#!/bin/bash

# Large model: auto-select mmap based on GPU layers
# <= 45 layers: use mmap (less VRAM usage, faster startup)
# > 45 layers: disable mmap (avoids SVM limits)
LAYERS=${CODER_TOP_GPU_LAYERS:-55}
MMAP_OPT=""
if [ "$LAYERS" -gt 45 ]; then
  MMAP_OPT="--no-mmap"
  echo "Using --no-mmap (layers=$LAYERS > 45)"
else
  echo "Using mmap (layers=$LAYERS <= 45)"
fi
exec /app/llama-server $CODER_TOP_MODEL \
  -c $CODER_CONTEXT_SIZE -ngl $LAYERS -n $CODER_MAX_TOKENS \
  --temp 0.45 --top-p 0.9 --top-k 40 --repeat-penalty 1.12 \
  $MMAP_OPT \
  --flash-attn auto --threads -1 --threads-batch -1 --threads-http -1 \
  --jinja \
  --timeout 1200 --host 0.0.0.0 --port 8096 &
PID=$!

cleanup() {
    echo "Stopping llama-server..."
    kill $PID 2>/dev/null
    wait $PID 2>/dev/null
    exit 0
}

trap cleanup SIGTERM SIGINT

wait $PID