scripts/deploy_gemma.sh

#!/usr/bin/env bash
# deploy_gemma.sh — Build and launch llama-server for Kansas cognitive router
#
# Target: RX 6700 XT (gfx1030) on ROCm 7.2.1
# Model: Gemma 4 E2B (Q4_K_M GGUF)
# Pin:   llama.cpp latest master (Gemma 4n compatible, ROCm 7.2.1 fixes)
#
# Usage:
#   bash scripts/deploy_gemma.sh          # Build + launch (blocking)
#   bash scripts/deploy_gemma.sh --build  # Build only
#   bash scripts/deploy_gemma.sh --start  # Launch only (requires prior build)
#   bash scripts/deploy_gemma.sh --stop   # Stop running server

set -euo pipefail
SERVER_PID=""

# ── Paths ──────────────────────────────────────────────────────────────────
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
VENDOR_DIR="$ROOT_DIR/vendor/llama.cpp"
BUILD_DIR="$VENDOR_DIR/build"
MODELS_LINK="$ROOT_DIR/models"
PID_FILE="$ROOT_DIR/runtime/.llama-server.pid"
LOG_FILE="$ROOT_DIR/runtime/logs/llama-server.log"

# GGUF model — unsloth/gemma-4-E2B-it-GGUF (Q4_K_M)
GGUF_REPO="unsloth/gemma-4-E2B-it-GGUF"
GGUF_FILENAME="gemma-4-E2B-it-Q4_K_M.gguf"
GGUF_MODEL_DIR="$ROOT_DIR/models"
GGUF_MODEL_PATH="$GGUF_MODEL_DIR/$GGUF_FILENAME"

# llama.cpp — track latest master for Gemma 4n compatibility and ROCm fixes
LLAMA_REPO="https://github.com/ggerganov/llama.cpp.git"
LLAMA_TAG=""  # empty = latest master

# Server config
LLAMA_HOST="127.0.0.1"
LLAMA_PORT=8080
LLAMA_CTX=16384
LLAMA_GPU_LAYERS=35
LLAMA_CACHE_TYPE="q8_0"

# ROCm override (mandatory for RX 6700 XT)
export HSA_OVERRIDE_GFX_VERSION="10.3.0"

# ── Helpers ────────────────────────────────────────────────────────────────
log() { echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*"; }
die() { log "ERROR: $*" >&2; exit 1; }

# ── Stop ───────────────────────────────────────────────────────────────────
stop_server() {
    # 1. Try global variable first (fastest)
    if [[ -n "${SERVER_PID:-}" ]]; then
        if kill -0 "$SERVER_PID" 2>/dev/null; then
            log "Stopping llama-server (PID $SERVER_PID) via global variable..."
            kill -TERM "$SERVER_PID" 2>/dev/null || true
        fi
    fi

    # 2. Fallback to PID file logic (handles orphaned processes)
    if [[ -f "$PID_FILE" ]]; then
        local pid
        pid=$(cat "$PID_FILE")
        if kill -0 "$pid" 2>/dev/null; then
            log "Stopping llama-server (PID $pid) via PID file..."
            kill -TERM "$pid" 2>/dev/null || true
            # Wait up to 10s for graceful shutdown
            for i in $(seq 1 10); do
                if ! kill -0 "$pid" 2>/dev/null; then
                    log "llama-server stopped."
                    rm -f "$PID_FILE"
                    SERVER_PID=""
                    return 0
                fi
                sleep 1
            done
            log "Force killing llama-server (PID $pid)..."
            kill -9 "$pid" 2>/dev/null || true
            rm -f "$PID_FILE"
        else
            log "Stale PID file (PID $pid not running). Cleaning up."
            rm -f "$PID_FILE"
        fi
    else
        log "No PID file found. Checking for running llama-server on port $LLAMA_PORT..."
        local existing_pid
        existing_pid=$(lsof -ti:"$LLAMA_PORT" 2>/dev/null || true)
        if [[ -n "$existing_pid" ]]; then
            log "Killing llama-server (PID $existing_pid) on port $LLAMA_PORT..."
            kill -TERM "$existing_pid" 2>/dev/null || true
            sleep 2
            kill -9 "$existing_pid" 2>/dev/null || true
        else
            log "No running llama-server found."
        fi
    fi
    SERVER_PID=""
}


# ── Download ───────────────────────────────────────────────────────────────
download_model() {
    log "=== Phase 0: Downloading $GGUF_FILENAME ==="

    mkdir -p "$GGUF_MODEL_DIR"

    if [[ -f "$GGUF_MODEL_PATH" ]]; then
        log "Model already exists at $GGUF_MODEL_PATH. Skipping download."
        return 0
    fi

    if ! command -v huggingface-cli &> /dev/null; then
        log "Installing huggingface_hub CLI..."
        pip install -q huggingface_hub
    fi

    if [[ -z "${HF_TOKEN:-}" ]]; then
        # Try loading from .env
        if [[ -f "$ROOT_DIR/.env" ]]; then
            export HF_TOKEN=$(grep '^HF_TOKEN=' "$ROOT_DIR/.env" | cut -d'=' -f2-)
        fi
    fi

    if [[ -z "${HF_TOKEN:-}" ]]; then
        die "HF_TOKEN not set. Set it in .env or export HF_TOKEN=..."
    fi

    log "Downloading $GGUF_FILENAME from $GGUF_REPO..."
    huggingface-cli download "$GGUF_REPO" "$GGUF_FILENAME" \
        --local-dir "$GGUF_MODEL_DIR" \
        --token "$HF_TOKEN"

    if [[ ! -f "$GGUF_MODEL_PATH" ]]; then
        die "Download failed: $GGUF_MODEL_PATH not found"
    fi

    local size
    size=$(du -h "$GGUF_MODEL_PATH" | cut -f1)
    log "Download complete: $GGUF_FILENAME ($size)"
}

# ── Build ──────────────────────────────────────────────────────────────────
build_server() {
    if [[ -z "$LLAMA_TAG" ]]; then
        log "=== Phase 1: Building llama.cpp (latest master) ==="
    else
        log "=== Phase 1: Building llama.cpp (tag $LLAMA_TAG) ==="
    fi

    # Clone if not present
    if [[ -d "$VENDOR_DIR/.git" ]]; then
        log "llama.cpp already cloned. Skipping."
    else
        if [[ -z "$LLAMA_TAG" ]]; then
            log "Cloning llama.cpp (latest master)..."
            git clone --depth 1 "$LLAMA_REPO" "$VENDOR_DIR"
        else
            log "Cloning llama.cpp at tag $LLAMA_TAG..."
            git clone --depth 1 --branch "$LLAMA_TAG" "$LLAMA_REPO" "$VENDOR_DIR"
        fi
    fi

    # CMake configuration
    log "Configuring CMake with ROCm (gfx1030)..."
    cmake -B "$BUILD_DIR" \
        -DGGML_HIP=ON \
        -DAMDGPU_TARGETS=gfx1030 \
        -DCMAKE_BUILD_TYPE=Release \
        -DCMAKE_C_COMPILER=/opt/rocm/llvm/bin/clang \
        -DCMAKE_CXX_COMPILER=/opt/rocm/llvm/bin/clang++ \
        -S "$VENDOR_DIR"

    # Build llama-server binary only
    log "Compiling llama-server (this may take 5-10 minutes)..."
    cmake --build "$BUILD_DIR" -j"$(nproc)" --target llama-server

    # Verify binary
    if [[ ! -x "$BUILD_DIR/bin/llama-server" ]]; then
        die "Build failed: llama-server binary not found at $BUILD_DIR/bin/llama-server"
    fi

    local version
    version=$("$BUILD_DIR/bin/llama-server" --version 2>&1 || echo "unknown")
    log "Build complete: $version"
}

# ── Start ──────────────────────────────────────────────────────────────────
start_server() {
    log "=== Phase 2: Launching llama-server ==="

    # Check binary exists
    local server_bin="$BUILD_DIR/bin/llama-server"
    [[ -x "$server_bin" ]] || die "llama-server binary not found. Run: $0 --build"

    # Verify model file
    if [[ ! -f "$GGUF_MODEL_PATH" ]]; then
        die "GGUF model not found at: $GGUF_MODEL_PATH. Run: $0 --download"
    fi

    # Ensure log directory exists
    mkdir -p "$(dirname "$LOG_FILE")"

    # Stop any existing server
    stop_server

    # Launch llama-server in background
    log "Starting llama-server on $LLAMA_HOST:$LLAMA_PORT..."
    log "  Model:  $GGUF_FILENAME"
    log "  Layers: $LLAMA_GPU_LAYERS GPU / ctx=$LLAMA_CTX"
    log "  Cache:  $LLAMA_CACHE_TYPE (K+V)"

    HSA_OVERRIDE_GFX_VERSION=10.3.0 \
        "$server_bin" \
        --model "$GGUF_MODEL_PATH" \
        --ctx-size "$LLAMA_CTX" \
        --n-gpu-layers "$LLAMA_GPU_LAYERS" \
        --cache-type-k "$LLAMA_CACHE_TYPE" \
        --cache-type-v "$LLAMA_CACHE_TYPE" \
        --port "$LLAMA_PORT" \
        --host "$LLAMA_HOST" \
        > "$LOG_FILE" 2>&1 &

    SERVER_PID=$!
    echo "$SERVER_PID" > "$PID_FILE"
    log "llama-server started with PID $SERVER_PID"

    # Health check loop
    log "Waiting for server to become healthy..."
    local max_attempts=120
    local attempt=0
    while (( attempt < max_attempts )); do
        attempt=$((attempt + 1))
        if curl -sf "http://$LLAMA_HOST:$LLAMA_PORT/health" > /dev/null 2>&1; then
            log "✓ Server is healthy! (attempt $attempt)"
            log "  Health: $(curl -s "http://$LLAMA_HOST:$LLAMA_PORT/health")"
            log "  Logs:   tail -f $LOG_FILE"
            log "  Stop:   bash $0 --stop"
            return 0
        fi
        # Check if process is still alive
        if ! kill -0 "$SERVER_PID" 2>/dev/null; then
            log "ERROR: llama-server (PID $SERVER_PID) exited unexpectedly."
            log "Last 20 lines of log:"
            tail -20 "$LOG_FILE" 2>/dev/null || true
            rm -f "$PID_FILE"
            exit 1
        fi
        sleep 1
    done

    die "Server did not become healthy within ${max_attempts}s. Check $LOG_FILE"
}

# ── Main ───────────────────────────────────────────────────────────────────
trap 'log "Received interrupt signal. Shutting down..."; stop_server; exit 0' INT TERM

case "${1:-}" in
    --download)
        download_model
        ;;
    --build)
        download_model
        build_server
        ;;
    --start)
        start_server
        ;;
    --stop)
        stop_server
        ;;
    --status)
        if [[ -f "$PID_FILE" ]]; then
            pid=$(cat "$PID_FILE")
            if kill -0 "$pid" 2>/dev/null; then
                log "llama-server running (PID $pid)"
                curl -s "http://$LLAMA_HOST:$LLAMA_PORT/health" 2>/dev/null || echo "Health check failed"
            else
                log "Stale PID file (PID $pid not running)"
            fi
        else
            log "No PID file found"
        fi
        ;;
    "")
        # Default: download + build + launch
        download_model
        build_server
        start_server
        # Block until interrupted
        log "Server running. Press Ctrl+C to stop."
        wait "$SERVER_PID" 2>/dev/null || true
        ;;
    *)
        echo "Usage: $0 [--download|--build|--start|--stop|--status]"
        exit 1
        ;;
esac