# GPU Fit benchmark for Windows: measures how fast a model runs in your local Ollama and, if you
# say yes, shares the result at https://gpufit.dev so others can see real speeds for their hardware.
#
#   irm https://gpufit.dev/bench.ps1 | iex
#
# Sends: GPU name, GPU memory, GPU count, OS name, Ollama version, model tag, quantization, how much
# of the model was on the GPU, context length and measured speeds. Nothing else: no hostname,
# username or files. Set $env:GPUFIT_MODEL to skip the model picker.
& {
  $ErrorActionPreference = 'Stop'
  $Site = if ($env:GPUFIT_SITE) { $env:GPUFIT_SITE } else { 'https://gpufit.dev' }
  $OllamaHost = if ($env:OLLAMA_HOST) { $env:OLLAMA_HOST } else { 'http://127.0.0.1:11434' }
  if ($OllamaHost -notmatch '^https?://') { $OllamaHost = "http://$OllamaHost" }
  $OllamaHost = $OllamaHost.TrimEnd('/') -replace '0\.0\.0\.0', '127.0.0.1'
  $Runs = 3; $Tokens = 256; $Ctx = 4096
  function Post($path, $body) { Invoke-RestMethod -Method Post -Uri "$OllamaHost$path" -ContentType 'application/json' -Body ($body | ConvertTo-Json -Depth 5 -Compress) }
  function Median($xs) { $s = @($xs | Sort-Object); if (-not $s.Count) { return $null }; $m = [math]::Floor($s.Count / 2); if ($s.Count % 2) { $s[$m] } else { ($s[$m - 1] + $s[$m]) / 2 } }

  try { $version = (Invoke-RestMethod "$OllamaHost/api/version").version }
  catch { Write-Host "Ollama isn't answering at $OllamaHost. Open the Ollama app and try again." -ForegroundColor Red; return }

  $model = $env:GPUFIT_MODEL
  if (-not $model) {
    $list = @((Invoke-RestMethod "$OllamaHost/api/tags").models | ForEach-Object { $_.name } | Where-Object { $_ -notmatch 'embed' })
    if (-not $list.Count) { Write-Host 'You have no models in Ollama yet. Try: ollama pull qwen3:8b' -ForegroundColor Red; return }
    Write-Host 'Your Ollama models:'
    for ($i = 0; $i -lt $list.Count; $i++) { Write-Host ("  {0}) {1}" -f ($i + 1), $list[$i]) }
    $pick = Read-Host 'Benchmark which one? [1]'
    if (-not $pick) { $pick = '1' }
    $model = if ($pick -match '^\d+$') { $list[[int]$pick - 1] } else { $pick }
    if (-not $model) { Write-Host 'No model picked.' -ForegroundColor Red; return }
  }
  try { $show = Post '/api/show' @{ model = $model } }
  catch { Write-Host "Ollama doesn't have '$model'. Pull it first: ollama pull $model" -ForegroundColor Red; return }
  $quant = if ($show.details.quantization_level) { $show.details.quantization_level } else { 'unknown' }

  # GPU: nvidia-smi gives exact memory; otherwise the display adapter name (AMD/Intel).
  $gpu = 'unknown'; $mem = $null; $count = 1
  $smi = Get-Command nvidia-smi -ErrorAction SilentlyContinue
  if ($smi) {
    $lines = @(& $smi.Source --query-gpu=name,memory.total --format=csv,noheader,nounits | Where-Object { $_ })
    if ($lines.Count) { $count = $lines.Count; $p = $lines[0].Split(','); $gpu = $p[0].Trim(); $mem = [math]::Round([double]$p[1] / 1024, 1) }
  } else {
    $vc = Get-CimInstance Win32_VideoController | Where-Object { $_.Name -notmatch 'Microsoft|Intel|Virtual|Parsec' } | Select-Object -First 1
    if ($vc) { $gpu = $vc.Name }
    # AdapterRAM is capped at 4 GB, so read the real size from the driver registry key when it's there.
    try {
      $key = Get-ChildItem 'HKLM:\SYSTEM\ControlSet001\Control\Class\{4d36e968-e325-11ce-bfc1-08002be10318}' -ErrorAction Stop | Where-Object { (Get-ItemProperty $_.PSPath).DriverDesc -eq $gpu } | Select-Object -First 1
      $q = (Get-ItemProperty $key.PSPath).'HardwareInformation.qwMemorySize'
      if ($q) { $mem = [math]::Round([double]$q / 1GB, 1) }
    } catch {}
  }
  $gpu = ($gpu -replace '[\"\\]', '').Trim()

  $prompt = 'Write a detailed, practical guide to growing tomatoes on a balcony. Cover choosing a variety, containers, soil, watering, feeding, pruning, pests and harvesting.'
  $gen = { param($n) Post '/api/generate' @{ model = $model; prompt = $prompt; stream = $false; options = @{ num_predict = $n; num_ctx = $Ctx; temperature = 0; seed = 42 } } }
  Write-Host "Loading $model ($quant)..."
  $null = & $gen 1
  $share = $null
  try { $ps = (Invoke-RestMethod "$OllamaHost/api/ps").models | Where-Object { $_.name -eq $model -or $_.name -eq "$($model):latest" } | Select-Object -First 1
        if ($ps -and $ps.size -gt 0) { $share = [math]::Round($ps.size_vram / $ps.size, 3) } } catch {}
  $gens = @(); $pps = @()
  for ($i = 1; $i -le $Runs; $i++) {
    Write-Host "Run $i of ${Runs}: generating $Tokens tokens..."
    $r = & $gen $Tokens
    if ($r.eval_duration -gt 0) { $gens += [math]::Round($r.eval_count / ($r.eval_duration / 1e9), 2) }
    if ($r.prompt_eval_duration -gt 0 -and $r.prompt_eval_count -gt 8) { $pps += [math]::Round($r.prompt_eval_count / ($r.prompt_eval_duration / 1e9), 1) }
  }
  $genTps = Median $gens; $promptTps = Median $pps
  if (-not $genTps) { Write-Host "Ollama didn't report timings; update Ollama and try again." -ForegroundColor Red; return }
  Write-Host ''
  Write-Host "$model ($quant) on ${gpu}: $genTps tokens/s generating$(if ($promptTps) { ", $promptTps tokens/s reading the prompt" })." -ForegroundColor Green
  if ($share -ne $null -and $share -lt 0.99) { Write-Host "Note: only $([int]($share * 100))% of the model is on the GPU, so part of it runs from system RAM." }

  $body = [ordered]@{ gpu = $gpu; mem_gb = $mem; gpu_count = $count; model = $model; quant = $quant; gen_tps = $genTps; prompt_tps = $promptTps; ctx = $Ctx; gpu_share = $share; runtime = 'ollama'; runtime_version = $version; os = 'windows'; client = 'bench.ps1/1' } | ConvertTo-Json -Compress
  Write-Host ''
  Write-Host "This is everything that would be sent to ${Site}:"
  Write-Host "  $body"
  $ok = Read-Host 'Share it so others can see real speeds for this hardware? [Y/n]'
  if ($ok -match '^[nN]') { Write-Host 'Not sent.'; return }
  try {
    $res = Invoke-RestMethod -Method Post -Uri "$Site/api/bench" -ContentType 'application/json' -Body $body
    Write-Host $res.message -ForegroundColor Green
  } catch {
    $msg = $_.ErrorDetails.Message; try { $msg = ($msg | ConvertFrom-Json).error } catch {}
    Write-Host "Not accepted: $msg" -ForegroundColor Red
  }
}