unsloth/studio/setup.sh

#!/usr/bin/env bash
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
REPO_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"

# ── Helper: run command quietly, show output only on failure ──
_run_quiet() {
    local on_fail=$1
    local label=$2
    shift 2

    local tmplog
    tmplog=$(mktemp) || {
        printf '%s\n' "Failed to create temporary file" >&2
        [ "$on_fail" = "exit" ] && exit 1 || return 1
    }

    if "$@" >"$tmplog" 2>&1; then
        rm -f "$tmplog"
        return 0
    else
        local exit_code=$?
        printf 'Failed: %s (exit code %s):\n' "$label" "$exit_code" >&2
        cat "$tmplog" >&2
        rm -f "$tmplog"

        if [ "$on_fail" = "exit" ]; then
            exit "$exit_code"
        else
            return "$exit_code"
        fi
    fi
}

run_quiet() {
    _run_quiet exit "$@"
}

run_quiet_no_exit() {
    _run_quiet return "$@"
}

if [ "${SKIP_STUDIO_BASE:-0}" = "1" ]; then
    echo "╔══════════════════════════════════════╗"
    echo "║     Unsloth Studio Setup Script      ║"
    echo "╚══════════════════════════════════════╝"
else
    echo "╔══════════════════════════════════════╗"
    echo "║     Unsloth Studio Update Script     ║"
    echo "╚══════════════════════════════════════╝"
fi

# ── Clean up stale Unsloth compiled caches ──
rm -rf "$REPO_ROOT/unsloth_compiled_cache"
rm -rf "$SCRIPT_DIR/backend/unsloth_compiled_cache"
rm -rf "$SCRIPT_DIR/tmp/unsloth_compiled_cache"

# ── Detect Colab (like unsloth does) ──
IS_COLAB=false
keynames=$'\n'$(printenv | cut -d= -f1)
if [[ "$keynames" == *$'\nCOLAB_'* ]]; then
    IS_COLAB=true
fi

# ── Detect whether frontend needs building ──
# Skip if dist/ exists AND no tracked input is newer than dist/.
# Checks top-level config/entry files and src/, public/ recursively.
# This handles: PyPI installs (dist/ bundled), repeat runs (no changes),
# and upgrades/pulls (source newer than dist/ triggers rebuild).
_NEED_FRONTEND_BUILD=true
if [ -d "$SCRIPT_DIR/frontend/dist" ]; then
    # Check all top-level files (package.json, bun.lock, vite.config.ts, index.html, etc.)
    _changed=$(find "$SCRIPT_DIR/frontend" -maxdepth 1 -type f \
        ! -name 'bun.lock' \
        -newer "$SCRIPT_DIR/frontend/dist" -print -quit 2>/dev/null)
    # Check src/ and public/ recursively (|| true guards against set -e when dirs are missing)
    if [ -z "$_changed" ]; then
        _changed=$(find "$SCRIPT_DIR/frontend/src" "$SCRIPT_DIR/frontend/public" \
            -type f -newer "$SCRIPT_DIR/frontend/dist" -print -quit 2>/dev/null) || true
    fi
    if [ -z "$_changed" ]; then
        _NEED_FRONTEND_BUILD=false
    fi
fi
if [ "$_NEED_FRONTEND_BUILD" = false ]; then
    echo "✅ Frontend already built and up to date -- skipping Node/npm check."
else
NEED_NODE=true
if command -v node &>/dev/null && command -v npm &>/dev/null; then
    NODE_MAJOR=$(node -v | sed 's/v//' | cut -d. -f1)
    NODE_MINOR=$(node -v | sed 's/v//' | cut -d. -f2)
    NPM_MAJOR=$(npm -v | cut -d. -f1)
    # Vite 8 requires Node ^20.19.0 || >=22.12.0
    NODE_OK=false
    if [ "$NODE_MAJOR" -eq 20 ] && [ "$NODE_MINOR" -ge 19 ]; then NODE_OK=true; fi
    if [ "$NODE_MAJOR" -eq 22 ] && [ "$NODE_MINOR" -ge 12 ]; then NODE_OK=true; fi
    if [ "$NODE_MAJOR" -ge 23 ]; then NODE_OK=true; fi
    if [ "$NODE_OK" = true ] && [ "$NPM_MAJOR" -ge 11 ]; then
        echo "✅ Node $(node -v) and npm $(npm -v) already meet requirements. Skipping nvm install."
        NEED_NODE=false
    else
        if [ "$IS_COLAB" = true ] && [ "$NODE_OK" = true ]; then
            echo "✅ Node $(node -v) and npm $(npm -v) detected in Colab."
            # In Colab, just upgrade npm directly - nvm doesn't work well
            if [ "$NPM_MAJOR" -lt 11 ]; then
                echo "   Upgrading npm to latest..."
                npm install -g npm@latest > /dev/null 2>&1
            fi
            NEED_NODE=false
        else
            echo "⚠️  Node $(node -v) / npm $(npm -v) too old. Installing via nvm..."
        fi
    fi
else
    echo "⚠️  Node/npm not found. Installing via nvm..."
fi

if [ "$NEED_NODE" = true ]; then
    # ── 2. Install nvm ──
    export NODE_OPTIONS=--dns-result-order=ipv4first # or else fails on colab.
    echo "Installing nvm..."
    curl -so- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash > /dev/null 2>&1

    # Load nvm (source ~/.bashrc won't work inside a script)
    export NVM_DIR="$HOME/.nvm"
    set +u
    [ -s "$NVM_DIR/nvm.sh" ] && \. "$NVM_DIR/nvm.sh"

    # ── Fix npmrc conflict with nvm ──
    # System npm (apt, conda, etc.) may have written `prefix` or `globalconfig`
    # to ~/.npmrc, which is incompatible with nvm and causes "nvm use" to fail
    # with: "has a `globalconfig` and/or a `prefix` setting, which are
    # incompatible with nvm."
    if [ -f "$HOME/.npmrc" ]; then
        if grep -qE '^\s*(prefix|globalconfig)\s*=' "$HOME/.npmrc"; then
            echo "   Removing incompatible prefix/globalconfig from ~/.npmrc for nvm..."
            sed -i.bak '/^\s*\(prefix\|globalconfig\)\s*=/d' "$HOME/.npmrc"
        fi
    fi

    # ── 3. Install Node LTS ──
    echo "Installing Node LTS..."
    run_quiet "nvm install" nvm install --lts
    nvm use --lts > /dev/null 2>&1
    set -u
    # ── 4. Verify versions ──
    NODE_MAJOR=$(node -v | sed 's/v//' | cut -d. -f1)
    NPM_MAJOR=$(npm -v | cut -d. -f1)

    if [ "$NODE_MAJOR" -lt 20 ]; then
        echo "❌ ERROR: Node version must be >= 20 (got $(node -v))"
        exit 1
    fi
    if [ "$NPM_MAJOR" -lt 11 ]; then
        echo "⚠️  npm version is $(npm -v), expected >= 11. Updating..."
        run_quiet "npm update" npm install -g npm@latest
    fi
fi

echo "✅ Node $(node -v) | npm $(npm -v)"

# ── Install bun (optional, faster package installs) ──
# Uses npm to install bun globally — Node is already guaranteed above,
# avoids platform-specific installers, PATH issues, and admin requirements.
if ! command -v bun &>/dev/null; then
    echo "   Installing bun (faster frontend package installs)..."
    if npm install -g bun > /dev/null 2>&1 && command -v bun &>/dev/null; then
        echo "✅ bun installed ($(bun --version))"
    else
        echo "   bun install skipped (npm will be used instead)"
    fi
else
    echo "✅ bun already installed ($(bun --version))"
fi

# ── 5. Build frontend ──
cd "$SCRIPT_DIR/frontend"

# Tailwind v4's oxide scanner respects .gitignore in parent directories.
# Python venvs create a .gitignore with "*" (ignore everything), which
# prevents Tailwind from scanning .tsx source files for class names.
# Temporarily hide any such .gitignore during the build, then restore it.
_HIDDEN_GITIGNORES=()
_dir="$(pwd)"
while [ "$_dir" != "/" ]; do
    _dir="$(dirname "$_dir")"
    if [ -f "$_dir/.gitignore" ] && grep -qx '\*' "$_dir/.gitignore" 2>/dev/null; then
        mv "$_dir/.gitignore" "$_dir/.gitignore._twbuild"
        _HIDDEN_GITIGNORES+=("$_dir/.gitignore")
    fi
done

_restore_gitignores() {
    for _gi in "${_HIDDEN_GITIGNORES[@]+"${_HIDDEN_GITIGNORES[@]}"}"; do
        mv "${_gi}._twbuild" "$_gi" 2>/dev/null || true
    done
}
trap _restore_gitignores EXIT

# Use bun for install if available (faster), fall back to npm.
# Build always uses npm (Node runtime — avoids bun runtime issues on some platforms).
# NOTE: We intentionally avoid run_quiet for the bun install attempt because
# run_quiet calls exit on failure, which would kill the script before the npm
# fallback can run. Instead we capture output manually and only show it on failure.
if command -v bun &>/dev/null; then
    echo "   Using bun for package install (faster)"
    _bun_log=$(mktemp)
    if bun install >"$_bun_log" 2>&1; then
        rm -f "$_bun_log"
    else
        echo "   ⚠️  bun install failed, falling back to npm"
        echo "   bun install output:"
        sed 's/^/   | /' "$_bun_log" >&2
        rm -f "$_bun_log"
        rm -rf node_modules
        run_quiet "npm install" npm install
    fi
else
    run_quiet "npm install" npm install
fi
run_quiet "npm run build" npm run build

_restore_gitignores
trap - EXIT

# Validate CSS output -- catch truncated Tailwind builds
_MAX_CSS=$(find "$SCRIPT_DIR/frontend/dist/assets" -name '*.css' -exec wc -c {} + 2>/dev/null | sort -n | tail -1 | awk '{print $1}')
if [ -z "$_MAX_CSS" ]; then
    echo "⚠️  WARNING: No CSS files were emitted. The frontend build may have failed."
elif [ "$_MAX_CSS" -lt 100000 ]; then
    echo "⚠️  WARNING: Largest CSS file is only $((_MAX_CSS / 1024))KB (expected >100KB)."
    echo "   Tailwind may not have scanned all source files. Check for .gitignore interference."
fi

cd "$SCRIPT_DIR"
echo "✅ Frontend built to frontend/dist"

fi  # end frontend build check

# ── oxc-validator runtime (needs npm -- skip if not available) ──
if [ -d "$SCRIPT_DIR/backend/core/data_recipe/oxc-validator" ] && command -v npm &>/dev/null; then
    cd "$SCRIPT_DIR/backend/core/data_recipe/oxc-validator"
    run_quiet "npm install (oxc validator runtime)" npm install
    cd "$SCRIPT_DIR"
fi

# ── 6. Python venv + deps ──

# The venv must already exist (created by install.sh).
# This script (setup.sh / "unsloth studio update") only updates packages.
STUDIO_HOME="$HOME/.unsloth/studio"
VENV_DIR="$STUDIO_HOME/unsloth_studio"
VENV_T5_DIR="$STUDIO_HOME/.venv_t5"

# Clean up legacy in-repo venvs if they exist
[ -d "$REPO_ROOT/.venv" ] && rm -rf "$REPO_ROOT/.venv"
[ -d "$REPO_ROOT/.venv_overlay" ] && rm -rf "$REPO_ROOT/.venv_overlay"
[ -d "$REPO_ROOT/.venv_t5" ] && rm -rf "$REPO_ROOT/.venv_t5"
# Note: do NOT delete $STUDIO_HOME/.venv here — install.sh handles migration

if [ ! -x "$VENV_DIR/bin/python" ]; then
    echo "❌ ERROR: Virtual environment not found at $VENV_DIR"
    echo "   Run install.sh first to create the environment:"
    echo "   curl -fsSL https://unsloth.ai/install.sh | sh"
    exit 1
fi

source "$VENV_DIR/bin/activate"

install_python_stack() {
    python "$SCRIPT_DIR/install_python_stack.py"
}

# ── Ensure uv is available (much faster than pip) ──
USE_UV=false
if command -v uv &>/dev/null; then
    USE_UV=true
elif curl -LsSf https://astral.sh/uv/install.sh | sh > /dev/null 2>&1; then
    export PATH="$HOME/.local/bin:$PATH"
    command -v uv &>/dev/null && USE_UV=true
fi

# Helper: install a package, preferring uv with pip fallback
fast_install() {
    if [ "$USE_UV" = true ]; then
        uv pip install --python "$(command -v python)" "$@" && return 0
    fi
    python -m pip install "$@"
}

cd "$SCRIPT_DIR"

# ── Check if Python deps need updating ──
# Compare installed package version against PyPI latest.
# Skip all Python dependency work if versions match (fast update path).
_PKG_NAME="${STUDIO_PACKAGE_NAME:-unsloth}"
_SKIP_PYTHON_DEPS=false
if [ "${SKIP_STUDIO_BASE:-0}" != "1" ] && [ "${STUDIO_LOCAL_INSTALL:-0}" != "1" ]; then
    # Only check when NOT called from install.sh (which just installed the package)
    INSTALLED_VER=$("$VENV_DIR/bin/python" -c "
from importlib.metadata import version
print(version('$_PKG_NAME'))
" 2>/dev/null || echo "")

    LATEST_VER=$(curl -fsSL --max-time 5 "https://pypi.org/pypi/$_PKG_NAME/json" 2>/dev/null \
        | "$VENV_DIR/bin/python" -c "import sys,json; print(json.load(sys.stdin)['info']['version'])" 2>/dev/null \
        || echo "")

    if [ -n "$INSTALLED_VER" ] && [ -n "$LATEST_VER" ] && [ "$INSTALLED_VER" = "$LATEST_VER" ]; then
        echo "✅ $_PKG_NAME $INSTALLED_VER is up to date (matches PyPI latest)"
        _SKIP_PYTHON_DEPS=true
    elif [ -n "$INSTALLED_VER" ] && [ -n "$LATEST_VER" ]; then
        echo "⬆️  $_PKG_NAME $INSTALLED_VER → $LATEST_VER available, updating dependencies..."
    elif [ -z "$LATEST_VER" ]; then
        echo "⚠️  Could not reach PyPI, updating dependencies to be safe..."
    fi
fi

if [ "$_SKIP_PYTHON_DEPS" = false ]; then
    install_python_stack

    # ── 6b. Pre-install transformers 5.x into .venv_t5/ ──
    # Models like GLM-4.7-Flash need transformers>=5.3.0. Instead of pip-installing
    # at runtime (slow, ~10-15s), we pre-install into a separate directory.
    # The training subprocess just prepends .venv_t5/ to sys.path -- instant switch.
    echo ""
    echo "   Pre-installing transformers 5.x for newer model support..."
    mkdir -p "$VENV_T5_DIR"
    run_quiet "install transformers 5.x" fast_install --target "$VENV_T5_DIR" --no-deps "transformers==5.3.0"
    run_quiet "install huggingface_hub for t5" fast_install --target "$VENV_T5_DIR" --no-deps "huggingface_hub==1.7.1"
    run_quiet "install hf_xet for t5" fast_install --target "$VENV_T5_DIR" --no-deps "hf_xet==1.4.2"
    # tiktoken is needed by Qwen-family tokenizers. Install with deps since
    # regex/requests may be missing on Windows.
    run_quiet "install tiktoken for t5" fast_install --target "$VENV_T5_DIR" "tiktoken"
    echo "✅ Transformers 5.x pre-installed to $VENV_T5_DIR/"
else
    echo "✅ Python dependencies up to date — skipping"
fi

# ── 7. Prefer prebuilt llama.cpp bundles before any source build path ──
UNSLOTH_HOME="$HOME/.unsloth"
mkdir -p "$UNSLOTH_HOME"
LLAMA_CPP_DIR="$UNSLOTH_HOME/llama.cpp"
LLAMA_SERVER_BIN="$LLAMA_CPP_DIR/build/bin/llama-server"
_NEED_LLAMA_SOURCE_BUILD=false
_LLAMA_FORCE_COMPILE="${UNSLOTH_LLAMA_FORCE_COMPILE:-0}"
_REQUESTED_LLAMA_TAG="${UNSLOTH_LLAMA_TAG:-latest}"
_HELPER_RELEASE_REPO="${UNSLOTH_LLAMA_RELEASE_REPO:-unslothai/llama.cpp}"
_RESOLVE_LLAMA_LOG="$(mktemp)"
set +e
python "$SCRIPT_DIR/install_llama_prebuilt.py" \
    --resolve-install-tag "$_REQUESTED_LLAMA_TAG" \
    --published-repo "$_HELPER_RELEASE_REPO" >"$_RESOLVE_LLAMA_LOG" 2>&1
_RESOLVE_LLAMA_STATUS=$?
set -e
if [ "$_RESOLVE_LLAMA_STATUS" -eq 0 ]; then
    _RESOLVED_LLAMA_TAG="$(tail -n 1 "$_RESOLVE_LLAMA_LOG" | tr -d '\r')"
else
    _RESOLVED_LLAMA_TAG=""
fi
if [ -z "$_RESOLVED_LLAMA_TAG" ]; then
    echo ""
    echo "⚠️  Failed to resolve an installable prebuilt llama.cpp tag via $_HELPER_RELEASE_REPO"
    cat "$_RESOLVE_LLAMA_LOG" >&2 || true
    set +e
    _RESOLVED_LLAMA_TAG="$(python "$SCRIPT_DIR/install_llama_prebuilt.py" --resolve-llama-tag "$_REQUESTED_LLAMA_TAG" 2>/dev/null)"
    _RESOLVE_UPSTREAM_STATUS=$?
    set -e
    if [ "$_RESOLVE_UPSTREAM_STATUS" -ne 0 ] || [ -z "$_RESOLVED_LLAMA_TAG" ]; then
        if [ "$_REQUESTED_LLAMA_TAG" = "latest" ]; then
            # Try Unsloth release repo first, then fall back to ggml-org upstream
            _RESOLVED_LLAMA_TAG="$(curl -fsSL "https://api.github.com/repos/${_HELPER_RELEASE_REPO}/releases/latest" 2>/dev/null | python -c "import sys,json; print(json.load(sys.stdin)['tag_name'])" 2>/dev/null)" || _RESOLVED_LLAMA_TAG=""
            if [ -z "$_RESOLVED_LLAMA_TAG" ]; then
                _RESOLVED_LLAMA_TAG="$(curl -fsSL https://api.github.com/repos/ggml-org/llama.cpp/releases/latest 2>/dev/null | python -c "import sys,json; print(json.load(sys.stdin)['tag_name'])" 2>/dev/null)" || _RESOLVED_LLAMA_TAG=""
            fi
        fi
        if [ -z "$_RESOLVED_LLAMA_TAG" ]; then
            _RESOLVED_LLAMA_TAG="$_REQUESTED_LLAMA_TAG"
        fi
    fi
    _NEED_LLAMA_SOURCE_BUILD=true
    _SKIP_PREBUILT_INSTALL=true
fi
rm -f "$_RESOLVE_LLAMA_LOG"

echo ""
echo "Resolved llama.cpp release tag: $_RESOLVED_LLAMA_TAG"

if [ "$_LLAMA_FORCE_COMPILE" = "1" ]; then
    echo ""
    echo "⚠️  UNSLOTH_LLAMA_FORCE_COMPILE=1 -- skipping prebuilt llama.cpp install"
    _NEED_LLAMA_SOURCE_BUILD=true
else
    echo ""
    echo "Installing prebuilt llama.cpp bundle (preferred path)..."
    if [ -d "$LLAMA_CPP_DIR" ]; then
        echo "Existing llama.cpp install detected -- validating staged prebuilt update before replacement"
    fi
    if [ "${_SKIP_PREBUILT_INSTALL:-false}" = true ]; then
        echo "⚠️  Skipping prebuilt install because prebuilt tag resolution failed -- falling back to source build"
    else
        _PREBUILT_CMD=(
            python "$SCRIPT_DIR/install_llama_prebuilt.py"
            --install-dir "$LLAMA_CPP_DIR"
            --llama-tag "$_RESOLVED_LLAMA_TAG"
            --published-repo "$_HELPER_RELEASE_REPO"
        )
        if [ -n "${UNSLOTH_LLAMA_RELEASE_TAG:-}" ]; then
            _PREBUILT_CMD+=(--published-release-tag "$UNSLOTH_LLAMA_RELEASE_TAG")
        fi
        set +e
        "${_PREBUILT_CMD[@]}"
        _PREBUILT_STATUS=$?
        set -e

        if [ "$_PREBUILT_STATUS" -eq 0 ]; then
            echo "✅ Prebuilt llama.cpp installed and validated"
        else
            if [ -d "$LLAMA_CPP_DIR" ]; then
                echo "⚠️  Prebuilt update failed; existing install was restored or cleaned before source build fallback"
            fi
            echo "⚠️  Prebuilt llama.cpp path unavailable or failed validation -- falling back to source build"
            _NEED_LLAMA_SOURCE_BUILD=true
        fi
    fi
fi

# ── 8. WSL: pre-install GGUF build dependencies for fallback source builds ──
# On WSL, sudo requires a password and can't be entered during GGUF export
# (runs in a non-interactive subprocess). Install build deps here instead.
if [ "$_NEED_LLAMA_SOURCE_BUILD" = true ] && grep -qi microsoft /proc/version 2>/dev/null; then
    echo ""
    echo "⚠️  WSL detected -- installing build dependencies for GGUF export..."
    _GGUF_DEPS="pciutils build-essential cmake curl git libcurl4-openssl-dev"

    # Try without sudo first (works when already root)
    apt-get update -y >/dev/null 2>&1 || true
    apt-get install -y $_GGUF_DEPS >/dev/null 2>&1 || true

    # Check which packages are still missing
    _STILL_MISSING=""
    for _pkg in $_GGUF_DEPS; do
        case "$_pkg" in
            build-essential) command -v gcc >/dev/null 2>&1 || _STILL_MISSING="$_STILL_MISSING $_pkg" ;;
            pciutils) command -v lspci >/dev/null 2>&1 || _STILL_MISSING="$_STILL_MISSING $_pkg" ;;
            libcurl4-openssl-dev) dpkg -s "$_pkg" >/dev/null 2>&1 || _STILL_MISSING="$_STILL_MISSING $_pkg" ;;
            *) command -v "$_pkg" >/dev/null 2>&1 || _STILL_MISSING="$_STILL_MISSING $_pkg" ;;
        esac
    done
    _STILL_MISSING=$(echo "$_STILL_MISSING" | sed 's/^ *//')

    if [ -z "$_STILL_MISSING" ]; then
        echo "✅ GGUF build dependencies installed"
    elif command -v sudo >/dev/null 2>&1; then
        echo ""
        echo "   !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"
        echo "   WARNING: We require sudo elevated permissions to install:"
        echo "   $_STILL_MISSING"
        echo "   If you accept, we'll run sudo now, and it'll prompt your password."
        echo "   !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"
        echo ""
        printf "   Accept? [Y/n] "
        if [ -r /dev/tty ]; then
            read -r REPLY </dev/tty || REPLY="y"
        else
            REPLY="y"
        fi
        case "$REPLY" in
            [nN]*)
                echo ""
                echo "   Please install these packages first, then re-run Unsloth Studio setup:"
                echo "   sudo apt-get update -y && sudo apt-get install -y $_STILL_MISSING"
                _SKIP_GGUF_BUILD=true
                ;;
            *)
                sudo apt-get update -y
                sudo apt-get install -y $_STILL_MISSING
                echo "✅ GGUF build dependencies installed"
                ;;
        esac
    else
        echo "   sudo is not available on this system."
        echo "   Please install as root, then re-run setup:"
        echo "   apt-get install -y $_STILL_MISSING"
        _SKIP_GGUF_BUILD=true
    fi
fi

# ── 9. Build llama.cpp binaries for GGUF inference + export when prebuilt install fails ──
# Builds at ~/.unsloth/llama.cpp — a single shared location under the user's
# home directory. This is used by both the inference server and the GGUF
# export pipeline (unsloth-zoo).
#   - llama-server: for GGUF model inference
#   - llama-quantize: for GGUF export quantization (symlinked to root for check_llama_cpp())
if [ "$_NEED_LLAMA_SOURCE_BUILD" = false ]; then
    :
elif [ "${_SKIP_GGUF_BUILD:-}" = true ]; then
    echo ""
    echo "Skipping llama-server build (missing dependencies)"
    echo "   Install the missing packages and re-run setup to enable GGUF inference."
else
{
    # Check prerequisites
    if ! command -v cmake &>/dev/null; then
        echo ""
        echo "⚠️  cmake not found — skipping llama-server build (GGUF inference won't be available)"
        echo "   Install cmake and re-run setup.sh to enable GGUF inference."
    elif ! command -v git &>/dev/null; then
        echo ""
        echo "⚠️  git not found — skipping llama-server build (GGUF inference won't be available)"
    else
        echo ""
        echo "Building llama-server for GGUF inference..."

        BUILD_OK=true
        _CLONE_BRANCH_ARGS=()
        if [ "$_RESOLVED_LLAMA_TAG" != "latest" ] && [ -n "$_RESOLVED_LLAMA_TAG" ]; then
            _CLONE_BRANCH_ARGS=(--branch "$_RESOLVED_LLAMA_TAG")
        fi
        _BUILD_TMP="${LLAMA_CPP_DIR}.build.$$"
        rm -rf "$_BUILD_TMP"
        run_quiet_no_exit "clone llama.cpp" git clone --depth 1 "${_CLONE_BRANCH_ARGS[@]}" https://github.com/ggml-org/llama.cpp.git "$_BUILD_TMP" || BUILD_OK=false

        if [ "$BUILD_OK" = true ]; then
            # Skip tests/examples we don't need (faster build)
            CMAKE_ARGS="-DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_NATIVE=ON"

            # Use ccache if available (dramatically faster rebuilds)
            if command -v ccache &>/dev/null; then
                CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache -DCMAKE_CUDA_COMPILER_LAUNCHER=ccache"
                echo "   Using ccache for faster compilation"
            fi

            # Detect GPU backend: CUDA (NVIDIA) or ROCm (AMD)
            GPU_BACKEND=""

            # Check for CUDA: check nvcc on PATH, then common install locations
            NVCC_PATH=""
            if command -v nvcc &>/dev/null; then
                NVCC_PATH="$(command -v nvcc)"
                GPU_BACKEND="cuda"
            elif [ -x /usr/local/cuda/bin/nvcc ]; then
                NVCC_PATH="/usr/local/cuda/bin/nvcc"
                export PATH="/usr/local/cuda/bin:$PATH"
                GPU_BACKEND="cuda"
            elif ls /usr/local/cuda-*/bin/nvcc &>/dev/null 2>&1; then
                # Pick the newest cuda-XX.X directory
                NVCC_PATH="$(ls -d /usr/local/cuda-*/bin/nvcc 2>/dev/null | sort -V | tail -1)"
                export PATH="$(dirname "$NVCC_PATH"):$PATH"
                GPU_BACKEND="cuda"
            fi

            # Check for ROCm (AMD) only if CUDA was not already selected
            ROCM_HIPCC=""
            if [ -z "$GPU_BACKEND" ]; then
                if command -v hipcc &>/dev/null; then
                    ROCM_HIPCC="$(command -v hipcc)"
                    GPU_BACKEND="rocm"
                elif [ -x /opt/rocm/bin/hipcc ]; then
                    ROCM_HIPCC="/opt/rocm/bin/hipcc"
                    export PATH="/opt/rocm/bin:$PATH"
                    GPU_BACKEND="rocm"
                elif ls /opt/rocm-*/bin/hipcc &>/dev/null 2>&1; then
                    ROCM_HIPCC="$(ls -d /opt/rocm-*/bin/hipcc 2>/dev/null | sort -V | tail -1)"
                    export PATH="$(dirname "$ROCM_HIPCC"):$PATH"
                    GPU_BACKEND="rocm"
                fi
            fi

            if [ -n "$NVCC_PATH" ]; then
                echo "   Building with CUDA support (nvcc: $NVCC_PATH)..."
                CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON"

                # Detect GPU compute capability and limit CUDA architectures
                # Without this, cmake builds for ALL default archs (very slow)
                CUDA_ARCHS=""
                if command -v nvidia-smi &>/dev/null; then
                    # Read all GPUs, deduplicate (handles mixed-GPU hosts)
                    _raw_caps=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)
                    while IFS= read -r _cap; do
                        _cap=$(echo "$_cap" | tr -d '[:space:]')
                        if [[ "$_cap" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
                            _arch="${BASH_REMATCH[1]}${BASH_REMATCH[2]}"
                            # Append if not already present
                            case ";$CUDA_ARCHS;" in
                                *";$_arch;"*) ;;
                                *) CUDA_ARCHS="${CUDA_ARCHS:+$CUDA_ARCHS;}$_arch" ;;
                            esac
                        fi
                    done <<< "$_raw_caps"
                fi

                if [ -n "$CUDA_ARCHS" ]; then
                    echo "   GPU compute capabilities: ${CUDA_ARCHS//;/, } -- limiting build to detected archs"
                    CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHS}"
                else
                    echo "   Could not detect GPU arch -- building for all default CUDA architectures (slower)"
                fi

                # Multi-threaded nvcc compilation (uses all CPU cores per .cu file)
                CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_FLAGS=--threads=0"
            elif [ "$GPU_BACKEND" = "rocm" ]; then
                # Resolve hipcc symlinks to find the real ROCm root
                _HIPCC_REAL="$(readlink -f "$ROCM_HIPCC" 2>/dev/null || printf '%s' "$ROCM_HIPCC")"
                ROCM_ROOT=""
                if command -v hipconfig &>/dev/null; then
                    ROCM_ROOT="$(hipconfig -R 2>/dev/null || true)"
                fi
                if [ -z "$ROCM_ROOT" ]; then
                    ROCM_ROOT="$(cd "$(dirname "$_HIPCC_REAL")/.." 2>/dev/null && pwd)"
                fi

                echo "   Building with ROCm support (AMD GPU, hipcc: $_HIPCC_REAL)..."
                CMAKE_ARGS="$CMAKE_ARGS -DGGML_HIP=ON"
                export ROCM_PATH="$ROCM_ROOT"
                export HIP_PATH="$ROCM_ROOT"

                # Use upstream-recommended HIP compiler (not legacy hipcc-as-CXX)
                if command -v hipconfig &>/dev/null; then
                    _HIP_CLANG_DIR="$(hipconfig -l 2>/dev/null || true)"
                    [ -n "$_HIP_CLANG_DIR" ] && export HIPCXX="$_HIP_CLANG_DIR/clang"
                fi

                # Detect AMD GPU architecture (gfx target)
                GPU_TARGETS=""
                if command -v rocminfo &>/dev/null; then
                    _gfx_list=$(rocminfo 2>/dev/null | grep -oE 'gfx[0-9]{2,4}[a-z]?' | sort -u || true)
                    _valid_gfx=""
                    for _gfx in $_gfx_list; do
                        if [[ "$_gfx" =~ ^gfx[0-9]{2,4}[a-z]?$ ]]; then
                            _valid_gfx="${_valid_gfx}${_valid_gfx:+;}$_gfx"
                        fi
                    done
                    [ -n "$_valid_gfx" ] && GPU_TARGETS="$_valid_gfx"
                fi

                if [ -n "$GPU_TARGETS" ]; then
                    echo "   AMD GPU architectures: ${GPU_TARGETS//;/, } -- limiting build to detected targets"
                    CMAKE_ARGS="$CMAKE_ARGS -DGPU_TARGETS=${GPU_TARGETS}"
                else
                    echo "   Could not detect AMD GPU arch -- building for default targets (cmake will auto-detect)"
                fi
            elif [ -d /usr/local/cuda ] || nvidia-smi &>/dev/null; then
                echo "   CUDA driver detected but nvcc not found — building CPU-only"
                echo "   To enable GPU: install cuda-toolkit or add nvcc to PATH"
            elif [ -d /opt/rocm ] || command -v rocm-smi &>/dev/null; then
                echo "   ROCm driver detected but hipcc not found — building CPU-only"
                echo "   To enable GPU: install rocm-dev or add hipcc to PATH"
            else
                echo "   Building CPU-only (no CUDA detected)..."
            fi

            NCPU=$(nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 4)

            # Use Ninja if available (faster parallel builds than Make)
            CMAKE_GENERATOR_ARGS=""
            if command -v ninja &>/dev/null; then
                CMAKE_GENERATOR_ARGS="-G Ninja"
            fi

            run_quiet_no_exit "cmake llama.cpp" cmake $CMAKE_GENERATOR_ARGS -S "$_BUILD_TMP" -B "$_BUILD_TMP/build" $CMAKE_ARGS || BUILD_OK=false
        fi

        if [ "$BUILD_OK" = true ]; then
            run_quiet_no_exit "build llama-server" cmake --build "$_BUILD_TMP/build" --config Release --target llama-server -j"$NCPU" || BUILD_OK=false
        fi

        # Also build llama-quantize (needed by unsloth-zoo's GGUF export pipeline)
        if [ "$BUILD_OK" = true ]; then
            run_quiet_no_exit "build llama-quantize" cmake --build "$_BUILD_TMP/build" --config Release --target llama-quantize -j"$NCPU" || true
        fi

        # Swap only after build succeeds -- preserves existing install on failure
        if [ "$BUILD_OK" = true ]; then
            rm -rf "$LLAMA_CPP_DIR"
            mv "$_BUILD_TMP" "$LLAMA_CPP_DIR"
            # Symlink to llama.cpp root -- check_llama_cpp() looks for the binary there
            QUANTIZE_BIN="$LLAMA_CPP_DIR/build/bin/llama-quantize"
            if [ -f "$QUANTIZE_BIN" ]; then
                ln -sf build/bin/llama-quantize "$LLAMA_CPP_DIR/llama-quantize"
            fi
        else
            rm -rf "$_BUILD_TMP"
        fi

        if [ "$BUILD_OK" = true ]; then
            if [ -f "$LLAMA_SERVER_BIN" ]; then
                echo "✅ llama-server built at $LLAMA_SERVER_BIN"
            else
                echo "⚠️  llama-server binary not found after build — GGUF inference won't be available"
            fi
            if [ -f "$LLAMA_CPP_DIR/llama-quantize" ]; then
                echo "✅ llama-quantize available for GGUF export"
            fi
        else
            echo "⚠️  llama-server build failed — GGUF inference won't be available, but everything else works"
        fi
    fi
}
fi  # end _SKIP_GGUF_BUILD check

echo ""
if [ "${SKIP_STUDIO_BASE:-0}" = "1" ]; then
    _DONE_LINE="║          Setup Complete!             ║"
else
    _DONE_LINE="║          Update Complete!            ║"
fi

if [ "$IS_COLAB" = true ]; then
    echo "╔══════════════════════════════════════╗"
    echo "$_DONE_LINE"
    echo "╠══════════════════════════════════════╣"
    echo "║ Unsloth Studio is ready to start     ║"
    echo "║ in your Colab notebook!              ║"
    echo "║                                      ║"
    echo "║ from colab import start              ║"
    echo "║ start()                              ║"
    echo "╚══════════════════════════════════════╝"
else
    echo "╔══════════════════════════════════════╗"
    echo "$_DONE_LINE"
    echo "╠══════════════════════════════════════╣"
    echo "║ Launch with:                         ║"
    echo "║                                      ║"
    echo "║ unsloth studio -H 0.0.0.0 -p 8888    ║"
    echo "╚══════════════════════════════════════╝"
fi