A high-performance inference engine for AI models
#!/usr/bin/env bash
set -euo pipefail
PROJECT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
"$PROJECT_DIR/bootstrap.sh"
ARGS=()
while [[ $# -gt 0 ]]; do
case "$1" in
-m|--model|-d|--draft-model)
FLAG="$1"
if [[ $# -lt 2 || "$2" == -* ]]; then
echo "$FLAG requires a model directory or Hugging Face repository" >&2
exit 1
fi
MODEL_PATH="$2"
if [[ ! -d "$MODEL_PATH" ]]; then
# Resolve both checkpoints before native inference. Python only downloads.
MODEL_PATH="$(uv run --project "$PROJECT_DIR/.." python -c \
'import sys; from common import get_model_path; print(get_model_path(sys.argv[1]))' "$MODEL_PATH")"
fi
ARGS+=("$FLAG" "$MODEL_PATH")
shift 2
;;
*)
ARGS+=("$1")
shift
;;
esac
done
exec "$PROJECT_DIR/zig-out/bin/engine-mlxserve" "${ARGS[@]}"