llama.cpp updater on Windows

README.md

This is my approach to self-updating llama.cpp.

Initial setup

1. Create all the files listed in the folder_structure.txt file from this gist.
2. Open the latest llama.cpp release on GitHub, for example: https://github.com/ggml-org/llama.cpp/releases/tag/b9587
3. Download the required archives. In my case, these were:
   • llama-b9587-bin-win-cuda-13.3-x64.zip
   • cudart-llama-bin-win-cuda-13.3-x64.zip
4. Extract both archives into the files folder.
5. Copy the SHA of cudart-llama-bin-win-cuda-13.3-x64.zip to cuda-dll.sha256. In my case, it was: sha256:1462a050eb4c684921ba51dcc4cc488a036674c3e73e9945ee705b854808d03e
6. In version.txt, place the name of the most recent release. In my case, this was: b9587

Updating

To update, simply run the update.bat script. It automatically checks for a new version and skips downloading cudart-llama-bin-win-cuda-13.3-x64.zip if the SHA has not changed. This saves roughly 370 MB of unnecessary downloads.

cuda-dll.sha256

sha256:1462a050eb4c684921ba51dcc4cc488a036674c3e73e9945ee705b854808d03e

folder_structure.txt

files (dir)
cuda-dll.sha256
update.bat
version.txt

update.bat

@echo off
setlocal
powershell -NoProfile -ExecutionPolicy Bypass -Command "$p='%~f0'; $c=Get-Content -Raw -LiteralPath $p; $m='### POWERSHELL ###'; $i=$c.LastIndexOf($m); if ($i -lt 0) { throw 'PowerShell payload marker was not found.' }; Invoke-Expression $c.Substring($i + $m.Length)"
if errorlevel 1 (
    echo Update failed.
    exit /b 1
)
exit /b 0

### POWERSHELL ###
$ErrorActionPreference = 'Stop'
$ProgressPreference = 'SilentlyContinue'

$root = (Get-Location).Path
$filesDir = Join-Path $root 'files'
$versionFile = Join-Path $root 'version.txt'
$cudaShaFile = Join-Path $root 'cuda-dll.sha256'
$repo = 'ggml-org/llama.cpp'
$headers = @{ 'User-Agent' = 'llama-cpp-windows-cuda-updater' }

function Get-AssetDigest {
    param([object] $Asset)

    if ($Asset.PSObject.Properties.Name -contains 'digest' -and $Asset.digest) {
        return [string] $Asset.digest
    }

    return ''
}

function Find-OneAsset {
    param(
        [object[]] $Assets,
        [string] $Kind,
        [string] $Pattern
    )

    $matches = @(
        $Assets |
            Where-Object { $_.name -like $Pattern } |
            Sort-Object name -Descending
    )

    if ($matches.Count -eq 0) {
        throw ('Could not find {0} asset matching {1}' -f $Kind, $Pattern)
    }

    if ($matches.Count -gt 1) {
        Write-Host ('Found several {0} assets; using {1}' -f $Kind, $matches[0].name)
    }

    return $matches[0]
}

if (-not (Test-Path -LiteralPath $filesDir)) {
    New-Item -ItemType Directory -Path $filesDir | Out-Null
}

$currentVersion = ''
if (Test-Path -LiteralPath $versionFile) {
    $currentVersion = (Get-Content -Raw -LiteralPath $versionFile).Trim()
}

$currentCudaDigest = ''
if (Test-Path -LiteralPath $cudaShaFile) {
    $currentCudaDigest = (Get-Content -Raw -LiteralPath $cudaShaFile).Trim()
}

Write-Host ('Checking latest release for {0}...' -f $repo)
$release = Invoke-RestMethod `
    -Headers $headers `
    -Uri ('https://api.github.com/repos/{0}/releases/latest' -f $repo)

$latestVersion = [string] $release.tag_name
if ([string]::IsNullOrWhiteSpace($latestVersion)) {
    throw 'GitHub response did not contain tag_name'
}

$assets = @($release.assets)
$llamaAsset = Find-OneAsset `
    -Assets $assets `
    -Kind 'llama.cpp CUDA 13 Windows x64' `
    -Pattern ('llama-{0}-bin-win-cuda-13.*-x64.zip' -f $latestVersion)
$cudaAsset = Find-OneAsset `
    -Assets $assets `
    -Kind 'CUDA 13 DLL' `
    -Pattern 'cudart-llama-bin-win-cuda-13.*-x64.zip'

$cudaDigest = Get-AssetDigest $cudaAsset
$versionChanged = ($latestVersion -ne $currentVersion)
$cudaChanged = $true

if ($cudaDigest -and $currentCudaDigest -and ($cudaDigest -eq $currentCudaDigest)) {
    $cudaChanged = $false
}

if (-not $versionChanged -and -not $cudaChanged) {
    Write-Host ('Already up to date: {0}' -f $latestVersion)
    exit 0
}

if (-not $versionChanged -and $cudaDigest -and -not $currentCudaDigest) {
    Set-Content -Encoding ASCII -NoNewline -Path $cudaShaFile -Value $cudaDigest
    Write-Host ('Version is current ({0}); recorded CUDA DLL digest.' -f $latestVersion)
    exit 0
}

$tmp = Join-Path $env:TEMP ('llama-cpp-update-' + [guid]::NewGuid().ToString('N'))
New-Item -ItemType Directory -Path $tmp | Out-Null

try {
    if ($versionChanged) {
        $llamaZip = Join-Path $tmp $llamaAsset.name
        Write-Host ('Downloading {0}...' -f $llamaAsset.name)
        Invoke-WebRequest `
            -Headers $headers `
            -Uri $llamaAsset.browser_download_url `
            -OutFile $llamaZip

        Write-Host 'Extracting llama.cpp binaries...'
        Expand-Archive -Force -Path $llamaZip -DestinationPath $tmp
        Copy-Item `
            -Force `
            -Recurse `
            -Path (Join-Path $tmp '*') `
            -Destination $filesDir `
            -Exclude '*.zip'
    }

    if ($cudaChanged) {
        if (-not $cudaDigest) {
            Write-Host 'GitHub did not provide a digest for the CUDA DLL asset; downloading DLL archive.'
        }

        $cudaZip = Join-Path $tmp $cudaAsset.name
        Write-Host ('Downloading {0}...' -f $cudaAsset.name)
        Invoke-WebRequest `
            -Headers $headers `
            -Uri $cudaAsset.browser_download_url `
            -OutFile $cudaZip

        Write-Host 'Extracting CUDA DLLs...'
        Expand-Archive -Force -Path $cudaZip -DestinationPath $tmp
        Copy-Item `
            -Force `
            -Recurse `
            -Path (Join-Path $tmp '*') `
            -Destination $filesDir `
            -Exclude '*.zip'
    }

    Set-Content -Encoding ASCII -NoNewline -Path $versionFile -Value $latestVersion
    if ($cudaDigest) {
        Set-Content -Encoding ASCII -NoNewline -Path $cudaShaFile -Value $cudaDigest
    }

    Write-Host ('Updated to {0}' -f $latestVersion)
}
finally {
    if (Test-Path -LiteralPath $tmp) {
        Remove-Item -Recurse -Force -LiteralPath $tmp
    }
}

version.txt

b9567