make local-llama service

local-llama.service

@@ -0,0 +1,25 @@
+[Unit]
+Description=llama-server (local-llama)
+After=network.target
+
+[Service]
+Type=simple
+User=grayhook
+Group=grayhook
+WorkingDirectory=/home/grayhook/Documents/llama-space
+EnvironmentFile=/home/grayhook/Documents/llama-space/.env
+ExecStart=/home/grayhook/Documents/llama-space/local-llama.sh
+Restart=on-failure
+RestartSec=5
+
+# CPU pinning: taskset -c 0-11
+CPUAffinity=0-11
+
+# ulimit -l unlimited → lock memory (mlockall / MADV_POPULATE_READ + mlock)
+LimitMEMLOCK=infinity
+
+# Не убивать OOM killer — модель критична
+OOMScoreAdjust=-900
+
+[Install]
+WantedBy=multi-user.target

local-llama.sh

@@ -1,12 +1,8 @@
 #!/bin/bash
 
-set -a            # Automatically export all variables defined next
+set -euo pipefail
-source .env       # Load the file
-set +a            # Disable automatic export
 
-sudo ulimit -l unlimited
+/home/bazzite/Документы/llama-cpp-turboquant/build/bin/llama-server \
-
-taskset -c 0-11 /home/bazzite/Документы/llama-cpp-turboquant/build/bin/llama-server \
   -m "models/${MODEL_FILE}" \
   --host 0.0.0.0 \
   --port 8080 \