# Beam-1.0-501B-A23B lokal ausprobieren – Windows (PowerShell)
# Automatisch erzeugt von Tokenstrom (tokenstrom.com).
# 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 "== Beam-1.0-501B-A23B =="
if (Get-Command nvidia-smi -ErrorAction SilentlyContinue) { nvidia-smi --query-gpu=name,memory.total --format=csv,noheader }
Write-Host "WARNUNG: Groß (501,3 B parameters, MoE): running entirely on the GPU takes a lot of VRAM (video memory). With MoE offloading, a regular graphics card plus plenty of RAM is enough (see table). Colab Free isn't enough." -ForegroundColor Yellow

if ($env:MODE -ne "hf") {
  Write-Host "== Ollama =="
  if (-not (Get-Command ollama -ErrorAction SilentlyContinue)) {
    Write-Host "Installing Ollama ..."
    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 }
  # Pick the right variant for your hardware (to choose your own:  $env:MODEL="<name>")
  if (-not $env:MODEL) {
    $choices = @(@{need=207.0; ref="hf.co/unsloth/Beam-1.0-501B-A23B-GGUF:UD-Q2_K_XL"; def=$false}, @{need=268.0; ref="hf.co/unsloth/Beam-1.0-501B-A23B-GGUF:UD-Q3_K_XL"; def=$false}, @{need=338.5; ref="hf.co/unsloth/Beam-1.0-501B-A23B-GGUF:UD-Q4_K_XL"; def=$true}, @{need=407.5; ref="hf.co/unsloth/Beam-1.0-501B-A23B-GGUF:UD-Q5_K_XL"; def=$false}, @{need=500.0; ref="hf.co/unsloth/Beam-1.0-501B-A23B-GGUF:UD-Q6_K_XL"; def=$false}, @{need=587.5; ref="hf.co/unsloth/Beam-1.0-501B-A23B-GGUF:Q8_0"; def=$false}, @{need=1104.0; ref="hf.co/unsloth/Beam-1.0-501B-A23B-GGUF:BF16"; def=$false})
    $mem = 0
    if (Get-Command nvidia-smi -ErrorAction SilentlyContinue) {
      $mem = [math]::Floor(((nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits) | ForEach-Object { [int]$_ } | Measure-Object -Maximum).Maximum / 1024)
    } else {
      $mem = [math]::Floor((Get-CimInstance Win32_ComputerSystem).TotalPhysicalMemory / 1GB) - 4
    }
    $pick = $choices | Where-Object { $_.def -and $_.need -le $mem } | Select-Object -First 1
    if (-not $pick) { $pick = $choices | Where-Object { $_.need -le $mem } | Select-Object -Last 1 }
    if (-not $pick) { $pick = $choices[0] }
    $env:MODEL = $pick.ref
    Write-Host "Available memory: approx. $mem GB  ->  selected variant: $($env:MODEL)" -ForegroundColor Cyan
  }
  if (-not $env:MODEL) { $env:MODEL = "beam:cloud" }
  Write-Host "Downloading $($env:MODEL) (the first time this takes a few minutes, depending on size) ..."
  ollama pull $env:MODEL
  Write-Host ""
  Write-Host "Done! Now type your question and press Enter. Quit with /bye" -ForegroundColor Green
  ollama run $env:MODEL
  exit
}

Write-Host "== Python / Hugging Face =="
python -m venv .venv-beam-1-0-501b-a23b
& .\.venv-beam-1-0-501b-a23b\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 bitsandbytes
} else { pip install -q -U torch transformers accelerate }

python demo.py