#!/bin/zsh set -euo pipefail support_root=${0:A:h} model_root=${support_root:h} models_root=${model_root:h} base_path=${OMLX_BASE_PATH:-$model_root/.omlx} cache_dir=${OMLX_CACHE_DIR:-$base_path/cache} api_key=${OMLX_API_KEY:-omlx} port=${OMLX_PORT:-8766} mkdir -p "$cache_dir" if [[ ! -f "$base_path/model_settings.json" ]]; then cp "$support_root/model_settings.json" "$base_path/model_settings.json" fi export OMLX_QWEN4_PLE_MODE=mmap export OMLX_QWEN4_PLE_MODEL_PATH="$model_root" exec /bin/zsh "$support_root/with-omlx-python" -m omlx.cli serve \ --model-dir "$models_root" \ --host 127.0.0.1 \ --port "$port" \ --max-concurrent-requests 1 \ --memory-guard balanced \ --paged-ssd-cache-dir "$cache_dir" \ --paged-ssd-cache-max-size 128GB \ --hot-cache-max-size 0 \ --initial-cache-blocks 1 \ --no-hf-cache \ --base-path "$base_path" \ --api-key "$api_key"