90 lines
3.3 KiB
Bash
90 lines
3.3 KiB
Bash
|
#!/usr/bin/env bash
|
||
|
|
set -euo pipefail
|
||
|
|
|
||
|
|
# Startet das zentrale, lokale LLM des Voice-Assistants als llama.cpp-Server
|
||
|
|
# (OpenAI-kompatibel). Dieses grosse, unzensierte Modell ist die Haupt-KI fuer
|
||
|
|
# den Provider "local-openai-compatible".
|
||
|
|
#
|
||
|
|
# Alles ueber ENV ueberschreibbar, z. B.:
|
||
|
|
# HOST_PORT=8101 GPU_DEVICE=2 ./start-llm-server.sh
|
||
|
|
# MODEL_REL_PATH="models/qwen3/anderes-modell.gguf" ./start-llm-server.sh
|
||
|
|
|
||
|
|
HF_HOME="${HF_HOME:-/home/dschlueter/nvme2n1p7_home/huggingface}"
|
||
|
|
MODEL_REL_PATH="${MODEL_REL_PATH:-models/qwen3/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf}"
|
||
|
|
IMAGE="${IMAGE:-ghcr.io/ggml-org/llama.cpp:server-cuda}"
|
||
|
|
CONTAINER_NAME="${CONTAINER_NAME:-va_llm}"
|
||
|
|
HOST_PORT="${HOST_PORT:-8001}"
|
||
|
|
CONTAINER_PORT="${CONTAINER_PORT:-8000}"
|
||
|
|
MODEL_ALIAS="${MODEL_ALIAS:-va_llm}"
|
||
|
|
GPU_DEVICE="${GPU_DEVICE:-1}"
|
||
|
|
|
||
|
|
echo "[*] Verwende HF_HOME = $HF_HOME"
|
||
|
|
echo "[*] Modell = $MODEL_REL_PATH"
|
||
|
|
echo "[*] GPU device = $GPU_DEVICE (ueberschreibbar: GPU_DEVICE=2 ./start-llm-server.sh)"
|
||
|
|
echo "[*] Port = $HOST_PORT | Alias = $MODEL_ALIAS | Container = $CONTAINER_NAME"
|
||
|
|
if [ ! -f "$HF_HOME/$MODEL_REL_PATH" ]; then
|
||
|
|
echo "[!] Modell-Datei nicht gefunden: $HF_HOME/$MODEL_REL_PATH" >&2
|
||
|
|
exit 1
|
||
|
|
fi
|
||
|
|
|
||
|
|
if docker ps -a --format '{{.Names}}' | grep -q "^${CONTAINER_NAME}\$"; then
|
||
|
|
echo "[*] Stoppe existierenden Container $CONTAINER_NAME ..."
|
||
|
|
docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
|
||
|
|
fi
|
||
|
|
|
||
|
|
echo "[*] Starte llama.cpp-Server (zentrale Voice-Assistant-KI) ..."
|
||
|
|
docker run -d \
|
||
|
|
--gpus "\"device=${GPU_DEVICE}\"" \
|
||
|
|
--name "$CONTAINER_NAME" \
|
||
|
|
--restart unless-stopped \
|
||
|
|
-e HF_HOME="/hf_home" \
|
||
|
|
-v "$HF_HOME:/hf_home:ro" \
|
||
|
|
-p "${HOST_PORT}:${CONTAINER_PORT}" \
|
||
|
|
"$IMAGE" \
|
||
|
|
-m "/hf_home/${MODEL_REL_PATH}" \
|
||
|
|
--alias "${MODEL_ALIAS}" \
|
||
|
|
-c 262144 \
|
||
|
|
-n 16384 \
|
||
|
|
--jinja \
|
||
|
|
--reasoning on \
|
||
|
|
--no-context-shift \
|
||
|
|
--temp 0.65 \
|
||
|
|
--top-p 0.80 \
|
||
|
|
--top-k 20 \
|
||
|
|
--min-p 0.01 \
|
||
|
|
--repeat-penalty 1.05 \
|
||
|
|
--main-gpu 0 \
|
||
|
|
-ngl 999 \
|
||
|
|
-fa on \
|
||
|
|
--kv-unified \
|
||
|
|
--cache-type-k q4_0 \
|
||
|
|
--cache-type-v q4_0 \
|
||
|
|
--batch-size 1024 \
|
||
|
|
--ubatch-size 512 \
|
||
|
|
--parallel 1 \
|
||
|
|
--cont-batching \
|
||
|
|
--host 0.0.0.0 \
|
||
|
|
--port "$CONTAINER_PORT"
|
||
|
|
|
||
|
|
echo "[*] Warte auf Modell-Bereitschaft (Completion-Check, max. 300 s) ..."
|
||
|
|
MODEL_READY=0
|
||
|
|
for i in {1..150}; do
|
||
|
|
HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" --max-time 10 \
|
||
|
|
-X POST "http://localhost:${HOST_PORT}/v1/chat/completions" \
|
||
|
|
-H "Content-Type: application/json" \
|
||
|
|
-d "{\"model\":\"${MODEL_ALIAS}\",\"messages\":[{\"role\":\"user\",\"content\":\"ping\"}],\"max_tokens\":1,\"temperature\":0.0,\"stream\":false}") || HTTP_CODE="000"
|
||
|
|
if [ "$HTTP_CODE" = "200" ]; then MODEL_READY=1; break; fi
|
||
|
|
echo " [${i}/150] HTTP ${HTTP_CODE:-000} — Modell laedt noch, warte 2s ..."
|
||
|
|
sleep 2
|
||
|
|
done
|
||
|
|
|
||
|
|
if [ "$MODEL_READY" -ne 1 ]; then
|
||
|
|
echo "[!] Modell wurde nicht rechtzeitig bereit (kein HTTP 200 auf Completion)." >&2
|
||
|
|
docker logs --tail 200 "$CONTAINER_NAME" || true
|
||
|
|
exit 1
|
||
|
|
fi
|
||
|
|
|
||
|
|
echo "[*] Modell bereit — erster Completion-Request erfolgreich (HTTP 200)."
|
||
|
|
echo "[*] Server laeuft auf http://0.0.0.0:${HOST_PORT}/v1 (Alias: ${MODEL_ALIAS})"
|
||
|
|
echo "[*] Setze im Gateway: LOCAL_LLM_BASE_URL=http://127.0.0.1:${HOST_PORT}/v1 LOCAL_LLM_MODEL=${MODEL_ALIAS}"
|
||
|
|
echo "[*] Stoppen mit: scripts/llm-server/stop-llm-server.sh"
|