单元文件（/etc/systemd/system/llama-server.service）

[Unit]
Description=llama.cpp router server
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
WorkingDirectory=/service/ai/llama.cpp
ExecStart=/service/ai/llama.cpp/build/bin/llama-server \
  --models-preset /service/ai/model/models.ini \
  --models-max 1 \
  --host 127.0.0.91 \
  --port 9192 \
  --api-key-file /service/ai/api.key

Restart=always
RestartSec=5
LimitMEMLOCK=infinity
LimitNOFILE=65535

[Install]
WantedBy=multi-user.target
root@serNCnJRbYWqCtG:~# cat /service/ai/model/models.ini
[*]
threads = 16
threads-batch = 16
batch-size = 512
ubatch-size = 128
parallel = 1
jinja = 1

temp = 0.6
top-k = 40
top-p = 0.85
min-p = 0.05

repeat-penalty = 1.2
repeat-last-n = 256
presence-penalty = 0.3
frequency-penalty = 0.3

n-predict = 4096
no-warmup = 1
mmap = 1

[✏️Coder-Qwen3-1.7B-Q5]
model = /service/ai/model/Coder/Qwen3-1.7b-coder-distilled-sft-Q5_K_M.gguf
ctx-size = 32768

[✏️Coder-Qwen3-30B-A3B-Q4]
model = /service/ai/model/Coder/Qwen3-Coder-30B-A3B-Instruct-Pruned-Q4_K_M.gguf
ctx-size = 16384

[🗣️Talk-Qwen3-1.7B-Q4]
model = /service/ai/model/Qwen/Qwen3-1.7B-Q4_K_M.gguf
ctx-size = 32768

[🗣️Talk-GLM4.7-23B-A3B-Q4]
model = /service/ai/model/GLM/GLM-4.7-Flash-REAP-23B-A3B-Q4_K_M.gguf
ctx-size = 16384

api.key文件生成

mkdir -p /service/ai
openssl rand -hex 32 > /service/ai/api.key
chmod 600 /service/ai/api.key
cat /service/ai/api.key