39 lines
1.8 KiB
Bash
Executable File
39 lines
1.8 KiB
Bash
Executable File
#!/bin/bash
|
||
# llama-server 模型切换开关：minicpm ↔ off（2026-09-08：主力已由 Qwen3.5-4B 切 MiniCPM5-2B）
|
||
# 用法: llama-switch.sh minicpm|4b|off|status
|
||
# 服务占用 :8080 端口
|
||
# 铁律来源: 3B 曾以 ctx 16384 CPU 推理致 RSS 7.6GB 内存黑洞; 7B/3B 均不再部署。4B 已由 MiniCPM5-2B 替换(省显存1.27G+快一倍)。
|
||
|
||
case "$1" in
|
||
3b|7b)
|
||
echo "🛑 3B/7B 已禁用（牧尘 2026-09-07 铁律）——本机只允许 MiniCPM5-2B (llama-server-4b unit)。"
|
||
echo " 原因: 3B CPU 推理 RSS 7.6GB 内存黑洞; MiniCPM5-2B GPU 部署仅需 VRAM 1.9G + host ~500MB。"
|
||
exit 1
|
||
;;
|
||
4b)
|
||
echo "ℹ️ 4B 已由 MiniCPM5-2B 替换（2026-09-08）——启动 MiniCPM（同 unit llama-server-4b 跑 MiniCPM 权重）。"
|
||
systemctl --user start llama-server-4b
|
||
echo "✅ 已启动 MiniCPM5-2B (Q4_K_M, ctx 90112/88K, KV q8_0, reasoning off) @ :8080"
|
||
;;
|
||
minicpm|start)
|
||
systemctl --user start llama-server-4b
|
||
echo "✅ 已启动 MiniCPM5-2B (Q4_K_M, ctx 90112/88K, KV q8_0, reasoning off) @ :8080"
|
||
;;
|
||
off)
|
||
systemctl --user stop llama-server-4b 2>/dev/null
|
||
echo "🛑 本地 LLM 服务已停止"
|
||
;;
|
||
status)
|
||
_s4b=$(systemctl --user is-active llama-server-4b 2>/dev/null)
|
||
if [ "$_s4b" = "active" ]; then
|
||
echo "🟢 当前: MiniCPM5-2B (Q4_K_M, ctx 90112/88K, KV q8_0, reasoning off) @ :8080"
|
||
echo " 显存: $(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits) MiB / 4096 MiB"
|
||
else
|
||
echo "⚪ 本地 LLM 服务未运行（llama-server-4b inactive）"
|
||
fi
|
||
;;
|
||
*)
|
||
echo "用法: llama-switch.sh minicpm|off|status （3B/7B 已禁用）"
|
||
;;
|
||
esac
|