#!/usr/bin/env bash
set -euo pipefail

HF_HOME="${HF_HOME:-/home/dschlueter/nvme2n1p7_home/huggingface}"
MODEL_REL_PATH="models/qwen3/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-IQ4_XS.gguf"
IMAGE="ghcr.io/ggml-org/llama.cpp:server-cuda"
CONTAINER_NAME="qwen36-27b-judge"
HOST_PORT=8002
CONTAINER_PORT=8000
MODEL_ALIAS="qwen3.5-judge"

echo "[*] Verwende HF_HOME = $HF_HOME"
if [ ! -f "$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 für Judge ..."
docker run -d \
  --gpus '"device=1,2"' \
  --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 "${MODEL_ALIAS}" \
    -c 262144 \
    -n 16384 \
    --jinja \
    --chat-template-kwargs '{"enable_thinking":true}' \
    --no-context-shift \
    --temp 0.7 \
    --top-p 0.80 \
    --top-k 20 \
    --min-p 0.01 \
    --repeat-penalty 1.05 \
    --main-gpu 0 \
    --tensor-split 0.5,0.5 \
    -ngl 999 \
    -fa on \
    --kv-unified \
    --cache-type-k q4_0 \
    --cache-type-v q4_0 \
    --batch-size 512 \
    --ubatch-size 256 \
    --parallel 1 \
    --cont-batching \
    --host 0.0.0.0 \
    --port "$CONTAINER_PORT"

echo "[*] Warte auf Modell-Bereitschaft (Completion-Check, max. 180 s) ..."
MODEL_READY=0
for i in {1..90}; 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\":\"${MODEL_ALIAS}\",\"messages\":[{\"role\":\"user\",\"content\":\"ping\"}],\"max_tokens\":1,\"temperature\":0.0,\"stream\":false}")
  if [ "$HTTP_CODE" = "200" ]; then MODEL_READY=1; break; fi
  echo "  [${i}/90] HTTP ${HTTP_CODE:-000} — Modell lädt 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 läuft auf http://0.0.0.0:${HOST_PORT}"
echo "[*] Stoppen mit: docker rm -f ${CONTAINER_NAME}"