# NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 lokal ausprobieren – Windows (PowerShell)
# Automatisch erzeugt von Tokenstrom (tokenstrom.de).
# Aufruf:  powershell -ExecutionPolicy Bypass -File try_windows.ps1
#         $env:MODE='hf'; powershell -ExecutionPolicy Bypass -File try_windows.ps1   (Python statt Ollama)
$ErrorActionPreference = "Stop"
Set-Location $PSScriptRoot
Write-Host "== NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 =="
if (Get-Command nvidia-smi -ErrorAction SilentlyContinue) { nvidia-smi --query-gpu=name,memory.total --format=csv,noheader }
Write-Host "WARNUNG: Groß (65.8 Mrd. Parameter): braucht auch quantisiert ~41.0 GB VRAM/RAM. Colab Free (T4, 15 GB) reicht nicht." -ForegroundColor Yellow

if ($env:MODE -ne "hf") {
  Write-Host "== Ollama =="
  if (-not (Get-Command ollama -ErrorAction SilentlyContinue)) {
    winget install -e --id Ollama.Ollama --accept-source-agreements --accept-package-agreements
    $env:Path = [Environment]::GetEnvironmentVariable("Path","Machine") + ";" + [Environment]::GetEnvironmentVariable("Path","User")
  }
  try { Invoke-WebRequest -UseBasicParsing http://localhost:11434/api/tags | Out-Null } catch { Start-Process -WindowStyle Hidden ollama -ArgumentList "serve"; Start-Sleep 3 }
  ollama pull "hf.co/lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr-GGUF"
  ollama run "hf.co/lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr-GGUF"
  exit
}

Write-Host "== Python / Hugging Face =="
python -m venv .venv-nvidia-nemotron-3-5-super-vl-1
& .\.venv-nvidia-nemotron-3-5-super-vl-1\Scripts\Activate.ps1
python -m pip install -q -U pip
if (Get-Command nvidia-smi -ErrorAction SilentlyContinue) {
  pip install -q -U torch --index-url https://download.pytorch.org/whl/cu124
  pip install -q -U transformers accelerate pillow scipy
} else { pip install -q -U torch transformers accelerate pillow scipy }

python demo.py