{"id":686,"date":"2026-10-26T21:14:52","date_gmt":"2026-10-26T13:14:52","guid":{"rendered":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/2026\/10\/26\/wsl2-llamacpp-qwen38\/"},"modified":"2026-10-26T21:59:36","modified_gmt":"2026-10-26T13:59:36","slug":"wsl2-llamacpp-qwen38","status":"publish","type":"post","link":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/2026\/10\/26\/wsl2-llamacpp-qwen38\/","title":{"rendered":"3080+WSL2 + llama.cpp \u90e8\u7f72 Qwen3.8-27B-Coder"},"content":{"rendered":"

27B \u7684 Coder \u6a21\u578b\u585e\u8fdb 20GB \u663e\u5b58\u662f\u4ec0\u4e48\u4f53\u9a8c\uff1fQ3 \u91cf\u5316 + KV cache 8bit + Flash Attention \u4e09\u4ef6\u5957\u4e0b\u6765\uff0c64K \u4e0a\u4e0b\u6587\u53ea\u5403 15.7GB\uff0cnetsh \u7aef\u53e3\u8f6c\u53d1\u4e00\u914d\uff0c\u5c40\u57df\u7f51\u91cc\u53e6\u4e00\u53f0\u4e3b\u673a\u7684 WorkBuddy \u76f4\u63a5\u628a\u5b83\u5f53 OpenAI API \u7528\u3002<\/p>\n

WSL2 + llama.cpp \u90e8\u7f72 Qwen3.8-27B-Coder\uff1aRTX 3080 20GB \u5b9e\u6218 OpenAI \u517c\u5bb9 API<\/h1>\n

\u73af\u5883\uff1aWindows + WSL2 (Ubuntu 26.04) \u00b7 NVIDIA RTX 3080 20GB \u00b7 llama.cpp\uff08CUDA \u6784\u5efa\uff09\u00b7 Qwen3.8-27B-Coder390 Q3LynnStyle16G\uff0813.28GB GGUF\uff09
\u751f\u6210\u65e5\u671f\uff1a2026-10-26<\/p><\/blockquote>\n

\n

\u4e00\u3001\u73af\u5883\u6982\u89c8\u4e0e\u5173\u952e\u8def\u5f84<\/h2>\n

\u6574\u5957\u65b9\u6848\u7684\u76ee\u6807\u5f88\u660e\u786e\uff1aWSL2 \u91cc\u8dd1\u63a8\u7406\uff0c\u5bf9\u5916\u63d0\u4f9b OpenAI \u517c\u5bb9 API\uff0c\u7ed9\u5c40\u57df\u7f51\u5185\u53e6\u4e00\u53f0\u4e3b\u673a\u4e0a\u7684 WorkBuddy \u5f53\u6a21\u578b\u540e\u7aef\u3002\u5148\u770b\u603b\u7684\u73af\u5883\uff1a<\/p>\n

\"\u73af\u5883\u6982\u89c8\"<\/figure>\n

\u6d89\u53ca\u7684\u51e0\u4e2a\u5173\u952e\u8def\u5f84\uff0c\u540e\u9762\u6240\u6709\u547d\u4ee4\u90fd\u56f4\u7ed5\u5b83\u4eec\u5c55\u5f00\uff1a<\/p>\n

\"\u5173\u952e\u8def\u5f84\"<\/figure>\n
\n

\u4e8c\u3001\u6a21\u578b\u4e0b\u8f7d\uff1a\u7ed5\u8fc7 XetHub \u8ba4\u8bc1\u5751<\/h2>\n

\u6a21\u578b\u4ece HuggingFace \u955c\u50cf\u7ad9\uff08hf-mirror.com\uff09\u76f4\u63a5\u4e0b\u8f7d\u2014\u2014\u6ce8\u610f\u8fd9\u91cc\u7684\u5751\uff1ahuggingface_hub \u9ed8\u8ba4\u8d70\u7684 XetHub CAS \u540e\u7aef\uff0c\u955c\u50cf\u7ad9\u4e0d\u652f\u6301 Xet \u8ba4\u8bc1\uff0c\u6240\u4ee5\u76f4\u63a5\u7528 wget<\/code> \u6253\u955c\u50cf\u7684 resolve \u94fe\u63a5\u6700\u7a33\u3002\u4e0b\u8f7d\u811a\u672c\u5982\u4e0b\uff1a<\/p>\n

#!\/usr\/bin\/env bash\n# \u4ece hf-mirror.com \u76f4\u63a5\u4e0b\u8f7d Qwen3.8-27B-Coder390 \u7684 Q3LynnStyle GGUF \u6587\u4ef6\n# \u7ed5\u8fc7 huggingface_hub \u7684 XetHub CAS \u540e\u7aef\uff08\u955c\u50cf\u4e0d\u652f\u6301 Xet \u8ba4\u8bc1\uff09\nset -uo pipefail\n\nREPO=\"nerkyor\/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2\"\nBASE=\"https:\/\/hf-mirror.com\/${REPO}\/resolve\/main\/GGUF\"\nDIR=\/home\/antares\/models\/qwen38-27b\/GGUF\nmkdir -p \"$DIR\"\n\nfiles=(\n  \"Qwen3.8-27B-Coder390-EfficientThink-Q3LynnStyle-MTP.gguf\"\n  \"Qwen3.8-27B-Coder390-EfficientThink-Q3LynnStyle-Q8MTP.gguf\"\n  \"Qwen3.8-27B-Coder390-Q3LynnStyle16G-MTP.gguf\"\n)\n\nfor f in \"${files[@]}\"; do\n    echo \"[$(date +%H:%M:%S)] === \u4e0b\u8f7d $f ===\"\n    if wget -c --progress=dot:giga -O \"$DIR\/$f\" \"$BASE\/$f\"; then\n        echo \"[$(date +%H:%M:%S)] OK  $f\"\n    else\n        echo \"[$(date +%H:%M:%S)] FAIL $f (exit $?)\"\n    fi\ndone\necho \"[$(date +%H:%M:%S)] \u5168\u90e8\u4efb\u52a1\u7ed3\u675f\uff0c\u7ed3\u679c\uff1a\"<\/code><\/pre>\n
\u6267\u884c\u5c31\u662f\u4e00\u884c\uff1a<\/p>\n
bash ~\/models\/downloadQwen38Coder.sh<\/code><\/pre>\n
\u6ce8\uff1a\u811a\u672c\u91cc\u7528\u7684\u662f wget -c<\/code>\uff0c\u65ad\u70b9\u7eed\u4f20\u81ea\u52a8\u751f\u6548\u2014\u201413GB \u7ea7\u522b\u7684\u6587\u4ef6\u4e0b\u8f7d\u4e2d\u65ad\u4e86\u76f4\u63a5\u91cd\u8dd1\u5373\u53ef\uff0c\u5df2\u4e0b\u8f7d\u7684\u90e8\u5206\u4e0d\u4f1a\u6d6a\u8d39\u3002<\/p>\n
\n
\u4e09\u3001\u542f\u52a8\u670d\u52a1<\/h2>\n
3.1 \u4e00\u952e\u542f\u52a8\u547d\u4ee4<\/h3>\n
nohup ~\/llama.cpp\/build\/bin\/llama-server \\\n  -m ~\/models\/qwen38-27b\/GGUF\/Qwen3.8-27B-Coder390-Q3LynnStyle16G-MTP.gguf \\\n  -ngl 99 \\\n  -c 65536 \\\n  --parallel 1 \\\n  --flash-attn on \\\n  --cache-type-k q8_0 \\\n  --cache-type-v q8_0 \\\n  --host 0.0.0.0 \\\n  --port 8000 \\\n  --alias qwen3.8-27b-coder \\\n  > ~\/llamaServer.log 2>&1 &<\/code><\/pre>\n
3.2 \u53c2\u6570\u9010\u6761\u89e3\u8bfb<\/h3>\n
\u8fd9\u5957\u53c2\u6570\u7ec4\u5408\u662f 20GB \u663e\u5b58\u8dd1 27B Q3 \u7684\u6838\u5fc3\uff0c\u9010\u6761\u8bf4\u6e05\u695a\uff1a<\/p>\n
\"\u542f\u52a8\u53c2\u6570\u8bf4\u660e\"<\/figure>\n
\u5176\u4e2d --cache-type-k\/v q8_0<\/code> \u548c --flash-attn on<\/code> \u662f\u663e\u5b58\u4f18\u5316\u7684\u5173\u952e\u4e8c\u4ef6\u5957\uff1aKV cache 8bit \u91cf\u5316\u76f4\u63a5\u628a KV \u663e\u5b58\u780d\u534a\uff0cFlash Attention \u518d\u7701\u4e00\u5c42\u2014\u2014\u4e24\u8005\u53e0\u52a0\u540e 64K \u4e0a\u4e0b\u6587\u603b\u5360\u7528\u63a7\u5236\u5728 15.7GB\u3002\u800c --parallel 1<\/code> \u662f\u4e2a\u5bb9\u6613\u5ffd\u7565\u7684\u70b9\uff1a\u591a slot \u4f1a\u5747\u5206\u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u5355\u8bf7\u6c42\u53ef\u7528 token \u76f4\u63a5\u8170\u65a9\uff0c\u72ec\u5360\u6a21\u5f0f\u624d\u80fd\u5403\u6ee1 64K\u3002<\/p>\n
3.3 \u542f\u52a8\u540e\u9a8c\u8bc1\u56db\u8fde<\/h3>\n
# \u68c0\u67e5\u8fdb\u7a0b\npgrep -af llama-server\n\n# \u68c0\u67e5 API\ncurl -s http:\/\/localhost:8000\/v1\/models | python3 -m json.tool\n\n# \u63a8\u7406\u6d4b\u8bd5\ncurl -s http:\/\/localhost:8000\/v1\/chat\/completions \\\n  -H \"Content-Type: application\/json\" \\\n  -d '{\"model\":\"qwen3.8-27b-coder\",\"messages\":[{\"role\":\"user\",\"content\":\"\u4f60\u597d\"}],\"max_tokens\":50}'\n\n# \u68c0\u67e5\u663e\u5b58\nnvidia-smi --query-gpu=memory.used,memory.free --format=csv,noheader<\/code><\/pre>\n
\u6b63\u5e38\u8fd0\u884c\u65f6\u663e\u5b58\u5360\u7528\u7ea6 15.7GB \/ 20GB<\/strong>\uff0c\u7559\u6709\u4f59\u91cf\u3002<\/p>\n
\n
\u56db\u3001\u7f51\u7edc\u914d\u7f6e\uff1aWSL2 \u7aef\u53e3\u8f6c\u53d1<\/h2>\n
WSL2 \u9ed8\u8ba4\u4e0d\u5bf9\u5916\u66b4\u9732\u7aef\u53e3\uff0cLAN \u91cc\u5176\u4ed6\u673a\u5668\u662f\u8fde\u4e0d\u8fdb\u6765\u7684\uff0c\u9700\u8981\u5728 Windows \u5bbf\u4e3b\u673a<\/strong>\u4e0a\u505a\u4e00\u5c42\u7aef\u53e3\u8f6c\u53d1\u3002<\/p>\n
4.1 \u67e5\u8be2 WSL2 IP<\/h3>\n
# \u5728 WSL2 \u5185\u6267\u884c\nhostname -I | awk '{print $1}'\n# \u793a\u4f8b\u8f93\u51fa\uff1a172.27.7.50<\/code><\/pre>\n
4.2 \u914d\u7f6e\u7aef\u53e3\u8f6c\u53d1\u4e0e\u9632\u706b\u5899<\/h3>\n
# PowerShell \u7ba1\u7406\u5458\u8fd0\u884c\nnetsh interface portproxy add v4tov4 listenport=8000 listenaddress=0.0.0.0 connectport=8000 connectaddress=172.27.7.50\nnetsh advfirewall firewall add rule name=\"llama-server\" dir=in action=allow protocol=TCP localport=8000<\/code><\/pre>\n
4.3 \u67e5\u8be2 Windows \u5bbf\u4e3b\u673a\u5c40\u57df\u7f51 IP<\/h3>\n
# PowerShell \u8fd0\u884c\nGet-NetIPAddress -AddressFamily IPv4 |\n  Where-Object {$_.InterfaceAlias -notlike \"*WSL*\" -and $_.IPAddress -notlike \"127.*\" -and $_.IPAddress -notlike \"169.254.*\" -and $_.IPAddress -notlike \"172.*\"} |\n  Select-Object IPAddress, InterfaceAlias\n# \u793a\u4f8b\u8f93\u51fa\uff1a10.10.10.216<\/code><\/pre>\n
\u6ce8\uff1aWSL2 \u6bcf\u6b21\u91cd\u542f\u540e IP \u53ef\u80fd\u53d8\u5316\uff0cconnectaddress<\/code> \u5c31\u4f1a\u5931\u6548\u3002\u56fe\u7701\u4e8b\u53ef\u4ee5\u5199\u4e2a\u5f00\u673a\u811a\u672c\u81ea\u52a8\u67e5\u8be2\u5e76\u66f4\u65b0\u8f6c\u53d1\u89c4\u5219\uff0c\u4e00\u52b3\u6c38\u9038\u3002<\/p>\n
4.4 \u66f4\u65b0\u7aef\u53e3\u8f6c\u53d1\uff08WSL2 IP \u53d8\u5316\u65f6\uff09<\/h3>\n
# \u5148\u5220\u9664\u65e7\u89c4\u5219\u518d\u6dfb\u52a0\nnetsh interface portproxy delete v4tov4 listenport=8000 listenaddress=0.0.0.0\nnetsh interface portproxy add v4tov4 listenport=8000 listenaddress=0.0.0.0 connectport=8000 connectaddress=<\u65b0\u7684WSL2_IP><\/code><\/pre>\n
\n
\u4e94\u3001WorkBuddy \u5ba2\u6237\u7aef\u63a5\u5165<\/h2>\n
\u5728\u5c40\u57df\u7f51\u53e6\u4e00\u53f0\u4e3b\u673a\u7684 WorkBuddy \u91cc\u6309\u4e0b\u9762\u7684\u53c2\u6570\u914d\u7f6e OpenAI API \u8fde\u63a5\uff1a<\/p>\n
\"WorkBuddy<\/figure>\n
\u6ce8\uff1a\u628a 10.10.10.216<\/code> \u6362\u6210\u4f60\u81ea\u5df1\u7684 Windows \u5bbf\u4e3b\u673a\u5b9e\u9645\u5c40\u57df\u7f51 IP\uff1bAPI Key \u968f\u4fbf\u586b\uff0cllama-server \u9ed8\u8ba4\u4e0d\u6821\u9a8c\u3002<\/p>\n
\n
\u516d\u3001\u65e5\u5e38\u8fd0\u7ef4<\/h2>\n
6.1 \u505c\u6b62 \/ \u91cd\u542f\u670d\u52a1<\/h3>\n
# \u505c\u6b62\npkill -f \"llama-server\"\n\n# \u91cd\u542f\uff1a\u6740\u6389\u540e sleep 3 \u518d\u62c9\u8d77\npkill -f \"llama-server\"; sleep 3\nnohup ~\/llama.cpp\/build\/bin\/llama-server \\\n  -m ~\/models\/qwen38-27b\/GGUF\/Qwen3.8-27B-Coder390-Q3LynnStyle16G-MTP.gguf \\\n  -ngl 99 -c 65536 --parallel 1 --flash-attn on \\\n  --cache-type-k q8_0 --cache-type-v q8_0 \\\n  --host 0.0.0.0 --port 8000 --alias qwen3.8-27b-coder \\\n  > ~\/llamaServer.log 2>&1 &<\/code><\/pre>\n
\u5b9e\u9645\u64cd\u4f5c\u4e2d\u628a\u91cd\u542f\u547d\u4ee4\u5b58\u6210\u4e00\u952e\u811a\u672c\u6700\u7701\u5fc3\u3002<\/p>\n
6.2 \u65e5\u5fd7\u4e0e\u72b6\u6001<\/h3>\n
# \u5b9e\u65f6\u65e5\u5fd7\ntail -f ~\/llamaServer.log\n\n# \u8fdb\u7a0b\npgrep -af llama-server\n\n# \u663e\u5b58\nnvidia-smi --query-gpu=memory.used,memory.free --format=csv,noheader\n\n# API \u53ef\u7528\u6027\ncurl -s http:\/\/localhost:8000\/v1\/models | head -5<\/code><\/pre>\n
\n
\u4e03\u3001\u6545\u969c\u6392\u67e5\u624b\u518c<\/h2>\n
7.1 request exceeds the available context size<\/h3>\n
\u539f\u56e0\uff1a<\/strong>\u8bf7\u6c42 token \u6570\u8d85\u8fc7\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002\u89e3\u51b3\uff1a<\/strong>\u589e\u5927 -c<\/code> \u5e76\u786e\u4fdd --parallel 1<\/code>\u3002\u5f53\u524d 64K \u5df2\u591f\u7528\uff0c\u518d\u5f80\u4e0a\u52a0\u8981\u76ef\u7740\u663e\u5b58\uff1a<\/p>\n
\"\u4e0a\u4e0b\u6587\u4e0e\u663e\u5b58\u5bf9\u7167\"<\/figure>\n
7.2 \u7aef\u53e3\u88ab\u5360\u7528 \/ \u91cd\u542f\u5931\u8d25<\/h3>\n
\u539f\u56e0\uff1a<\/strong>\u65e7\u8fdb\u7a0b\u672a\u5b8c\u5168\u91ca\u653e\u7aef\u53e3\u3002\u89e3\u51b3\uff1a<\/strong>\u5f3a\u6740\u540e\u786e\u8ba4\u7aef\u53e3\u7a7a\u95f2\u518d\u62c9\u8d77\uff1a<\/p>\n
pkill -9 -f \"llama-server\"\nsleep 5\n# \u786e\u8ba4\u7aef\u53e3\u7a7a\u95f2\u540e\u518d\u542f\u52a8\nss -tlnp | grep :8000<\/code><\/pre>\n
7.3 \u53e6\u4e00\u53f0\u4e3b\u673a\u65e0\u6cd5\u8fde\u63a5<\/h3>\n
\u6309\u4e0b\u9762\u4e94\u6b65\u9010\u5c42\u6392\u67e5\uff0c\u4ece\u5185\u5230\u5916\uff1a<\/p>\n
    \n
  • \u2713<\/span> WSL2 \u5185\u670d\u52a1\u6b63\u5e38\uff1acurl http:\/\/localhost:8000\/v1\/models<\/code><\/li>\n
  • \u2713<\/span> Windows \u7aef\u53e3\u8f6c\u53d1\u5b58\u5728\uff1anetsh interface portproxy show v4tov4<\/code><\/li>\n
  • \u2713<\/span> WSL2 IP \u672a\u53d8\u5316\uff1ahostname -I<\/code> \u4e0e connectaddress<\/code> \u6bd4\u5bf9<\/li>\n
  • \u2713<\/span> Windows \u9632\u706b\u5899\u89c4\u5219\u5b58\u5728\uff1anetsh advfirewall firewall show rule name=\"llama-server\"<\/code><\/li>\n
  • \u2713<\/span> Windows \u5c40\u57df\u7f51 IP \u6b63\u786e\uff1a\u4ece\u53e6\u4e00\u53f0\u4e3b\u673a ping 10.10.10.216<\/code><\/li>\n<\/ul>\n
    7.4 \u663e\u5b58 OOM<\/h3>\n
    \u89e3\u51b3\uff1a<\/strong>\u964d\u4f4e\u4e0a\u4e0b\u6587 -c<\/code>\uff0c\u6216\u6362\u66f4\u5c0f\u7684\u91cf\u5316\u3002~\/models\/qwen38-27b\/GGUF\/<\/code> \u4e0b\u7684\u4e09\u4e2a\u91cf\u5316\u6587\u4ef6\u6309\u663e\u5b58\u91cf\u529b\u800c\u884c\uff1a<\/p>\n
    \"\u53ef\u9009\u6a21\u578b\u6587\u4ef6\u5bf9\u6bd4\"<\/figure>\n
    \u6ce8\uff1a20GB \u663e\u5b58\u65e0\u8111\u9009 Q3LynnStyle16G\uff0813.28GB\uff09\uff0c\u662f\u5b9e\u6d4b\u6700\u7a33\u7684\u4e00\u6863\uff1bQ8 \u7cbe\u5ea6\u66f4\u9ad8\u4f46 17.52GB \u88c5\u4e0d\u4e0b 64K \u4e0a\u4e0b\u6587\uff0c\u9664\u975e\u4f60\u663e\u5b58 24GB \u5f80\u4e0a\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
    WSL2 + llama.cpp \u5b9e\u6218\uff1aRTX 3080 20GB \u8dd1 Qwen3.8-27B-Coder Q3 \u91cf\u5316\u7248\uff0c64K \u4e0a\u4e0b\u6587\u5360 15.7GB\uff0cnetsh \u7aef\u53e3\u8f6c\u53d1\u540e\u5c40\u57df\u7f51\u76f4\u8fde OpenAI \u517c\u5bb9 API\u3002<\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[44,26,45],"tags":[],"class_list":["post-686","post","type-post","status-publish","format-standard","hentry","category-ai","category-tech-share","category-small-model-app"],"_links":{"self":[{"href":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/wp-json\/wp\/v2\/posts\/686","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/wp-json\/wp\/v2\/comments?post=686"}],"version-history":[{"count":2,"href":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/wp-json\/wp\/v2\/posts\/686\/revisions"}],"predecessor-version":[{"id":688,"href":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/wp-json\/wp\/v2\/posts\/686\/revisions\/688"}],"wp:attachment":[{"href":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/wp-json\/wp\/v2\/media?parent=686"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/wp-json\/wp\/v2\/categories?post=686"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nelliemylove.cn\/nelliemylove\/index.php\/wp-json\/wp\/v2\/tags?post=686"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}