#!/usr/bin/env bash
# NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 lokal ausprobieren – macOS
# Automatisch erzeugt von Tokenstrom (tokenstrom.de).
# Aufruf: bash try_mac.sh            # Ollama (Standard)
#        MODE=hf bash try_mac.sh    # Python/transformers
set -e
cd "$(dirname "$0")"
echo "== NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 =="
echo "WARNUNG: Groß (65.8 Mrd. Parameter): braucht auch quantisiert ~41.0 GB VRAM/RAM. Colab Free (T4, 15 GB) reicht nicht."

if [ "${MODE:-ollama}" = "ollama" ]; then
  echo "== Ollama =="
  if ! command -v ollama >/dev/null 2>&1; then
    if command -v brew >/dev/null 2>&1; then brew install ollama; else
      echo "Bitte Ollama installieren: https://ollama.com/download/mac"; exit 1; fi
  fi
  curl -s http://localhost:11434/api/tags >/dev/null 2>&1 || (ollama serve >/dev/null 2>&1 &)
  sleep 3
  ollama pull "hf.co/lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr-GGUF"
  echo "Starte Chat – beenden mit /bye"
  exec ollama run "hf.co/lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr-GGUF"
fi

echo "== Python / Hugging Face =="
python3 -m venv .venv-nvidia-nemotron-3-5-super-vl-1
source .venv-nvidia-nemotron-3-5-super-vl-1/bin/activate
pip install -q -U pip
pip install -q -U torch transformers accelerate pillow scipy

python3 demo.py