젬마 26b 번역용 llama.cpp 세팅 (VRAM 16기가)

$ErrorActionPreference = "Stop"

$MODEL_DIR = "__________"

$TargetModel = Join-Path $MODEL_DIR "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf"
$DraftModel  = Join-Path $MODEL_DIR "mtp-gemma-4-26B-A4B-it-Q4_0.gguf"

if (-not (Test-Path $TargetModel)) {
    throw "Target model not found: $TargetModel"
}

if (-not (Test-Path $DraftModel)) {
    throw "Draft model not found: $DraftModel"
}

$ServerCommand = Get-Command "llama-server.exe" -ErrorAction SilentlyContinue

if (-not $ServerCommand) {
    throw "llama-server.exe was not found in PATH."
}

$llamaArgs = @(
    # Models
    "-m"
    $TargetModel

    "-md"
    $DraftModel

    "-a"
    "gemma-4-26b"

    "--reasoning" 
    "off"

    # MTP speculative decoding
    "--spec-type"
    "draft-mtp"

    "--spec-draft-n-max"
    "2"

    "--no-spec-draft-backend-sampling"

    # Full GPU offload
    "-ngl"
    "all"

    "-ngld"
    "all"

    "--fit"
    "off"

    "-fa"
    "on"

    # Five slots sharing one 2048-token context pool
    "-c"
    "2048"

    "-np"
    "5"

    "--kv-unified"

    # Quantized target and draft KV caches
    # "-ctk"
    # "q8_0"

    # "-ctv"
    # "q8_0"

    # "-ctkd"
    # "q8_0"

    # "-ctvd"
    # "q8_0"

    # Batch configuration
    "-b"
    "1024"

    "-ub"
    "128"

    # Google-recommended Gemma 4 sampling
    "--temp"
    "1.0"

    "--top-p"
    "0.95"

    "--top-k"
    "64"

    # Disable additional llama.cpp sampling filters
    "--min-p"
    "0.0"

    "--typical"
    "1.0"

    "--top-n-sigma"
    "-1.0"

    "--xtc-probability"
    "0.0"

    "--dynatemp-range"
    "0.0"

    "--mirostat"
    "0"

    # Disable repetition and presence/frequency penalties
    "--repeat-penalty"
    "1.0"

    "--presence-penalty"
    "0.0"

    "--frequency-penalty"
    "0.0"

    "--dry-multiplier"
    "0.0"

    # Memory and server options
    "--ctx-checkpoints"
    "0"

    "--cache-ram"
    "0"

    "--no-cache-idle-slots"
    "--no-mmproj"

    "--host"
    "127.0.0.1"

    "--port"
    "8080"
)

Write-Host "Starting Gemma 4 llama.cpp server..." -ForegroundColor Cyan
Write-Host "Executable: $($ServerCommand.Source)"
Write-Host "Target:     $TargetModel"
Write-Host "MTP:        $DraftModel"
Write-Host "Endpoint:   http://127.0.0.1:8080"
Write-Host "Sampling:   temp=1.0, top-p=0.95, top-k=64, min-p=disabled"
Write-Host ""

& $ServerCommand.Source @llamaArgs

if ($LASTEXITCODE -ne 0) {
    throw "llama-server exited with code $LASTEXITCODE"
}