#!/usr/bin/env bash
# Starts or stops the two servers on their own, without the assistant.
#
# Handy while developing: loading the models takes more than a minute, and
# this way the Rust binary can be restarted as often as needed without paying
# for it again. The assistant detects they are already listening and reuses them.
#
#   scripts/servers.sh start | stop | status
set -euo pipefail
cd "$(dirname "$0")/.."

LOGS="${ASIST_LOGS:-logs}"
LLM_PORT="${ASIST_LLM_PORT:-${ASIST_LLM_PUERTO:-8012}}"
TTS_PORT="${ASIST_TTS_PORT:-${ASIST_TTS_PUERTO:-8013}}"
mkdir -p "$LOGS"

alive() { curl -sf -m 2 "http://127.0.0.1:$1/health" >/dev/null 2>&1; }

wait_for() { # port name seconds
  local end=$(( SECONDS + $3 ))
  while [ $SECONDS -lt $end ]; do
    alive "$1" && { echo "  $2 ready"; return 0; }
    sleep 1
  done
  echo "  $2 did NOT answer within $3 s; see $LOGS/$2.log" >&2
  return 1
}

start() {
  if alive "$LLM_PORT"; then
    echo "  llama-server was already up"
  else
    echo "  starting llama-server…"
    nohup vendor/llama.cpp/build/bin/llama-server \
      --model models/Qwen3.5-2B.Q5_K_M.gguf \
      --mmproj models/mmproj-BF16.gguf \
      --host 127.0.0.1 --port "$LLM_PORT" \
      --threads 10 --threads-batch 10 \
      --batch-size 512 --ubatch-size 256 \
      --gpu-layers 10 --split-mode layer --tensor-split 1 --main-gpu 0 \
      --no-mmap --ctx-size 8192 --parallel 2 --cache-ram 6144 \
      --rope-freq-base 1000000 --rope-freq-scale 0.25 \
      --jinja --chat-template-file config/qwen35-no-think.jinja \
      > "$LOGS/llama-server.log" 2>&1 &
  fi

  if alive "$TTS_PORT"; then
    echo "  tts-server was already up"
  else
    echo "  starting tts-server…"
    # --codec-chunk-dur 1.0 is what makes audio come out in blocks as it is
    # generated instead of all at the end. See docs/RENDIMIENTO.md.
    nohup vendor/qwentts.cpp/build/tts-server \
      --model models/qwen-talker-1.7b-base-Q8_0.gguf \
      --codec models/qwen-tokenizer-12hz-Q8_0.gguf \
      --host 127.0.0.1 --port "$TTS_PORT" \
      --lang spanish --codec-chunk-dur 1.0 \
      > "$LOGS/tts-server.log" 2>&1 &
  fi

  wait_for "$LLM_PORT" llama-server 240
  wait_for "$TTS_PORT" tts-server 240
}

stop() {
  # SIGTERM, not SIGKILL: they need a chance to release the GPU.
  pkill -TERM -f 'llama-server .*--port '"$LLM_PORT" 2>/dev/null && echo "  llama-server stopped" || true
  pkill -TERM -f 'tts-server .*--port '"$TTS_PORT"   2>/dev/null && echo "  tts-server stopped"   || true
}

status() {
  alive "$LLM_PORT" && echo "  llama-server  up  :$LLM_PORT" || echo "  llama-server  down"
  alive "$TTS_PORT" && echo "  tts-server    up  :$TTS_PORT" || echo "  tts-server    down"
  command -v nvidia-smi >/dev/null && \
    nvidia-smi --query-gpu=memory.used,memory.total --format=csv,noheader | sed 's/^/  GPU: /'
}

# The Spanish verbs (arrancar/parar/estado) are still accepted.
case "${1:-status}" in
  start|arrancar) start ;;
  stop|parar)     stop ;;
  status|estado)  status ;;
  *) echo "usage: $0 {start|stop|status}" >&2; exit 1 ;;
esac