u662f\u5c4e</h3>\n<ul>\n<li>CC / Cursor \u8981\u7a33\u5b9a\u8fde\u6d77\u5916\uff0c\u56fd\u5185\u65ad\u5230\u4f60\u6000\u7591\u4eba\u751f\u3002  </li>\n<li>Hermes \u975e\u8981 WSL2 \uff0cWindows \u539f\u751f\u7528\u6237\u5403\u95ed\u95e8\u7fb9\u3002  </li>\n<li>Web search \u8981\u4e48\u6ca1\u6709\uff0c\u8981\u4e48\u63a5\u5783\u573e\u5546\u5bb6 API \uff0c\u641c\u51fa\u6765\u7684\u5168\u662f SEO \u6c61\u67d3\u7684\u7ed3\u679c\u3002</li>\n</ul>\n<h3>3. \u672c\u5730\u6a21\u578b\u8fde\u5de5\u5177\u90fd\u4e0d\u4f1a\u7528</h3>\n<ul>\n<li>\u7528\u6237\u53cd\u9988\uff1a\u63a5 CC \u6216 Codex \uff0c\u6a21\u578b\u7b28\u5f97\u4e0d\u4f1a\u8c03 tool \u3002  </li>\n<li>8B \u6a21\u578b\u5e72\u5b8c\u6d3b\u4e22\u7ed9\u4f60\u4e00\u4e32\u4ee3\u7801\uff1a\u201c\u81ea\u5df1\u590d\u5236\u53bb\u8fd0\u884c\u201d\u3002  </li>\n<li>\u6211\u662f\u7528 CC \u4e60\u60ef\u7684\u4eba\uff0c\u8fd9\u4f53\u9a8c\u7b49\u4e8e\u8ba9\u6211\u56de\u53bb\u7528\u8bb0\u4e8b\u672c\u5199\u4ee3\u7801\u3002</li>\n</ul>\n<h3>4. \u5c0f\u6a21\u578b\u672c\u8eab\u80fd\u529b\u5c31\u90a3\u6837\uff0c\u4f46 API \u8fd8\u4e0d\u8ba9\u7528</h3>\n<ul>\n<li>8B/14B \u5931\u8bef\u7387\u9ad8\u3001\u7a97\u53e3\u5c0f\u3001\u6ca1\u8054\u7f51\u3001\u9047\u65b0\u95ee\u9898\u5c31\u6b7b\u673a\u3002  </li>\n<li>\u4f60\u4e0d\u53ef\u80fd\u6307\u671b\u5c0f\u5b66\u751f\u89e3\u5fae\u79ef\u5206\u2014\u2014\u8fd9\u662f\u7269\u7406\u89c4\u5f8b\u3002  </li>\n<li>\u53ef A \u5382\u4e0d\u7ed9\u56fd\u4eba\u6ce8\u518c\uff0c\u82b1\u94b1\u4e70\u6ce8\u6c34\u4e2d\u8f6c\uff0c\u50cf\u4ea4\u4fdd\u62a4\u8d39\u3002<br/>\n<strong>\u51ed\u4ec0\u4e48\uff1f</strong></li>\n</ul>\n<h3>5. \u660e\u660e\u672c\u5730\u8fd0\u884c\uff0c\u5374\u662f\u4e2a\u6ca1\u8bb0\u5fc6\u7684\u94a2\u94c1\u5e9f\u6599</h3>\n<ul>\n<li>\u5728\u4e91\u7aef\u4e0d\u8bb0\u4e8b\uff0c\u6211\u8ba4\u4e86\u2014\u2014\u6bd5\u7adf\u6ca1\u82b1\u94b1\u4e70\u5b58\u50a8\u3002  </li>\n<li><strong>\u6211\u90fd\u672c\u5730\u8dd1\u4e86</strong>\uff0c\u786c\u76d8 1T \u8fd8\u80fd\u52a0\uff0c\u4f60\u5374\u53ea\u7ed9\u6211\u4e00\u4e2a markdown \u6587\u4ef6\u5f53\u201c\u8bb0\u5fc6\u201d\uff1f<br/>\n\u8fd9\u5c31\u50cf\u4f60\u4e70\u4e86\u4e00\u53f0\u8d85\u7ea7\u8ba1\u7b97\u673a\uff0c\u7ed3\u679c\u5b83\u6bcf\u6b21\u91cd\u542f\u90fd\u5fd8\u5149\u3002</li>\n</ul>\n<h3>6. \u591a\u6a21\u6001\uff1f\u89c6\u9891\u56fe\u7247\uff1f\u4e0d\u5b58\u5728\u7684</h3>\n<ul>\n<li>\u6a21\u578b\u672c\u8eab\u5f31\uff0c\u4f46\u66f4\u5927\u7684\u95ee\u9898\u662f\u2014\u2014\u6ca1\u6709\u4e13\u95e8\u4f18\u5316\u3002  </li>\n<li>\u95ed\u6e90 API \u4e5f\u4e00\u6837\u70c2\uff0c\u4f46\u4eba\u5bb6\u6536\u94b1\u4e0d\u529e\u4e8b\u3002</li>\n</ul>\n<blockquote>\n<p>\u90e8\u7f72\u96be\u3001\u901f\u5ea6\u6162\u3001\u786c\u4ef6\u8981\u6c42\u9ad8\u8fd9\u4e9b\uff0c\u6211\u4e4b\u524d\u7684 kaiwu + LM + Turbo \u80fd\u89e3\u51b3\u3002<br/>\n\u4eca\u5929\u6211\u4eec\u4e0d\u804a\u8fd9\u4e9b\uff0c\u5c31\u804a<strong>\u600e\u4e48\u8ba9 8B \u6a21\u578b\u8dd1\u51fa Opus \u7684\u4f53\u9a8c</strong>\u3002</p>\n</blockquote>\n<hr/>\n<h2>\u6211\u7684\u9769\u547d\u601d\u8def\uff1a\u4e0d\u7528 CC \u7684\u4f9d\u8d56\u5f3a LLM \u4e32\u884c\uff0c\u6539\u7528 LLM \u505a Gate + \u786e\u5b9a\u6027\u4e13\u5bb6\u7684 MOE \u67b6\u6784</h2>\n<p><strong>\u6838\u5fc3\u7406\u5ff5</strong>\uff1a<br/>\nLLM \u53ea\u8d1f\u8d23\u5f53\u201c\u63a5\u7ebf\u5458\u201d\uff0c\u771f\u6b63\u5e72\u6d3b\u7684\u662f<strong>\u786e\u5b9a\u6027\u4e13\u5bb6</strong>\u2014\u2014<br/>\n\u4e0d\u4f9d\u8d56\u6a21\u578b\u201c\u5565\u90fd\u61c2\u201d\uff0c\u800c\u662f\u8ba9\u6a21\u578b\u53ea\u505a\u4e00\u4ef6\u6781\u5c0f\u3001\u6781\u660e\u786e\u7684\u4e8b\u3002</p>\n<h3>\u539f\u7406\u4e00\uff1aAgentless \u6d41\u6c34\u7ebf\uff08 ICSE 2025 \u6700\u4f73\u8bc1\u660e\uff09</h3>\n<blockquote>\n<p>\u4e0d\u8ba9 LLM \u778e\u51b3\u7b56\uff0c\u7528\u56fa\u5b9a\u6d41\u7a0b \u2192 SWE-bench \u4e0a<strong>\u901a\u8fc7\u7387\u6700\u9ad8\uff0c\u6210\u672c\u6700\u4f4e</strong>\u3002</p>\n</blockquote>\n<p>\u6211\u8bbe\u8ba1\u7684\u6d41\u7a0b\uff08 KWCode \uff09\uff1a\n\u7528\u6237\u8f93\u5165\n\u2514\u2500\u25ba Gate \uff08\u6beb\u79d2\u7ea7\u5206\u7c7b\uff09\n\u2514\u2500\u25ba Locator \uff08\u7cbe\u786e\u5b9a\u4f4d\u6587\u4ef6/\u51fd\u6570\uff09\n\u2514\u2500\u25ba Generator \uff08\u53ea\u6539\u8be5\u6539\u7684\u5730\u65b9\uff09\n\u2514\u2500\u25ba Verifier \uff08\u8bed\u6cd5 + pytest \uff0c\u5931\u8d25\u91cd\u8bd5\uff09</p>\n<p>\u5c0f\u6a21\u578b\u53ea\u9700\u8981\u5728\u5c0f\u7a97\u53e3\u91cc\u505a\u4e00\u4ef6\u4e8b\u2014\u2014<strong>\u5931\u8bef\u7387\u66b4\u8dcc\uff0c\u9519\u8bef\u53ef\u88ab\u5f53\u573a\u6293\u4f4f</strong>\u3002</p>\n<h3>\u539f\u7406\u4e8c\uff1aBM25 + AST \u8c03\u7528\u56fe\u5b9a\u4f4d\uff08\u4e13\u6cbb\u201c\u9690\u85cf\u4f9d\u8d56\u201d\uff09</h3>\n<blockquote>\n<p>\u8bba\u6587 CodeCompass \u53d1\u73b0\u4e00\u4e2a<strong>\u53cd\u5e38\u8bc6</strong>\u4e8b\u5b9e\uff1a<br/>\ncontext \u8d8a\u5927\u7684\u6a21\u578b\uff0c\u53cd\u800c\u8d8a\u5bb9\u6613\u6f0f\u6389<strong>\u67b6\u6784\u4e0a\u5173\u952e\u4f46\u8bed\u4e49\u4e0a\u9065\u8fdc\u7684\u6587\u4ef6</strong>\u2014\u2014\u8fd9\u53eb\u201c\u5bfc\u822a\u6096\u8bba\u201d\u3002</p>\n</blockquote>\n<p>\u5b9e\u9a8c\u6570\u636e\uff08 FastAPI \u771f\u5b9e\u9879\u76ee\uff09\uff1a</p>\n<table>\n<thead>\n<tr>\n<th>\u4efb\u52a1\u7c7b\u578b</th>\n<th>BM25</th>\n<th>\u56fe\u904d\u5386</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>\u6709\u660e\u786e\u5173\u952e\u8bcd</td>\n<td>100%</td>\n<td>\u2014</td>\n</tr>\n<tr>\n<td>\u53ef\u901a\u8fc7 import \u94fe\u627e\u5230</td>\n<td>~85%</td>\n<td>~85%</td>\n</tr>\n<tr>\n<td><strong>\u5b8c\u5168\u65e0\u5173\u952e\u8bcd\u7684\u9690\u85cf\u4f9d\u8d56</strong></td>\n<td>76.2%</td>\n<td><strong>99.4%</strong> \ud83d\ude80</td>\n</tr>\n</tbody></table><p><strong>\u6211\u4eec\u7684\u5b9e\u73b0</strong>\uff1a  </p>\n<ol>\n<li>BM25 \u79d2\u7ea7\u53ec\u56de top-20  </li>\n<li>AST \u8c03\u7528\u56fe\u5c55\u5f00 2 \u8df3\uff08\u5411\u4e0a\u627e\u8c03\u7528\u8005\uff0c\u5411\u4e0b\u627e\u88ab\u8c03\u7528\u8005\uff09  </li>\n<li>\u53d1\u73b0\u90a3\u4e9b\u201c\u540d\u5b57\u548c bug \u65e0\u5173\u4f46\u5b9e\u9645\u662f\u6839\u56e0\u201d\u7684\u9b54\u9b3c\u51fd\u6570  </li>\n</ol>\n<p>\u6280\u672f\u6808\uff1a<code>tree-sitter</code> + <code>rank-bm25</code> + <code>SQLite</code><br/>\n<strong>\u96f6\u4f9d\u8d56\u3001\u96f6 embedding \u3001\u96f6 Docker</strong>\u3002<br/>\n\u652f\u6301\uff1aPython \u00b7 JS \u00b7 TS \u00b7 Java \u00b7 Go \u00b7 Rust</p>\n<h3>\u539f\u7406\u4e09\uff1a\u4e13\u5bb6\u98de\u8f6e\u2014\u2014\u4f60\u7684\u5de5\u5177\u8d8a\u7528\u8d8a\u5f3a\uff0c\u5927\u5382\u6c38\u8fdc\u8ffd\u4e0d\u4e0a</h3>\n<blockquote>\n<p>\u6765\u81ea EE-MCP (NeurIPS 2025) + WLBS \u884c\u4e3a\u56fe\u3002</p>\n</blockquote>\n<p>\u9884\u7f6e 12 \u4e2a\u4e13\u5bb6\uff08\u901a\u7528 7 \u4e2a + \u4e2d\u56fd\u573a\u666f 5 \u4e2a\uff09\u3002<br/>\n<strong>\u7136\u540e\u5f00\u59cb\u98de\u8f6e</strong>\uff1a</p>\n<ul>\n<li>\u540c\u7c7b\u4efb\u52a1\u6210\u529f \u22655 \u6b21 \u2192 \u81ea\u52a8\u751f\u6210\u4e13\u5c5e\u4e13\u5bb6  </li>\n<li>\u65b0\u4e13\u5bb6\u7ecf\u8fc7<strong>\u56de\u6d4b + AB \u6d4b\u8bd5</strong>\u4e09\u9053\u9a8c\u8bc1\u95e8 \u2192 \u6295\u4ea7  </li>\n<li>\u4e0b\u6b21\u540c\u7c7b\u4efb\u52a1\uff0cGate \u76f4\u63a5\u8def\u7531 \u2192 \u66f4\u5feb\u3001\u66f4\u51c6  </li>\n</ul>\n<p>3 \u4e2a\u6708\u540e\uff0c\u4f60\u7684\u4e13\u5c5e\u4e13\u5bb6\u6c60\u2014\u2014<br/>\nCursor \u548c Hermes \u6c38\u8fdc\u8ffd\u4e0d\u4e0a\uff0c\u56e0\u4e3a\u5b83\u4eec<strong>\u65e0\u72b6\u6001</strong>\uff0c\u800c\u4f60\u6709<strong>\u6c38\u4e45\u8bb0\u5fc6</strong>\u3002</p>\n<p>\u4e13\u5bb6\u53ef\u4ee5\u5bfc\u51fa\u3001\u5206\u4eab\u5f62\u6210\u6211\u4eec\u7684\u793e\u533a\u6570\u636e\u8d44\u6e90\u3002</p>\n<h2>\u539f\u7406\u56db\uff1a\u5931\u8d25\u81ea\u52a8\u641c\u7d22\u2014\u2014\u5899\u5185\u7528 Bing \uff0c\u5899\u5916\u7528 DDG</h2>\n<p>Verifier \u8fde\u6302 2 \u6b21 \u2192 \u81ea\u52a8\u89e6\u53d1\u641c\u7d22\uff1a</p>\n<ul>\n<li>\u56fd\u5185\u7f51\u7edc \u2192 Bing \u4e2d\u6587\u7248\uff08 <a href=\"http://cn.bing.com\" rel=\"nofollow\">cn.bing.com</a> \u76f4\u8fde\uff09</li>\n<li>\u6b63\u5e38\u7f51\u7edc \u2192 DuckDuckGo</li>\n<li>\u63d0\u53d6\u6b63\u6587 \u2192 \u538b\u7f29 \u2192 \u6ce8\u5165 context</li>\n</ul>\n<p><strong>\u96f6 API key \uff0c\u96f6\u914d\u7f6e\uff0c\u88c5\u5b8c\u5373\u7528\u3002</strong><br/>\n\u60f3\u66f4\u9690\u79c1\uff1f\u81ea\u5df1\u90e8\u7f72 SearXNG \uff0c\u6570\u636e\u4e0d\u51fa\u7f51\u3002</p>\n<hr/>\n<h2>\u529f\u80fd\u4e00\u89c8\uff08\u4e0d\u662f\u4e3a\u4e86\u70ab\u6280\uff0c\u662f\u4e3a\u4e86\u89e3\u51b3\u4f60\u7684\u6bcf\u4e00\u5929\u7684\u75db\uff09</h2>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u5757</th>\n<th>\u505a\u4e86\u4ec0\u4e48</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>\u4ee3\u7801\u5b9a\u4f4d</td>\n<td>BM25 + AST \u8c03\u7528\u56fe\uff0c99.4% \u547d\u4e2d\u9690\u85cf\u4f9d\u8d56</td>\n</tr>\n<tr>\n<td>\u4ee3\u7801\u4fee\u6539</td>\n<td>\u53ea\u6539 patch \uff0c\u4e0d\u91cd\u5199\u5168\u6587\uff0c\u7cbe\u786e\u5339\u914d</td>\n</tr>\n<tr>\n<td>\u9a8c\u8bc1\u91cd\u8bd5</td>\n<td>\u8bed\u6cd5 + pytest \uff0c\u5931\u8d25\u56de\u6eda\uff0c\u5931\u8d25 2 \u6b21\u5f00\u641c\u7d22</td>\n</tr>\n<tr>\n<td>\u9879\u76ee\u8bb0\u5fc6</td>\n<td><a href=\"http://PROJECT.md\" rel=\"nofollow\">PROJECT.md</a> / <a href=\"http://EXPERT.md\" rel=\"nofollow\">EXPERT.md</a> / <a href=\"http://PATTERN.md\" rel=\"nofollow\">PATTERN.md</a> \u4e09\u5c42\u5206\u79bb\uff0c\u6309\u9700 BM25 \u6ce8\u5165</td>\n</tr>\n<tr>\n<td>\u4e13\u5bb6\u7cfb\u7edf</td>\n<td>12 \u9884\u7f6e + \u98de\u8f6e\u81ea\u751f\u6210 + \u53ef\u5206\u4eab\u5b89\u88c5</td>\n</tr>\n<tr>\n<td>\u4e2d\u56fd\u672c\u5730\u5316</td>\n<td>\u81ea\u52a8\u5207 ModelScope / \u6e05\u534e\u955c\u50cf / Bing \u641c\u7d22 / Windows \u539f\u751f</td>\n</tr>\n</tbody></table><hr/>\n<h2>\u6211\u4eec\u548c\u201c\u5b83\u4eec\u201d\u7684\u4e0d\u4e00\u6837</h2>\n<table>\n<thead>\n<tr>\n<th>\u573a\u666f</th>\n<th>\u5176\u4ed6\u5de5\u5177</th>\n<th><strong>KWCode \uff08\u6211\u4eec\uff09</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Windows</td>\n<td>\u903c\u4f60\u88c5 WSL2</td>\n<td><strong>cmd / PowerShell \u539f\u751f\u8dd1</strong></td>\n</tr>\n<tr>\n<td>\u6a21\u578b\u4e0b\u8f7d</td>\n<td>HuggingFace \u88ab\u5899</td>\n<td>\u81ea\u52a8\u5207 <strong>ModelScope</strong></td>\n</tr>\n<tr>\n<td>pip \u5b89\u88c5</td>\n<td>PyPI \u6162\u6b7b</td>\n<td>\u81ea\u52a8\u5207 <strong>\u6e05\u534e/\u963f\u91cc\u955c\u50cf</strong></td>\n</tr>\n<tr>\n<td>\u641c\u7d22\u589e\u5f3a</td>\n<td>DDG \u88ab\u5899</td>\n<td>\u81ea\u52a8\u5207 <strong>Bing \u4e2d\u6587\u7248</strong></td>\n</tr>\n<tr>\n<td>\u63a8\u8350\u6a21\u578b</td>\n<td>GPT / Claude \uff08\u8981\u94b1/\u8981\u68af\u5b50\uff09</td>\n<td><strong>DeepSeek \u00b7 Qwen \u00b7 GLM</strong>\uff08\u56fd\u4ea7\u514d\u8d39\uff09</td>\n</tr>\n</tbody></table><hr/>\n<h2>\u540c\u5fd7\u4eec\uff0c\u8fd9\u4e0d\u662f\u4e00\u4e2a\u4eba\u7684\u6218\u6597</h2>\n<p>\u6211\u53ea\u6709\u4e00\u53f0 5060 8G \u663e\u5b58 16G \u5185\u5b58\u5c0f\u7834\u7535\u8111\uff0c\u786c\u76d8\u8fd8\u65f6\u597d\u65f6\u574f\uff0c\u82b1\u94b1\u4e70 api \u4e00\u4e2a\u6708\u4e09\u56db\u5343\u3002\n\u6211\u60f3\u8981\u4eba\u4eba\u4e3a\u9f99\u65f6\u4ee3\uff0c\u800c\u4e0d\u662f api \u72ec\u5927\u65f6\u4ee3\u3002\n\u6240\u4ee5\u6211\u60f3\u6253\u9020\n\u4e00\u4e2a<strong>\u771f\u6b63\u5c5e\u4e8e\u5f00\u6e90\u793e\u533a\u3001\u4e0d\u4f9d\u8d56\u5927\u5382 API \u3001\u4e0d\u88ab\u5899\u3001\u8ba9 8B \u6a21\u578b\u4e5f\u80fd\u5e72\u7ffb Opus</strong> \u7684 Coding Agent \u3002</p>\n<p><strong>\u6211\u4eec\u6709\u8bba\u6587\u652f\u6491\uff0c\u6709\u539f\u578b\u4ee3\u7801\uff0c\u6709\u6ee1\u8154\u6012\u706b\u548c\u70ed\u8840\u3002</strong><br/>\n\u73b0\u5728\u8fd8\u7f3a\u4f60\u2014\u2014<br/>\n\u7f3a\u6bcf\u4e00\u4e2a\u53d7\u591f\u4e86\u88ab\u6536\u5272\u3001\u88ab\u6b67\u89c6\u3001\u88ab\u7f51\u7edc\u66b4\u529b\u7684\u5f00\u53d1\u8005\u3002</p>\n<p>GitHub \u4ed3\u5e93\u8fd1\u671f\u5f00\u653e\uff0c\u4ee3\u7801\u5b8c\u5168\u5f00\u6e90\u3002<br/>\n\u4f60\u53ef\u4ee5\uff1a</p>\n<ul>\n<li>\u8d21\u732e\u4ee3\u7801\uff08 Rust/Python/TS \u90fd\u884c\uff09</li>\n<li>\u5206\u4eab\u4f60\u7684\u4e13\u5c5e\u4e13\u5bb6\uff08.kwx \u6587\u4ef6\uff09</li>\n<li>\u63d0 bug \u3001\u5199\u6587\u6863\u3001\u5ba3\u4f20\u51fa\u53bb</li>\n<li>\u6216\u8005\u53ea\u662f\u53bb\u70b9\u4e00\u4e2a \u2b50\uff0c\u8ba9\u66f4\u591a\u4eba\u770b\u89c1</li>\n</ul>\n<p><strong>\u56fd\u9645\u5171\u4ea7\u4e3b\u4e49\u7cbe\u795e\uff0c\u4ece\u4e00\u884c\u5f00\u6e90\u4ee3\u7801\u5f00\u59cb\u3002</strong><br/>\n<strong>\u8ba9\u5927\u5382\u53bb\u5356 token \u5427\uff0c\u6211\u4eec\u6709\u81ea\u5df1\u7684\u5de5\u5177\u4e86\u3002</strong></p>\n<hr/>\n<h2>\u884c\u52a8\u53f7\u53ec</h2>\n<p>\ud83d\udc49 \u6709\u6ca1\u6709\u66f4\u597d\u7684\u601d\u8def\u548c\u8def\u5f84\uff0c\u4e0a\u8ff0\u53ea\u662f\u6211\u4e2a\u4eba\u7814\u7a76<br/>\n\ud83d\udc49 \u540e\u7eed\u5728\u672c\u94fe\u63a5\u53d1\u5e03 github \uff0c\u6b22\u8fce fork \u7ee7\u7eed\u6df1\u6316</p>\n<p><strong>\u4e0d\u8981\u8ba9\u8d44\u672c\u5b9a\u4e49\u201c\u53ef\u80fd\u201d\u4e0e\u201c\u4e0d\u53ef\u80fd\u201d\u3002</strong><br/>\n<strong>\u6211\u4eec\u8bf4\u4e86\u7b97\u3002</strong>\n<strong>\u6216\u8bb8\u5f88\u5feb\uff0c8B \u6a21\u578b\u771f\u80fd\u8dd1\u8d62 OPUS \uff0c\u6240\u6709\u4eba\u90fd\u80fd\u62e5\u6709\u72ec\u5c5e\u4e8e\u81ea\u5df1\u7684\u667a\u80fd\u4f53</strong></p>\n<p>\u8981\u4e0d\u8981\u5148\u5efa\u4e2a\u7fa4\uff0c\u7b97\u4e86 \u6211\u793e\u6050 \u4e0d\u4f1a\u7ef4\u62a4\uff0c\u6709\u4e8b\u54b1\u4eec\u8fd9\u4e2a\u94fe\u63a5\u804a\u628a</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/diudiuu", 
        "name": "diudiuu", 
        "avatar": "https://cdn.v2ex.com/avatar/53c0/4118/1055_large.png?m=1780363419"
      }, 
      "url": "https://www.v2ex.com/t/1208894", 
      "title": "\u81ea\u5df1\u505a\u4e86\u4e00\u6b3e\u5728\u7ebf GPU \u63a8\u7406\u901f\u5ea6\u8ba1\u7b97\u5668 \u00b7 TPS Calculator", 
      "id": "https://www.v2ex.com/t/1208894", 
      "date_published": "2026-04-27T09:34:17+00:00", 
      "content_html": "<h1>TPS Calculator \u00b7 GPU \u63a8\u7406\u901f\u5ea6\u8ba1\u7b97\u5668</h1>\n<blockquote>\n<p>\u4e70\u4e0d\u8d77\u673a\u5b50\uff0c\u6240\u4ee5\u505a\u4e86\u8fd9\u4e2a\u3002</p>\n</blockquote>\n<p>\u5728\u7ebf\u5730\u5740\uff1a<a href=\"https://tps.bunai.cc\" rel=\"nofollow\">tps.bunai.cc</a> </p>\n<hr/>\n<h2>\u7a81\u53d1\u5947\u60f3\u8d76\u7d27\u8bb0\u5f55\u4e0b\u6765\uff0c\u76f4\u63a5 vibe code \uff0c\u8bf4\u6572\u5c31\u5199</h2>\n<p>\u4e00\u4e2a vibe code \u51fa\u6765\u7684 GPU \u63a8\u7406\u6027\u80fd\u4f30\u7b97\u5de5\u5177\u3002</p>\n<p>\u8d77\u56e0\u5f88\u7b80\u5355\u2014\u2014\u663e\u5361\u592a\u8d35\uff0c\u4e70\u4e0d\u8d77\uff0c\u60f3\u8dd1\u4e2a\u6a21\u578b\u53c8\u4e0d\u77e5\u9053\u81ea\u5df1\u7684\u914d\u7f6e\u591f\u4e0d\u591f\uff0c\n\u4e8e\u662f\u628a\u7f51\u4e0a\u6563\u843d\u7684\u53c2\u6570\u548c\u516c\u5f0f\u6c47\u603b\u4e86\u4e00\u4e0b\uff0c\u505a\u6210\u4e86\u8fd9\u4e2a\u8ba1\u7b97\u5668\u3002</p>\n<p>\u8f93\u5165\u663e\u5361\u578b\u53f7\u3001\u6a21\u578b\u3001\u91cf\u5316\u65b9\u5f0f\u548c\u8fd0\u884c\u53c2\u6570\uff0c\u5feb\u901f\u4f30\u7b97\uff1a</p>\n<ul>\n<li>\u663e\u5b58\u5360\u7528\u4e0e OOM \u98ce\u9669</li>\n<li>Decode / Prefill token/s</li>\n<li>TTFT / TPOT / \u603b\u65f6\u5ef6</li>\n<li>\u5e26\u5bbd\u74f6\u9888\u8fd8\u662f\u7b97\u529b\u74f6\u9888</li>\n<li>\u591a\u5361 TP \u901a\u4fe1\u6548\u7387</li>\n</ul>\n<hr/>\n<h2>\u9002\u5408\u5e72\u4ec0\u4e48</h2>\n<p>\u2705 \u5728\u4e70\u673a\u5b50 / \u79df\u5361\u4e4b\u524d\uff0c\u5148\u5927\u6982\u9884\u4f30\u4e00\u4e0b\u8dd1\u4e0d\u8dd1\u5f97\u8d77\u6765<br/>\n\u2705 \u5b66\u4e60\u63a8\u7406\u6027\u80fd\u5efa\u6a21\uff0c\u7406\u89e3\u91cf\u5316\u3001KV Cache \u3001TP \u3001Roofline \u8fd9\u4e9b\u6982\u5ff5<br/>\n\u2705 \u505a\u65b9\u6848\u521d\u7b5b\u548c\u53c2\u6570\u5bf9\u6bd4  </p>\n<p>\u274c \u4e0d\u9002\u5408\u76f4\u63a5\u66ff\u4ee3\u771f\u5b9e benchmark<br/>\n\u274c \u4e0d\u9002\u5408\u628a\u4f30\u7b97\u503c\u5f53\u4f5c\u751f\u4ea7\u627f\u8bfa<br/>\n\u274c Mac \u7535\u8111\u6ca1\u6709\u653e\u51fa\u6765\uff0c\u9a8c\u8bc1\u4e86\u4e00\u4e0b\u5dee\u8ddd\u6709\u70b9\u5927\uff0c\u5148\u653e\u4e00\u653e</p>\n<hr/>\n<h2>\u53c2\u8003\u8d44\u6599</h2>\n<ul>\n<li>\u6a21\u578b\u53c2\u6570\u6765\u6e90\uff1a<a href=\"https://huggingface.co\" rel=\"nofollow\">HuggingFace</a> model cards \u53ca <a href=\"https://ollama.com\" rel=\"nofollow\">Ollama</a> \u5b98\u65b9\u9875\u9762  </li>\n<li>MoE CPU Offload \u573a\u666f\u53c2\u8003\uff1a<a href=\"https://github.com/val1813/kaiwu\" rel=\"nofollow\">val1813/kaiwu</a></li>\n<li>\u81ea\u5df1\u642d\u5efa\u6a21\u578b<a href=\"https://2libra.com/post/ai-applications/ovZiTd0\" rel=\"nofollow\">Gemma4 26b</a> </li>\n<li>\u81ea\u5df1\u642d\u5efa\u6a21\u578b<a href=\"https://2libra.com/post/ai-applications/KT_9AES\" rel=\"nofollow\">Gemma4 31b</a></li>\n<li>\u8fd8\u6709\u4e2a 4070ti \u5f97\u6570\u636e</li>\n</ul>\n<p>\u8fd9\u5957\u516c\u5f0f\u548c\u53c2\u6570\u662f\u6211\u81ea\u5df1\u6574\u7406\u6c47\u603b\u7684\uff0c\u6ca1\u6709\u5927\u91cf\u771f\u673a\u8dd1\u8fc7\u9a8c\u8bc1\u3002\n\u5982\u679c\u4f60\u624b\u4e0a\u6709\u771f\u5b9e\u7684\u6d4b\u8bd5\u6570\u636e\uff0c\u53d1\u73b0\u54ea\u91cc\u4f30\u7b97\u504f\u5dee\u5927\u3001\u516c\u5f0f\u6709\u95ee\u9898\uff0c\n<strong>\u6b22\u8fce\u5f00 Issue \u6216 PR \u6307\u51fa\u6765</strong>\uff0c\u5927\u5bb6\u4e00\u8d77\u5b66\u4e60\uff0c\u4e00\u8d77\u628a\u8fd9\u4e2a\u4e1c\u897f\u505a\u5f97\u66f4\u51c6\u3002</p>\n<p><strong>\u5e0c\u671b\u6709\u771f\u5b9e\u6570\u636e\u7684\u5927\u4f6c\u5e2e\u5fd9\u6307\u6b63</strong>\uff0c\u8c22\u8c22\uff01\ud83d\ude4f</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/zsj1029", 
        "name": "zsj1029", 
        "avatar": "https://cdn.v2ex.com/avatar/38a0/9f4d/189266_large.png?m=1777977403"
      }, 
      "url": "https://www.v2ex.com/t/1208824", 
      "date_modified": "2026-04-27T23:22:04+00:00", 
      "content_html": "\u641e\u4e86\u4e00\u4e0a\u5348\uff0c\u672c\u5730 a100 40g \uff0c\u8f93\u51fa\u4e5f\u6162 40t/s<br />\u5927\u6982\u7684\u63d0\u793a\u8bcd\u52b3\u529b\u58eb\u98ce\u683c\uff0c\u7f57\u9a6c\u6570\u5b57\uff0c\u6708\u76f8\u65e5\u5386\uff0c\u9ad8\u8d35\u5178\u96c5<br /><a target=\"_blank\" href=\"https://i.imgur.com/PbqajDk.png\" rel=\"nofollow noopener\" target=\"_blank\"><img src=\"https://i.imgur.com/PbqajDk.png\" class=\"embedded_image\" rel=\"noreferrer\"></a><br /><br />\u6708\u76f8\u90a3\u5757\u641e\u4e86\u597d\u591a\u8f6e<br /><br />\u7ed3\u8bba:<br />\u5c0f\u53c2\u6570\u7684\u6a21\u578b\u667a\u529b\u4e0d\u5dee\uff0cTrae IDE agent \u8fde\u63a5\u672c\u5730\u6a21\u578b\uff0ccoding \u5b8c\u5168\u53ef\u7528", 
      "date_published": "2026-04-27T06:36:20+00:00", 
      "title": "qwen3.6 27b \u672c\u5730\u7f16\u7801\u6d4b\u8bd5", 
      "id": "https://www.v2ex.com/t/1208824"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/joeue404", 
        "name": "joeue404", 
        "avatar": "https://cdn.v2ex.com/gravatar/7487abdb33e5648e84c20ea8dee7821a?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1208804", 
      "title": "xllm \u771f\u7684\u6bd4 vllm+plugin \u6027\u80fd\u597d\u4e48\uff1f", 
      "id": "https://www.v2ex.com/t/1208804", 
      "date_published": "2026-04-27T05:46:35+00:00", 
      "content_html": ""
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/Hermitist", 
        "name": "Hermitist", 
        "avatar": "https://cdn.v2ex.com/gravatar/da6c1e355f86d79cd2887cb34a3c864e?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1208567", 
      "date_modified": "2026-04-26T01:02:44+00:00", 
      "content_html": "27B/31B \u751a\u81f3 35B \u7684 4bit \u90fd\u53ef\u4ee5, \u6d4b\u8bd5\u4e86\u597d\u4e45, \u4e5f\u4e0b\u8f7d\u4e86\u51e0\u5341\u4e2a\u4e86,\u90fd\u4e0d\u592a\u884c, \u611f\u89c9\u964d\u667a\u4e86, \u8fd9\u4e9b\u521a\u51fa\u6765\u7684\u65f6\u5019\u6211\u8fd9\u4e2a\u914d\u7f6e\u80fd\u8dd1\u5230 35tokens/s.<br /><br /><br />\u51c6\u5907\u76f4\u63a5\u6284\u4f5c\u4e1a, \u8bf7\u7ed9 huggingface \u8fde\u63a5, \u6211\u7684\u672c\u5730\u63a8\u7406\u6846\u67b6\u662f omlx, \u611f\u8c22\u611f\u8c22.", 
      "date_published": "2026-04-26T00:16:06+00:00", 
      "title": "\u5404\u4f4d\u63a8\u8350\u4e00\u4e2a 32G Macbook air M5 \u53ef\u4ee5\u8dd1\u7684 moe \u6a21\u578b", 
      "id": "https://www.v2ex.com/t/1208567"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/evegod", 
        "name": "evegod", 
        "avatar": "https://cdn.v2ex.com/avatar/78f2/2b39/349991_large.png?m=1776690822"
      }, 
      "url": "https://www.v2ex.com/t/1208556", 
      "title": "\u6211\u7684\u5f00\u6e90\u9879\u76ee\uff0c\u6b22\u8fce\u5927\u5bb6\u4f7f\u7528\u548c\u6279\u8bc4\uff0c\u672c\u5730\u65e0\u5b57\u5178\u5b57\u7b26\u578b\u6a21\u578b\u8bad\u7ec3\u67b6\u6784\u4ee3\u7801\u5b8c\u5168\u5f00\u6e90\uff0c\u53ef\u5f62\u6210\u8bed\u4e49\u7ed3\u6784", 
      "id": "https://www.v2ex.com/t/1208556", 
      "date_published": "2026-04-25T17:17:46+00:00", 
      "content_html": "<p>\u6b22\u8fce\u6279\u8bc4\uff0c\u4e5f\u662f vibe coding \u7684\u4ea7\u7269\uff0c\u6211\u662f\u5728\u5c1d\u8bd5\u5b66\u4e60\u6570\u5b66\u548c\u7269\u7406\u76f8\u5173\u7406\u8bba\u7684\u65f6\u5019\u7ed3\u5408\u7f16\u7801\u5b66\u7684\u4e00\u4e9b\u81ea\u5df1\u7684\u770b\u6cd5\u5728\u505a\u5b9e\u9a8c\uff0c\u5f53\u7136\u5b9e\u9a8c\u5185\u5bb9\u5927\u90e8\u5206\u4e5f\u662f vibe coding \u7684\u4ea7\u7269\uff0c\u73b0\u6709\u57fa\u51c6\u662f\u8fd9\u4e2a\u6a21\u578b\u5728\u672c\u5730\u5b66\u4e60 fineweb \u6570\u636e\u96c6\uff0c\u67b6\u6784\u6ca1\u6709\u8bcd\u5178\u5c42\uff0c\u53ea\u6709\u5b57\u7b26\u5b66\u4e60\u548c\u76f8\u5173\u7eaf\u6570\u5b66\u67b6\u6784\u548c\u7f16\u7801\u5c1d\u8bd5\u7684\u60c5\u51b5\u4e0b\u53ef\u4ee5\u6d8c\u73b0\u7c7b\u82f1\u8bed\u8bed\u4e49\u7ed3\u6784\uff0c\u800c\u4e14\u8bad\u7ec3\u548c\u5c55\u5f00\u8f93\u51fa\u5747\u662f\u663e\u5b58\u548c\u5185\u5b58\u4f18\u5316\u5f62\u5f0f\u7684\uff0c\u5927\u5bb6\u53ef\u4ee5\u5c1d\u8bd5\u81ea\u5df1\u5206\u6790\u548c\u4f7f\u7528\u4e00\u4e0b\uff0c\u76f8\u5173\u7684\u601d\u8003\u65b9\u5f0f\u548c\u67b6\u6784\u672c\u8eab\u4e5f\u5728\u4ee3\u7801\u4e2d\u6ce8\u91ca\u4e86\uff0c\u5982\u679c\u7528\u5176\u4ed6 ai \u53bb\u5206\u6790\u8be5\u9879\u76ee\u4f1a\u5bf9\u5176\u6570\u5b66\u7ed3\u6784\u6709\u4e0d\u540c\u770b\u6cd5\uff0c\u5f53\u7136\u53ef\u80fd\u662f\u6211\u7684\u601d\u8003\u89d2\u5ea6\u5bfc\u81f4\u6211\u7684\u7528\u8bed\u548c\u63d0\u793a\u8bcd\u5bfc\u81f4\u5176\u7ed3\u6784\u504f\u79fb\u548c\u6211\u7684\u7528\u8bed\u6ca1\u6709\u5e7f\u6cdb\u88ab\u63a5\u53d7\u7684\u95ee\u9898\u3002\u8bf7\u5927\u5bb6\u6279\u8bc4\u6307\u6b63\uff0c\u6211\u5c3d\u529b\u63d0\u9ad8\u6211\u81ea\u5df1\u3002\n\u9879\u76ee\u5730\u5740\uff1a <a href=\"https://github.com/makai891124-prog/H2Q-MicroStream\" rel=\"nofollow\">https://github.com/makai891124-prog/H2Q-MicroStream</a></p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/jamme", 
        "name": "jamme", 
        "avatar": "https://cdn.v2ex.com/avatar/4ec7/2727/199115_large.png?m=1769231137"
      }, 
      "url": "https://www.v2ex.com/t/1208538", 
      "date_modified": "2026-04-25T15:16:27+00:00", 
      "content_html": "<p>\u4e3b\u8981\u662f\u7528\u6765\u90e8\u7f72<code>YOLO26</code>\u505a\u6570\u636e\u96c6\u8bad\u7ec3\u548c\u76ee\u6807\u68c0\u6d4b\u6216\u8ffd\u8e2a\u7684\uff0c\u56fe\u7247\u6570\u636e\u6682\u5b9a 5000 \u5f20\uff08\u5176\u5b9e\u6570\u636e\u6709\u5f88\u591a\uff0c\u4f46\u662f\u6682\u5b9a\u7528\u4e8e\u8bad\u7ec3\u7684\u6570\u636e\u4e0a\u9650\u662f 5000 \u5f20\uff09\u3002</p>\n<p>\u76ee\u524d\u6709\u4e00\u53f0 RX6600xt \uff0c\u4f46\u662f directML \u597d\u50cf\u4e5f\u4e0d\u80fd\u4f7f\u8fd9\u5f20\u5361\u53c2\u4e0e\u8bad\u7ec3\u8ba1\u7b97\uff0c\u4e0a\u7f51\u67e5\u4e86\u4e00\u4e0b\u597d\u50cf\u662f\u5bf9 7000 \u7cfb\u5217\u4ee5\u4e0a\u7684\u663e\u5361\u652f\u6301\u7684\u66f4\u597d\u4e00\u4e9b\u3002</p>\n<p>\u6240\u4ee5\u8001\u677f\u7684\u610f\u601d\u662f\u91cd\u65b0\u914d\u4e00\u53f0 N \u5361\u4e3b\u673a\uff0c\u4f46\u6211\u4e4b\u524d\u6ca1\u6709\u4f7f\u7528 YOLO \u8bad\u7ec3\u7684\u7ecf\u9a8c\uff0c\u4e0d\u77e5\u9053\u76ee\u524d\u8fd9\u4e2a\u6570\u91cf\u7ea7\u7684\u6570\u636e\u8bad\u7ec3\u4ee5\u53ca\u8fd9\u4e2a\u4f53\u91cf\u7684\u6a21\u578b\u8be5\u4f7f\u7528\u4ec0\u4e48\u5361\u3002\u54a8\u8be2\u5b98\u7f51 AI \u7684\u8bdd\uff0c\u5c31\u662f\u65e0\u8111\u63a8\u8350 4090 \u30015090 \u8fd9\u79cd\u5927\u663e\u5b58\u7684\u5361\u3002\u641e\u5f97\u6211\u5f88\u5934\u75bc~</p>\n<p>\u5173\u4e8e\u9884\u7b97\u7684\u8bdd\uff0c\u8001\u677f\u53ea\u8bf4\u4e86\u4e00\u53e5\u4f60\u770b\u7740\u529e\u5427\u3002\u4f46\u4e4b\u524d\u8001\u677f\u7684\u610f\u601d\u662f\u8ba9\u6211\u770b\u770b\u80fd\u4e0d\u80fd\u628a\u73b0\u5728\u8fd9\u53f0\u4e3b\u673a\u7684\u663e\u5361\u6362\u6210 RTX5070 \uff0c\u540e\u6765\u6211\u67e5\u4e86\u4e00\u4e0b\u73b0\u5728\u4e3b\u673a\u7684\u7535\u6e90\uff0c\u624d 500W \uff0c\u5e26\u4e0d\u52a8 5070 \uff0c\u624d\u6709\u4e86\u914d\u65b0\u4e3b\u673a\u7684\u8fd9\u4ef6\u4e8b\u3002\u6240\u4ee5\u6211\u60f3\u7740\u5199\u4e2a\u4e24\u4e09\u5957\u914d\u7f6e\u5355\u7ed9\u8001\u677f\u770b\uff0c\u4f4e\u914d\u9ad8\u914d\u90fd\u5199\u4e00\u4e0b\uff0c\u8ba9\u8001\u677f\u51b3\u5b9a\u9009\u4ec0\u4e48\u3002</p>\n<p>\u6709\u6ca1\u6709\u6709<code>YOLO \u8bad\u7ec3+\u76ee\u6807\u68c0\u6d4b\u7ecf\u9a8c</code>\u7684 V \u53cb\u7ed9\u70b9\u5efa\u8bae\uff1f\u8dea\u8c22\u4e86~</p>\n", 
      "date_published": "2026-04-25T15:02:26+00:00", 
      "title": "\u8bf7\u6559\u4e00\u4e2a\u5173\u4e8e\u6a21\u578b\u8bad\u7ec3\u4e3b\u673a\u914d\u7f6e\u7684\u95ee\u9898", 
      "id": "https://www.v2ex.com/t/1208538"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/KaiWuBOSS", 
        "name": "KaiWuBOSS", 
        "avatar": "https://cdn.v2ex.com/gravatar/243db3a31aa62a02d726471a3fd1782e?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1208365", 
      "title": "\u6211\u505a\u4e86\u4e2a\u5de5\u5177\u8ba9 8GB \u663e\u5361\u8dd1 30B \u6a21\u578b\u4ece 3 tok/s \u63d0\u5230 21 tok/s\uff0c\u8bb0\u5f55\u4e00\u4e0b\u6280\u672f\u53d1\u73b0", 
      "id": "https://www.v2ex.com/t/1208365", 
      "date_published": "2026-04-24T10:51:29+00:00", 
      "content_html": "<p>\u6700\u8fd1\u5728\u6298\u817e\u672c\u5730\u5927\u6a21\u578b\uff0c\u53d1\u73b0\u4e00\u4e2a\u6838\u5fc3\u95ee\u9898\uff1aOllama \u548c LM Studio \u80fd\u8ba9\u6a21\u578b\u8dd1\u8d77\u6765\uff0c\u4f46\u53c2\u6570\u5168\u9760\u731c\u2014\u2014\u4e0a\u4e0b\u6587\u957f\u5ea6\u3001KV cache \u7c7b\u578b\u3001MoE expert \u653e\u54ea\u3001ubatch \u591a\u5927\u2026\u2026\u7528\u9ed8\u8ba4\u53c2\u6570\u57fa\u672c\u662f\u5728\u6d6a\u8d39\u663e\u5361\u3002</p>\n<p>\u4e8e\u662f\u505a\u4e86\u4e2a\u5de5\u5177\u81ea\u52a8\u627e\u6700\u4f18\u914d\u7f6e\uff0c\u8fc7\u7a0b\u4e2d\u8e29\u4e86\u4e0d\u5c11\u5751\uff0c\u8bb0\u5f55\u4e00\u4e0b\u3002</p>\n<hr/>\n<h2>\u6838\u5fc3\u53d1\u73b0</h2>\n<h3>1. MoE \u6a21\u578b\u7684 offload \u7b56\u7565\u51b3\u5b9a\u4e86\u4e00\u5207</h3>\n<p>Qwen3-30B-A3B \u662f MoE \u67b6\u6784\uff0c\u5728 8GB \u663e\u5361\u4e0a\uff1a</p>\n<ul>\n<li>LM Studio \u9ed8\u8ba4\u628a\u6240\u6709\u5c42\u585e\u8fdb\u663e\u5b58 \u2192 7549MB \uff08 93%\uff09\uff0c3 tok/s</li>\n<li>\u53ea\u628a attention \u5c42\u653e GPU \uff0cMoE expert \u5c42\u8d70 CPU \u2192 2603MB \uff08 32%\uff09\uff0c21 tok/s</li>\n</ul>\n<p>\u5feb\u4e86 7 \u500d\uff0c\u663e\u5b58\u53cd\u800c\u7701\u4e86 65%\u3002\u5173\u952e\u662f llama.cpp \u652f\u6301\u8fd9\u4e2a\uff0c\u4f46\u4f60\u5f97\u81ea\u5df1\u8bc6\u522b\u54ea\u4e9b tensor \u662f MoE expert \uff08<code>.ffn_.*_exps.</code> \u8fd9\u7c7b\u547d\u540d\uff09\uff0c\u7136\u540e\u624b\u52a8\u914d\u3002</p>\n<h3>2. KV cache \u7c7b\u578b\u5f71\u54cd\u6bd4\u5927\u591a\u6570\u4eba\u60f3\u7684\u5927</h3>\n<p>\u540c\u4e00\u5f20 8GB \u663e\u5361\u8dd1 Llama 3.1 8B \uff0c\u4e0d\u540c KV cache \u914d\u7f6e\u901f\u5ea6\u5dee\u5f02\uff1a</p>\n<table>\n<thead>\n<tr>\n<th>\u914d\u7f6e</th>\n<th>ctx</th>\n<th>\u901f\u5ea6</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>iso3+iso3 \uff0c4 slot</td>\n<td>8K</td>\n<td>19.4 tok/s</td>\n</tr>\n<tr>\n<td>q8_0+q4_0 \uff0c1 slot</td>\n<td>8K</td>\n<td>38.2 tok/s</td>\n</tr>\n<tr>\n<td>f16+f16 \uff0c1 slot</td>\n<td>8K</td>\n<td>51.7 tok/s</td>\n</tr>\n<tr>\n<td>f16+f16 \uff0c1 slot \uff08\u81ea\u52a8\uff09</td>\n<td>64K</td>\n<td>26.2 tok/s</td>\n</tr>\n</tbody></table><p>f16 \u6bd4 iso3 \u5feb\u5c06\u8fd1 3 \u500d\u3002\u4f46 f16 \u663e\u5b58\u5360\u7528\u66f4\u5927\uff0c\u6240\u4ee5\u6b63\u786e\u7b56\u7565\u662f\uff1a\u5148\u7b97 f16 KV cache \u5360\u591a\u5c11\u663e\u5b58\uff0c\u88c5\u5f97\u4e0b\u5c31\u7528 f16 \uff0c\u88c5\u4e0d\u4e0b\u518d\u964d\u7ea7\u3002</p>\n<p>\u516c\u5f0f\uff1a<code>KV_MB = 2 \u00d7 layers \u00d7 kv_heads \u00d7 head_dim \u00d7 ctx \u00d7 bytes / 1024\u00b2</code></p>\n<h3>3. oobabooga \u516c\u5f0f\u7528\u6765\u9884\u6d4b ctx \u4e0a\u9650</h3>\n<p>\u793e\u533a\u91cc\u6d41\u4f20\u7684 oobabooga \u663e\u5b58\u4f30\u7b97\u516c\u5f0f\uff0c\u539f\u672c\u7528\u6765\u9884\u6d4b\u88c5\u8f7d\u6a21\u578b\u540e\u5269\u4f59\u663e\u5b58\u80fd\u652f\u6301\u591a\u5927 ctx \u3002\u4f46\u8fd9\u4e2a\u516c\u5f0f\u662f\u57fa\u4e8e q8_0/f16 \u62df\u5408\u7684\uff0c\u7528 iso3 \u7684\u65f6\u5019\u4f1a\u4e25\u91cd\u9ad8\u4f30\u663e\u5b58\u9700\u6c42\uff0c\u5bfc\u81f4 ctx \u53ea\u7b97\u51fa 4K \u3002</p>\n<p>\u6700\u540e\u653e\u5f03\u516c\u5f0f\u9884\u6d4b\uff0c\u6539\u6210\u4e8c\u5206\u63a2\u6d4b\uff1a\u4ece min(nativeCtx, 65536) \u5f00\u59cb\uff0cOOM \u5c31\u51cf\u534a\uff0c\u6700\u591a\u63a2 5 \u6b21\uff0c\u8ba9 llama-server \u81ea\u5df1\u544a\u8bc9\u6211\u80fd\u8dd1\u591a\u5c11\u3002Llama 3.1 8B \u7684 ctx \u4ece 4K \u76f4\u63a5\u5230 64K \u3002</p>\n<h3>4. parallel slot \u6570\u91cf\u5bf9\u5355\u7528\u6237\u573a\u666f\u5f71\u54cd\u5de8\u5927</h3>\n<p>llama.cpp \u9ed8\u8ba4\u5f00 4 \u4e2a\u5e76\u884c slot \uff08\u4e3a\u4e86\u591a\u7528\u6237\u5e76\u53d1\uff09\uff0c\u4f46\u5355\u7528\u6237\u573a\u666f\u4e0b\u8fd9\u4f1a\u628a VRAM \u5206\u6210 4 \u4efd\u3002</p>\n<p>\u5173\u6389\u591a\u4f59 slot \uff08<code>--parallel 1</code>\uff09\u4e4b\u540e\uff1a18.5 \u2192 38.2 tok/s \uff0c\u76f4\u63a5\u7ffb\u500d\u3002</p>\n<h3>5. ubatch \u5b9e\u6d4b\u6bd4\u7406\u8bba\u66f4\u53ef\u9760</h3>\n<p>ubatch 128 vs 512 \u7684\u6027\u80fd\u5dee\u5f02\u8ddf\u6a21\u578b\u548c\u663e\u5361\u90fd\u6709\u5173\u7cfb\uff0c\u6ca1\u6709\u901a\u7528\u6700\u4f18\u503c\u3002\u5b9e\u6d4b\u7ed3\u8bba\uff1a</p>\n<ul>\n<li>8K ctx\uff1aubatch 512 \u6bd4 128 \u5feb 7.6%</li>\n<li>64K ctx\uff1aubatch 512 \u6bd4 128 \u5feb 21.6%</li>\n</ul>\n<p>\u76f4\u63a5 benchmark \u4e24\u4e2a\u503c\u53d6\u5feb\u7684\uff0c\u6bd4\u67e5\u6587\u6863\u731c\u9760\u8c31\u3002</p>\n<h3>6. \u5bf9\u8bdd\u538b\u7f29\u4e0d\u8981\u7528\u6a21\u578b\u751f\u6210\u6458\u8981</h3>\n<p>\u6700\u521d\u65b9\u6848\u662f\u4e0a\u4e0b\u6587\u6ee1\u4e86\u4e4b\u540e\u8c03\u672c\u5730\u6a21\u578b\u751f\u6210\u6458\u8981\u2014\u2014\u7ed3\u679c\u5355 slot \u963b\u585e\uff0c\u76f4\u63a5\u8d85\u65f6\u3002</p>\n<p>\u6539\u6210\u7eaf\u7b97\u6cd5\u63d0\u53d6\uff1a\u4fdd\u7559\u5934\u90e8\uff08 system prompt + \u9996\u8f6e\u5bf9\u8bdd\uff09\u548c\u5c3e\u90e8\uff08\u6700\u8fd1 8K tokens \uff09\uff0c\u4e2d\u95f4\u90e8\u5206\u63d0\u53d6\u4ee3\u7801\u8def\u5f84\u3001\u51fd\u6570\u540d\u3001\u6587\u4ef6\u540d\u3001TODO \u7b49\u5173\u952e\u4fe1\u606f\u3002\u538b\u7f29\u7387 73%\uff0c\u8017\u65f6 &lt;1ms \u3002</p>\n<hr/>\n<h2>\u7528\u4e86\u54ea\u4e9b\u6280\u672f\uff0c\u5b9e\u73b0\u4e86\u4ec0\u4e48\u529f\u80fd</h2>\n<h3>llama.cpp \u2014 \u63a8\u7406\u5f15\u64ce\u6838\u5fc3</h3>\n<p>\u76f4\u63a5\u8c03\u7528 llama.cpp \u7684 llama-server \uff0c\u6240\u6709\u53c2\u6570\uff08 ctx \u3001KV cache \u7c7b\u578b\u3001\u7ebf\u7a0b\u6570\u3001ubatch \u3001mlock \u3001tensor split \uff09\u90fd\u901a\u8fc7\u542f\u52a8\u53c2\u6570\u6ce8\u5165\u3002Kaiwu \u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u53c2\u6570\u51b3\u7b56\u5c42\uff0c\u4e0d\u6539\u63a8\u7406\u5f15\u64ce\u672c\u8eab\u3002</p>\n<h3>IsoQuant / TurboQuant \u2014 3-bit KV cache \u538b\u7f29</h3>\n<p>\u96c6\u6210\u4e86 johndpope \u7684 turboquant fork \uff08<code>feature/planarquant-kv-cache</code>\uff09\uff0c\u652f\u6301 <code>-ctk iso3 -ctv iso3</code> \u53c2\u6570\u3002iso3 \u7684\u538b\u7f29\u7cfb\u6570\u5b9e\u6d4b 0.73 \uff0c\u7406\u8bba\u503c 0.75 \uff0c\u5728 VRAM \u7d27\u5f20\u7684\u8bbe\u5907\uff08 8GB \uff09\u4e0a\u53ef\u4ee5\u628a KV cache \u5360\u7528\u538b\u7f29\u5230 q8_0 \u7684\u4e00\u534a\u3002\u4f46\u6709\u7ea6 600MB \u56fa\u5b9a\u89e3\u7801 buffer \u5f00\u9500\uff0cVRAM \u5145\u88d5\u65f6\u53cd\u800c\u6bd4 f16 \u6162 8%\uff0c\u6240\u4ee5\u7b56\u7565\u662f VRAM &gt; 16GB \u624d\u9ed8\u8ba4\u5f00 iso3 \u3002</p>\n<h3>oobabooga \u663e\u5b58\u4f30\u7b97\u516c\u5f0f \u2014 ctx \u4e0a\u9650\u9884\u6d4b\uff08\u5df2\u653e\u5f03\uff09</h3>\n<p>\u793e\u533a\u6d41\u4f20\u7684\u516c\u5f0f\u7528\u6765\u9884\u6d4b\u5269\u4f59\u663e\u5b58\u80fd\u652f\u6301\u591a\u5927 ctx \uff0c\u57fa\u4e8e q8_0/f16 \u62df\u5408\u3002iso3 \u573a\u666f\u4e0b\u9ad8\u4f30\u663e\u5b58\u9700\u6c42\uff0c\u5bfc\u81f4 ctx \u53ea\u7b97\u51fa 4K \u3002\u6700\u7ec8\u6539\u6210\u4e8c\u5206\u63a2\u6d4b\u4ee3\u66ff\u516c\u5f0f\uff0c\u8ba9 llama-server \u81ea\u5df1\u51b3\u5b9a\u80fd\u8dd1\u591a\u5c11\u3002</p>\n<h3>GQA \u67b6\u6784\u8bc6\u522b \u2014 KV cache \u7cbe\u51c6\u4f30\u7b97</h3>\n<p>Qwen3 \u7b49\u65b0\u6a21\u578b\u7528 GQA \uff08 Grouped Query Attention \uff09\uff0ckv_heads \u8fdc\u5c0f\u4e8e attention_heads \u3002KV cache \u5927\u5c0f\u516c\u5f0f\u91cc\u7528\u7684\u662f kv_heads \u800c\u4e0d\u662f heads \uff0c\u4e0d\u8bc6\u522b\u8fd9\u4e00\u70b9\u4f1a\u9ad8\u4f30 3-4 \u500d\u3002\u901a\u8fc7\u8bfb GGUF metadata \u62ff\u5230\u51c6\u786e\u7684 kv_heads \u503c\u518d\u505a\u8ba1\u7b97\u3002</p>\n<h3>MoE tensor \u8bc6\u522b \u2014 \u81ea\u52a8 expert offload</h3>\n<p>\u8bfb\u53d6\u6a21\u578b\u7684 tensor \u540d\u79f0\u5217\u8868\uff0c\u5339\u914d <code>.ffn_.*_exps.</code> \u6a21\u5f0f\u8bc6\u522b\u51fa MoE expert \u5c42\uff0c\u81ea\u52a8\u51b3\u5b9a\u628a\u8fd9\u90e8\u5206\u8def\u7531\u5230 CPU \u3002\u4e0d\u9700\u8981\u7528\u6237\u624b\u52a8\u6307\u5b9a\uff0c\u4e5f\u4e0d\u9700\u8981\u63d0\u524d\u77e5\u9053\u6a21\u578b\u67b6\u6784\u3002</p>\n<h3>Extractive Summary \u2014 \u96f6\u5ef6\u8fdf\u5bf9\u8bdd\u538b\u7f29</h3>\n<p>\u4e0a\u4e0b\u6587\u5230 75% \u65f6\u89e6\u53d1\uff0c\u7eaf\u7b97\u6cd5\u63d0\u53d6\uff1a\u4fdd\u7559 system prompt \u3001\u9996\u8f6e\u5bf9\u8bdd\u3001\u6700\u8fd1 8K tokens \uff0c\u4e2d\u95f4\u90e8\u5206\u6309\u5173\u952e\u8bcd\u6743\u91cd\u4fdd\u7559\uff08\u4ee3\u7801\u8def\u5f84\u3001\u51fd\u6570\u540d\u3001\u6587\u4ef6\u540d\u3001TODO \u3001\u547d\u4ee4\u884c\u7b49\uff09\u3002\u4e0d\u8c03\u7528\u4efb\u4f55\u6a21\u578b\uff0c\u538b\u7f29\u8017\u65f6 &lt;1ms \uff0c73% \u538b\u7f29\u7387\u3002\u6700\u521d\u8bd5\u8fc7\u8c03\u672c\u5730\u6a21\u578b\u751f\u6210\u6458\u8981\uff0c\u5355 slot \u963b\u585e\u76f4\u63a5\u8d85\u65f6\uff0c\u8fd9\u6761\u8def\u8d70\u4e0d\u901a\u3002</p>\n<h3>GitHub Actions CI \u2014 \u8de8\u5e73\u53f0\u81ea\u52a8\u7f16\u8bd1</h3>\n<p>turboquant fork \u9700\u8981\u81ea\u5df1\u7f16\u8bd1\u5e26 iso3 \u652f\u6301\u7684 llama-server \u3002\u7528 GitHub Actions \u540c\u65f6\u7f16\u8bd1 Windows \uff08 MSVC \uff09\u548c Linux \uff08 GCC \uff09\u7248\u672c\uff0cCUDA 12.4 \uff0c\u8986\u76d6 sm_75/80/86/89 \u67b6\u6784\uff0cRTX 50 \u7cfb\u5217\u901a\u8fc7 PTX JIT \u8fd0\u884c\u65f6\u652f\u6301\u3002\u8e29\u4e86\u4e09\u4e2a MSVC \u7f16\u8bd1\u5751\uff08 extern \"C\" \u58f0\u660e\u6539\u5b9a\u4e49\u3001M_PI \u672a\u5b9a\u4e49\u3001\u5168\u5c40\u7b26\u53f7\u7f3a\u5931\uff09\uff0c\u8bb0\u5f55\u5728 <a href=\"http://PROGRESS.md\" rel=\"nofollow\">PROGRESS.md</a> \u91cc\u3002</p>\n<hr/>\n<h2>\u5de5\u5177</h2>\n<p>\u628a\u4e0a\u9762\u8fd9\u4e9b\u903b\u8f91\u90fd\u81ea\u52a8\u5316\u4e86\uff0c\u53eb\u5f00\u7269\uff08 Kaiwu \uff09\u3002\u4e00\u884c\u547d\u4ee4\u542f\u52a8\uff0c\u53c2\u6570\u5168\u90e8\u81ea\u52a8\u627e\uff0c\u7ed3\u679c\u7f13\u5b58\u8d77\u6765\uff0c\u7b2c\u4e8c\u6b21 2 \u79d2\u542f\u52a8\u3002</p>\n<p>GitHub\uff1a <a href=\"https://github.com/val1813/kaiwu\" rel=\"nofollow\">https://github.com/val1813/kaiwu</a></p>\n<p>OpenAI \u517c\u5bb9 API \uff0cContinue / Cursor / Claude Code \u76f4\u63a5\u63a5\u3002</p>\n<hr/>\n<p>\u6709\u9047\u5230\u7c7b\u4f3c\u95ee\u9898\u7684\u6b22\u8fce\u4ea4\u6d41\uff0c\u5c24\u5176\u662f MoE offload \u548c KV cache \u8fd9\u5757\u8e29\u5751\u633a\u6df1\u7684\u3002</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/archxm", 
        "name": "archxm", 
        "avatar": "https://cdn.v2ex.com/avatar/6fb7/882c/72419_large.png?m=1766119273"
      }, 
      "url": "https://www.v2ex.com/t/1208354", 
      "date_modified": "2026-04-25T02:19:04+00:00", 
      "content_html": "<ul>\n<li>\u6bd4\u5982\u6211\u4e0b\u8f7d\u4e86\u4e00\u4e2a\u6a21\u578b\u3002</li>\n<li>\u7136\u540e\u518d\u628a\u6211\u6240\u6709\u6587\u6863\u4ea4\u7ed9\u5b83\uff0c\u4e8c\u6b21\u8bad\u7ec3\u3002</li>\n<li>\u90a3\u4e48\uff0c\u662f\u4e0d\u662f\u5c31\u6ca1\u5fc5\u8981 RAG \u4e86\u3002</li>\n<li>\u901a\u8fc7\u8fd9\u4e2a\u6a21\u578b\uff0c\u6211\u5c31\u80fd\u63d0\u95ee\u4e86\u561b\uff0c\u6bd5\u7adf\uff0c\u6211\u7684\u57fa\u56e0\u5df2\u7ecf\u5d4c\u5165\u8fdb\u53bb\u4e86\u3002</li>\n</ul>\n", 
      "date_published": "2026-04-24T09:58:36+00:00", 
      "title": "\u5927\u4f19\u6709\u60f3\u8fc7\u4e8c\u6b21\u8bad\u7ec3\u5417\uff1f", 
      "id": "https://www.v2ex.com/t/1208354"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/qazwsxkevin", 
        "name": "qazwsxkevin", 
        "avatar": "https://cdn.v2ex.com/gravatar/bfefb99d6203d351791672a1d3fc936a?s=73&d=retro"
      }, 
      "url": "https://www.v2ex.com/t/1207819", 
      "date_modified": "2026-04-22T11:29:00+00:00", 
      "content_html": "<p>\u8fd9\u4e1c\u897f\u6211\u90fd\u6ca1\u89c1\u8fc7\u5b9e\u7269\uff0c\u5728 USA \u7684\u540c\u5b66\u6709\u4e00\u53f0\uff0c\u4f46\u662f\u4ed6\u4e0a\u7ebf\u540e\u5e94\u7528\u7684\u4e8b\u60c5\u90fd\u5feb\u62c9\u7206\u4e86\uff0c\u6682\u65f6\u6ca1\u7a7a\u6d4b\u6211\u7684\u60f3\u6cd5\uff0c\u53ea\u80fd\u60f3\u8c61\u7740\u6765\u95ee\u4e00\u4e0b\u8fd9\u91cc\u5404\u4f4d\u4f6c\u4e86:</p>\n<ul>\n<li>C++,Python \u7684\u4ee3\u7801 review;</li>\n<li>\u6839\u636e\u63d0\u793a\u8bcd,\u5904\u7406 MySQL \u8fd4\u56de\u6765 8k~13k \u6761\u5df2\u6210 JSON \u7684\u6570\u636e\u63d0\u53d6;(\u5b57\u6bb5\u548c\u5185\u5bb9\u4e0d\u591a,\u7206\u4e0d\u4e86\u4e0a\u4e0b\u6587)</li>\n<li>\u7ed9\u51fa\u521d\u9ad8\u4e2d\u7684\u6570\u5b66\u7269\u7406,\u67d0\u9898\u7684\u89e3\u9898\u601d\u8def;<br/>\n\u9700\u6c42\u5c31\u8fd9\u4e09\u7c7b\u4e8b\u60c5\u4e3a\u4e3b\u3002</li>\n</ul>\n<p>\u95ee\u9898:</p>\n<ul>\n<li>DGX Spark 128G \u8dd1\u4e2a\u4ec0\u4e48\u6a21\u578b\u80fd\u5e94\u4ed8\u4ee5\u4e0a\u4e09\u7c7b\u5f3a\u5ea6\u7684\u4e8b\u60c5?</li>\n<li>\u5982\u679c\u6709\u5408\u9002(\u6216\u8005\u5c06\u5c31)\u5e94\u4ed8\u7684\u6a21\u578b\uff0c90%\u989d\u5b9a\u5bb9\u91cf\u7684\u4e0a\u4e0b\u6587\u6253\u8fdb\u53bb,\u8981\u591a\u4e45\u65f6\u95f4\u6709\u53cd\u5e94\u5f00\u59cb\u51fa tokens?</li>\n<li>\u6bcf\u79d2\u80fd\u5410\u591a\u5c11 tokens?</li>\n</ul>\n<p>\u8003\u8651:</p>\n<ul>\n<li>\u573a\u5730\u7a7a\u95f4\u548c\u7269\u7406\u6761\u4ef6\u6240\u9650,\u53ea\u80fd\u627e\u8fd9\u7c7b\u5c0f\u673a.</li>\n<li>\u53ef\u4ee5\u8003\u8651 Mac Studio M3U 256G,\u518d\u65b0\u6b3e\u7684\u52a0\u94b1\u4e5f\u4e0d\u597d\u4e70,\u4e5f\u8d35.</li>\n</ul>\n", 
      "date_published": "2026-04-22T11:28:26+00:00", 
      "title": "\u7528 DGX Spark \u505a\u8fd9\u4e9b\u4e8b\u60c5\uff0c\u662f\u5426\u80fd\u529b\u5408\u9002/\u8db3\u591f\uff0c\u6709\u4f6c\u80fd\u89e3\u7b54\u5417?(\u4f30\u7b97\u4e5f\u884c)", 
      "id": "https://www.v2ex.com/t/1207819"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/mingtdlb", 
        "name": "mingtdlb", 
        "avatar": "https://cdn.v2ex.com/avatar/1067/fd49/525301_large.png?m=1742795473"
      }, 
      "url": "https://www.v2ex.com/t/1207541", 
      "title": "\u591a\u53f0 GPU \u4e4b\u95f4\u600e\u4e48\u7ec4\u7f51\u4e92\u8054\uff1f", 
      "id": "https://www.v2ex.com/t/1207541", 
      "date_published": "2026-04-21T09:52:14+00:00", 
      "content_html": "<p>\u6bd4\u5982\u8981\u90e8\u7f72 deepseek \u6ee1\u8840\u7248\uff0c\u603b\u4e0d\u80fd\u7528\u4e00\u53f0\u8dd1\u5bf9\u5427\uff0c\u90a3\u6bd4\u5982\u6709\u4e09\u53f0 SXM \u7248\u7684 8 \u5361 A100 \u7684 GPU \u670d\u52a1\u5668</p>\n<p>\u597d\u5947\u95ee\u4e00\u4e0b\uff0c\u60f3\u5b66\u4e60\u5b66\u4e60</p>\n"
    }, 
    {
      "author": {
        "url": "https://www.v2ex.com/member/diudiuu", 
        "name": "diudiuu", 
        "avatar": "https://cdn.v2ex.com/avatar/53c0/4118/1055_large.png?m=1780363419"
      }, 
      "url": "https://www.v2ex.com/t/1207254", 
      "date_modified": "2026-04-20T09:43:48+00:00", 
      "content_html": "<p>\u6bd4\u5982\u770b dgx spark \u8fd9\u53f0\u673a\u5b50\uff0c\u90e8\u7f72 31B BF16 gemma</p>\n<p>\u8fd9\u53f0\u673a\u5b50\u7684\u5e26\u5bbd 273 GB/s</p>\n<p><strong>31B \u53c2\u6570 \u00d7 2 bytes (BF16) \u00f7 273 GB/s = \u6bcf\u4e2a token 227 ms = \u7406\u8bba\u6700\u5927 4.4 token/s</strong></p>\n<p>\u5b9e\u9645\u80fd\u5230 3token/s \u5df2\u7ecf\u662f\u725b\u903c plus \uff0c\u9876\u591a 2.5token/s</p>\n<p>\u6240\u4ee5\u6709\u4e2a\u5173\u7cfb\uff0c\u4e0d\u8981\u95ee\u80fd\u4e0d\u80fd\u8fd0\u884c\u548b\u7684\uff0c\u81ea\u5df1\u5927\u6982\u7b97\u4e0b\u57fa\u672c\u5c31\u77e5\u9053\u80fd\u4e0d\u80fd\u7528</p>\n<p>\u7b80\u5355\u5f97\u63a8\u7406\u6211\u89c9\u5f97\u81f3\u5c11\u8981\u5230<strong>25token/s</strong>\uff0c\u770b\u8d77\u6765\u624d\u6b63\u5e38</p>\n<p><strong>1. \u6a21\u578b\u5fc5\u987b\u80fd\u52a0\u8f7d\u5b8c\uff0c\u663e\u5b58\u53ea\u662f\u57fa\u672c\u6761\u4ef6</strong></p>\n<p><strong>2. \u5fc5\u987b\u8981\u770b\u5185\u5b58\u5e26\u5bbd\uff08 Memory Bandwidth \uff09\uff0c\u8fd9\u4e2a\u592a\u4f4e\u5f97\u8bdd\u4f30\u8ba1\u5c31\u662f\u4e2a\u8ddb\u5b50\uff0c\u6211\u770b\u51e0\u4e4e\u5f88\u5c11\u6709\u4eba\u90e8\u7f72\u6a21\u578b\u65f6\u6ce8\u610f\u8fd9\u4e2a\u914d\u7f6e\uff0c\u8fd9\u4e2a\u4e5f\u662f\u975e\u5e38\u91cd\u8981\u5f97\u53c2\u6570</strong></p>\n<p><strong>3. \u4e0a\u9762\u5f97\u57fa\u672c\u662f\u6309\u7167\u82f1\u4f1f\u8fbe\u673a\u5b50\u7b97\u51fa\u6765\u5f97\uff0cmac \u673a\u5b50\u6bd4\u8f83\u7279\u6b8a\uff0c\u57fa\u672c\u53ea\u8981\u80fd\u52a0\u8f7d\u5230 gpu \u91cc\u9762\uff0c\u5269\u4f59\u4e00\u70b9\u5185\u5b58\uff0c\u5c31\u80fd\u7528\u901f\u5ea6\u4e0d\u4f1a\u5f88\u6162\uff08 20token/s \u5c06\u5c31\u80fd\u7528\uff09\uff0c\u51b7\u542f\u52a8\u7a0d\u5fae\u6162\u70b9</strong></p>\n<p>\u8fd8\u6709\u4e2a\u672c\u5730\u6a21\u578b\u90e8\u7f72\uff0c<strong>\u9664\u4e86\u82b1\u5927\u94b1</strong>\uff0c\u672c\u5730\u90e8\u7f72\u5c31\u662f\u73a9\u73a9\u53ef\u4ee5\uff0c\u8d77\u7801\u73b0\u5728\u4e0d\u8981\u5984\u60f3\u8d85\u8fc7\u7ebf\u4e0a\u5f97\u6a21\u578b\uff0c\u5c24\u5176\u5199\u4ee3\u7801\u65b9\u9762</p>\n<p>\u6211\u4e2a\u4eba\u8ba4\u4e3a\u73b0\u5728\u672c\u5730\u6a21\u578b\u80fd\u505a\u5f97\u4e8b</p>\n<ol>\n<li>ocr</li>\n<li>\u603b\u7ed3\u505a\u77e5\u8bc6\u5e93</li>\n<li>openclaw \u8fd8\u6709\u4ec0\u4e48\u7231\u9a6c\u4ed5\u8fd9\u4e2a\u63a8\u7406\u4e5f\u53ef\u4ee5\u505a\uff0c\u9700\u8981\u63d0\u524d\u7528\u7ebf\u4e0a\u6a21\u578b\u5b8c\u6210\u590d\u6742\u5f97\u4ee3\u7801\uff0c\u672c\u5730\u6267\u884c\u63a8\u7406\u4e00\u5b9a\u8981\u8bb0\u5f97\u505a\u597d\u673a\u5b50\u6563\u70ed\uff0c\u4e00\u5b9a\uff01\uff01\u4e00\u5b9a\u4e00\u5b9a\uff01\uff01\uff01</li>\n</ol>\n<p>\u5e0c\u671b\u5927\u5bb6\u6765\u4ea4\u6d41\u81ea\u5df1\u5f97\u5fc3\u5f97\uff0c\u5927\u5bb6\u5171\u540c\u5b66\u4e60\u8fdb\u6b65</p>\n", 
      "date_published": "2026-04-20T09:40:14+00:00", 
      "title": "\u90e8\u7f72\u672c\u5730\u6a21\u578b token \u8f93\u51fa\u4e07\u80fd\u516c\u5f0f", 
      "id": "https://www.v2ex.com/t/1207254"
    }
  ]
}