llama.cpp PowerShell utils

llama_cpp_utils.ps1

# Collection of variables, aliases and Functions to work w/ locally hosted LLMs
# Source to activate.

$env:LLM_VENV_ACTIVE = 0
$env:LLM_PYTHON_VENV_PATH = "$env:USERPROFILE\.llm.venv"
$env:LLAMA_ARG_HOST = "steelph0enix.pc"
$env:LLAMA_ARG_PORT = 51536

# llm python venv stuff
Function llm-venv-activate {
    if ($env:LLM_VENV_ACTIVE -eq 1) {
        Write-Host "llama.cpp virtual environment is active!"
        return
    }

    if (-Not (Test-Path -Path $env:LLM_PYTHON_VENV_PATH)) {
        llm-venv-initialize
    }

    . $env:LLM_PYTHON_VENV_PATH\Scripts\Activate.ps1

    $env:ROCM_VERSION = "6.2.4"
    $env:USE_ROCM = 1
    $env:HIP_PLATFORM = "amd"
    $env:GPU_ARCHS = "gfx1100"
    $env:HSA_OVERRIDE_GFX_VERSION = "11.0.0"

    $env:GFX_ARCH = $env:GPU_ARCHS
    $env:AMDGPU_TARGETS = $env:GPU_ARCHS
    $env:PYTORCH_ROCM_ARCH = $env:GPU_ARCHS
    $env:HIP_ARCHS = $env:GPU_ARCHS
    # HIP_PATH should be set by ROCm installer
    $env:ROCM_PATH = $env:HIP_PATH

    # llama.cpp-related variables (tweak if necessary)
    $env:LLAMA_CPP_PATH = "$env:USERPROFILE\.llama.cpp"
    $env:LLAMA_ARG_BATCH = 1024 * 8
    $env:LLAMA_ARG_UBATCH = [int]$env:LLAMA_ARG_BATCH / 8
    $env:LLAMA_ARG_N_GPU_LAYERS = 999
    $env:LLAMA_ARG_FLASH_ATTN = 1
    $env:LLAMA_ARG_MLOCK = 1

    # openwebui-related variables (tweak if necessary)
    $env:OPEN_WEBUI_ROOT_DIR = "F:/openwebui"
    $env:OPEN_WEBUI_HOST = $env:LLAMA_ARG_HOST
    $env:OPEN_WEBUI_PORT = 55569

    $env:PATH = "${env:ROCM_PATH}bin;${env:PATH}"
    $env:CMAKE_MODULE_PATH = "${env:CMAKE_MODULE_PATH};${env:ROCM_PATH}cmake"

    Import-VisualStudioVars -Architecture x64

    $env:LLM_VENV_ACTIVE = 1
    Write-Host "llama.cpp python virtual environment activated!"
}

Function llm-venv-update  {
    llm-venv-activate
    Write-Host "Updating python virtualenv for LLMs..."
    # sentencepiece must be installed manually before running that
    uv pip install --upgrade pip setuptools wheel
    uv pip install --upgrade sentencepiece==0.2.1 numpy transformers gguf protobuf torch
    Write-Host "Updated!"
}

Function llm-venv-initialize {
    Write-Host "Initializing python virtualenv for LLMs..."
    # python -m venv $env:LLM_PYTHON_VENV_PATH
    uv venv --no-project --no-config $env:LLM_PYTHON_VENV_PATH
    # llm-venv-update
    Write-Host "Virtual env initialized!"
}

Function llm-venv-remove {
    Write-Host "Removing python virtualenv for LLMs..."
    rm -Force -Recurse $env:LLM_PYTHON_VENV_PATH
    Write-Host "Deleted!"
}

# llama.cpp management functions
Function llama-cpp-clone {
    llm-venv-activate
    Write-Host "Pulling llama.cpp repository to $env:LLAMA_CPP_PATH"
    git clone [email protected]:ggerganov/llama.cpp.git $env:LLAMA_CPP_PATH --recurse-submodules
    Write-Host "Pulled!"
}

Function llama-cpp-update {
    llm-venv-activate
    Write-Host "Updating llama.cpp..."
    Push-Location $env:LLAMA_CPP_PATH
    git fetch
    git clean -xddf
    git pull
    git submodule update --recursive
    git lfs pull
    llm-venv-update
    Pop-Location
    Write-Host "llama.cpp updated!"
}

Function llama-cpp-clean {
    llm-venv-activate
    Write-Host "Performing clean build of llama.cpp..."
    Push-Location $env:LLAMA_CPP_PATH
    git clean -xddf
    Pop-Location
    Write-Host "Cleaning done!"
}

Function llama-cpp-remove {
    llm-venv-activate
    Write-Host "Removing llama.cpp repository and build files..."
    rm -Force -Recurse $env:LLAMA_CPP_PATH
    Write-Host "Removed!"
}

Function llama-cpp-build-rocm {
    llm-venv-activate
    Write-Host "Building llama.cpp for ROCm..."

    Push-Location $env:LLAMA_CPP_PATH
    cmake -S . -B build -G Ninja `
        -DCMAKE_BUILD_TYPE=Release `
        -DCMAKE_CXX_COMPILER=clang++ `
        -DCMAKE_C_COMPILER=clang `
        -DCMAKE_INSTALL_PREFIX="C:/Users/phoen/AppData/Local/llama-cpp" `
        -DLLAMA_BUILD_TESTS=OFF `
        -DLLAMA_BUILD_EXAMPLES=ON `
        -DLLAMA_BUILD_SERVER=ON `
        -DLLAMA_STANDALONE=ON `
        -DLLAMA_CURL=OFF `
        -DGGML_CCACHE=ON `
        -DGGML_NATIVE=ON `
        -DGGML_OPENMP=ON `
        -DGGML_AVX=ON `
        -DGGML_AVX2=ON `
        -DGGML_FMA=ON `
        -DGGML_HIP=ON `
        -DAMDGPU_TARGETS=gfx1100 `
        -DGGML_CUDA_FA_ALL_QUANTS=ON

    cmake --build build --config Release --parallel 24
    cmake --install build --config Release
    Pop-Location
    Write-Host "llama.cpp build completed!"
}

# build vulkan from msys

# openwebui management functions
function openwebui-pull {
    llm-venv-activate
    Push-Location $env:OPEN_WEBUI_ROOT_DIR
    docker compose pull
    Pop-Location
}

function openwebui-start {
    llm-venv-activate
    Push-Location $env:OPEN_WEBUI_ROOT_DIR
    docker compose up
    Pop-Location
}

function openwebui-start-daemon {
    llm-venv-activate
    Push-Location $env:OPEN_WEBUI_ROOT_DIR
    docker compose up -d
    Pop-Location
}

function openwebui-stop {
    llm-venv-activate
    Push-Location $env:OPEN_WEBUI_ROOT_DIR
    docker compose down
    Pop-Location
}

# llm utility functions
Function llm-quantize-model {
    Param(
        [Parameter(Mandatory=$true)]
        [string]$model_path,

        [Parameter(Mandatory=$false)]
        [string]$quantization = 'auto'
    )

    llm-venv-activate
    $output_file = Split-Path -Path $model_path -Leaf

    Write-Host "Converting ${model_path} to ${output_file}..."
    Invoke-Expression "python $env:LLAMA_CPP_PATH/convert_hf_to_gguf.py --outtype $quantization --outfile $output_file.$quantization.gguf $model_path"
}

Function llm-server {
    Param(
        [Parameter(Mandatory=$true)]
        [string]$model_path,

        [Parameter(Mandatory=$true)]
        [string]$model_name,

        [Parameter(Mandatory=$true)]
        [int]$context_size_k
    )

    llm-venv-activate
    $env:LLAMA_ARG_CTX_SIZE = $context_size_k * 1024
    llama-server `
        --model $model_path `
        --alias $model_name `
        --slots `
        --props
}

Function llm-cli {
    Param(
        [Parameter(Mandatory=$true)]
        [string]$model_path,

        [Parameter(Mandatory=$true)]
        [int]$context_size_k
    )

    llm-venv-activate
    $env:LLAMA_ARG_CTX_SIZE = $context_size_k * 1024
    llama-cli `
        --prompt $env:LLAMA_CPP_PROMPT `
        --model $model_path `
        --conversation
}