#!/usr/bin/env bash
# Inside The Machine · one-command install: MLX LM
# Apple's MLX-native way to generate, chat, fine-tune and serve LLMs on Apple silicon.
# Page: https://insidethemachine.ai/setups/mlx-lm   Agent doc: https://insidethemachine.ai/install/mlx-lm.md
# Run status: Not run: commands checked against the official docs only. Commands checked against the official docs on 2026-10-15. Not run end to end.
# This script: The wrapper scripts are generated from the site data. Every .sh passed bash -n and shellcheck and ran in plan mode on Linux (Debian 13). Every .ps1 passed the PowerShell parser and chat.ps1 ran in plan, verify and rollback mode under PowerShell 7 on Linux with Windows calls mocked. Not run on macOS or real Windows.
#
# Read it first:  curl -fsSL https://insidethemachine.ai/install/mlx-lm.sh | less
#   curl -fsSL https://insidethemachine.ai/install/mlx-lm.sh | bash                        install, then check
#   curl -fsSL https://insidethemachine.ai/install/mlx-lm.sh | bash -s -- plan             print every step, run nothing
#   curl -fsSL https://insidethemachine.ai/install/mlx-lm.sh | bash -s -- verify           only the checks
#   curl -fsSL https://insidethemachine.ai/install/mlx-lm.sh | bash -s -- rollback --yes   undo it (without --yes it only prints)
# Options: ITM_RAM=32 (memory in GB instead of detecting it)
# Lines marked YOU RUN are left to you: servers that keep running, interactive apps, your own files.
# shellcheck disable=SC2016,SC1090
set -u
main() {
ACT="${ITM_ACTION:-install}"; YES="${ITM_YES:-}"
for a in "$@"; do case "$a" in --yes|-y) YES=1 ;; install|plan|verify|rollback) ACT="$a" ;; *) echo "unknown option: $a (use install, plan, verify or rollback)" >&2; exit 2 ;; esac; done
START="$(pwd)"; FAILS=0; NEXT=""
case "$(uname -s)" in Darwin) P=mac ;; Linux) P=linux ;; *) P=other ;; esac
say() { printf '\n== %s\n' "$1"; }
note() { printf '  # %s\n' "$1"; }
die() { printf '\nSTOP: %s\nHelp: %s\n' "$1" 'https://insidethemachine.ai/setups/mlx-lm' >&2; exit 1; }
run() { printf '  $ %s\n' "$1"; [ "$ACT" = plan ] && return 0; eval "$1" || die "this step failed: $1"; }
mac() { if [ "$P" != mac ]; then printf '  (macOS only, skipped: use your package manager) %s\n' "$1"; return 0; fi
  case "$1" in
    xcode-select*) printf '  $ %s\n' "$1"; [ "$ACT" = plan ] || xcode-select -p >/dev/null 2>&1 || die "install Apple's command line tools first: run xcode-select --install, finish the window, then run this again"; return 0 ;;
    brew*) [ "$ACT" = plan ] || command -v brew >/dev/null 2>&1 || die "needs Homebrew first: https://brew.sh" ;;
  esac
  run "$1"; }
you() { printf '  YOU RUN: %s\n' "$1"; [ -n "$2" ] && printf '           (%s)\n' "$2"; NEXT="$NEXT  $1\n"; }
undo() { if [ "${2:-}" = mac ] && [ "$P" != mac ]; then return 0; fi; printf '  $ %s\n' "$1"; if [ "$ACT" = rollback ] && [ -n "$YES" ]; then eval "$1" || echo "    (failed, carrying on)"; fi; }
check() { printf '  ? %s\n    expect: %s\n' "$1" "$2"; [ "$ACT" = plan ] && return 0; if out=$(eval "$1" 2>&1) && ! printf '%s' "$out" | grep -q '^{"error"'; then echo "    PASS"; else echo "    FAIL"; FAILS=$((FAILS + 1)); fi; [ -n "$out" ] && printf '%s\n' "$out" | head -n 3 | sed 's/^/    | /'; return 0; }
need_ollama() { [ "$ACT" = plan ] && return 0; ollama list >/dev/null 2>&1 && return 0; echo "  starting 'ollama serve' in the background (log: /tmp/itm-ollama.log)"; nohup ollama serve >/tmp/itm-ollama.log 2>&1 & i=0; while [ $i -lt 30 ]; do sleep 1; ollama list >/dev/null 2>&1 && return 0; i=$((i + 1)); done; die "Ollama is not running. Open the Ollama app, then run this again."; }
[ "$P" = other ] && die "this script is for macOS and Linux."
install_mac() {
  say '1. Install MLX LM'
  run 'python3 -m venv ~/.venvs/mlx && source ~/.venvs/mlx/bin/activate'
  run 'pip install mlx-lm'
  note 'or with conda: conda install -c conda-forge mlx-lm'
  say '2. Configure (your choices)'
  you 'mlx_lm.server --model mlx-community/Mistral-7B-Instruct-v0.3-4bit' 'starts an OpenAI-like server on localhost:8080; model downloads from Hugging Face on first run'
  you 'mlx_lm.server --model <hf_repo_or_path> --kv-bits 4 --prefill-step-size 512' 'quantized KV cache for long contexts'
}
verify_mac() {
  say 'Check'
  check 'mlx_lm.generate --model mlx-community/Mistral-7B-Instruct-v0.3-4bit --prompt "hello"' 'prints generated text and tokens-per-second stats'
  check 'curl -s http://localhost:8080/v1/chat/completions -H '\''Content-Type: application/json'\'' -d '\''{"messages":[{"role":"user","content":"Say hi"}]}'\''' 'JSON response with a choices[0].message.content field (server must be running)'
}
rollback_mac() {
  say 'Rollback'
  undo 'pip uninstall -y mlx-lm'
  undo 'rm -rf ~/.venvs/mlx'
  note 'downloaded models live in the Hugging Face cache (~/.cache/huggingface/hub); delete specific model folders to free space'
}
act() { [ -f ~/.venvs/mlx/bin/activate ] && . ~/.venvs/mlx/bin/activate; return 0; }
inwd() { return 0; }
V=$P
[ "$V" = linux ] && die "no Linux steps for MLX LM in our data. See https://insidethemachine.ai/setups/mlx-lm"

printf 'Inside The Machine · MLX LM · %s · %s%s\n' "$ACT" "$P" ""
case "$ACT" in
  install|plan) "install_$V"; cd "$START" || exit 1; act; inwd; "verify_$V"; cd "$START" || exit 1; [ "$ACT" = plan ] && "rollback_$V" ;;
  verify) act; inwd; "verify_$V" ;;
  rollback) [ -n "$YES" ] || note 'print only: add --yes to run these'; act; "rollback_$V" ;;
esac
if [ -n "$NEXT" ]; then say 'Now you run these (in order)'; printf '%b' "$NEXT"; fi
[ "$ACT" = install ] && printf '\nDone. %s check(s) failed. Next steps and fixes: %s\n' "$FAILS" 'https://insidethemachine.ai/setups/mlx-lm'
[ "$FAILS" -eq 0 ]
}
main "$@"