# LFM2.5-24B-A2B 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 "== LFM2.5-24B-A2B =="
if (Get-Command nvidia-smi -ErrorAction SilentlyContinue) { nvidia-smi --query-gpu=name,memory.total --format=csv,noheader }

if ($env:MODE -ne "hf") {
  Write-Host "== Ollama =="
  if (-not (Get-Command ollama -ErrorAction SilentlyContinue)) {
    Write-Host "Ollama wird installiert ..."
    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 }
  # Passende Variante für deine Hardware auswählen (eigene Wahl:  $env:MODEL="<name>")
  if (-not $env:MODEL) {
    $choices = @(@{need=17.0; ref="hf.co/LiquidAI/LFM2.5-24B-A2B-GGUF:Q4_K_M"; def=$true}, @{need=20.0; ref="hf.co/LiquidAI/LFM2.5-24B-A2B-GGUF:Q5_K_M"; def=$false}, @{need=23.0; ref="hf.co/LiquidAI/LFM2.5-24B-A2B-GGUF:Q6_K"; def=$false}, @{need=29.5; ref="hf.co/LiquidAI/LFM2.5-24B-A2B-GGUF:Q8_0"; def=$false}, @{need=54.0; ref="hf.co/LiquidAI/LFM2.5-24B-A2B-GGUF:BF16"; def=$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 "Verfügbarer Speicher: ca. $mem GB  ->  gewählte Variante: $($env:MODEL)" -ForegroundColor Cyan
  }
  if (-not $env:MODEL) { $env:MODEL = "hf.co/LiquidAI/LFM2.5-24B-A2B-GGUF" }
  Write-Host "Lade $($env:MODEL) herunter (beim ersten Mal dauert das je nach Größe einige Minuten) ..."
  ollama pull $env:MODEL
  Write-Host ""
  Write-Host "Fertig! Tippe jetzt deine Frage ein und drücke Enter. Beenden mit /bye" -ForegroundColor Green
  ollama run $env:MODEL
  exit
}

Write-Host "== Python / Hugging Face =="
python -m venv .venv-lfm2-5-24b-a2b
& .\.venv-lfm2-5-24b-a2b\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