#!/bin/sh
# Serve Qwen 3.8 27B (4-bit, the official baseline build) for the harness on the Mac mini. Run this on the MacBook.
# Same server flags as train/serve.sh (the official baseline); only the host differs (0.0.0.0) and caffeinate keeps the
# MacBook awake as long as the server process lives (caffeinate -w <server pid>). Usage: ~/serve_remote.sh        Stop: kill $(cat ~/qwen_server.pid)
MODEL="$HOME/models/Qwen3.8-27B-4bit"
[ -f "$MODEL/config.json" ] || { echo "model not found: $MODEL" >&2; exit 1; }
# the mlx_lm of the harness venv if the repo is on this Mac, else the one on PATH
if [ -x "$HOME/projects/abap-llm/harness/train/.venv/bin/mlx_lm.server" ]; then
  SERVER="$HOME/projects/abap-llm/harness/train/.venv/bin/mlx_lm.server"
elif command -v mlx_lm.server >/dev/null 2>&1; then
  SERVER="$(command -v mlx_lm.server)"
else
  echo "mlx_lm.server not found (pip install mlx-lm==0.32.0)" >&2; exit 1
fi
"$SERVER" \
  --model "$MODEL" \
  --host 0.0.0.0 --port 8080 \
  --temp 0.2 --top-p 0.95 --top-k 20 --min-p 0 \
  --max-tokens 32768 \
  --prompt-cache-size 4 --prompt-cache-bytes 6000000000 \
  --chat-template-args '{"enable_thinking": true, "reasoning_effort": "medium"}' &
SP=$!
echo "$SP" > "$HOME/qwen_server.pid"          # the pid of the server itself
exec caffeinate -dimsu -w "$SP"                # keeps the Mac awake until the server process ends