start-backend.sh

#!/usr/bin/env bash
# ============================================================================
# LAION-Tunes Backend Startup Script
# ============================================================================
# Starts the FastAPI search server with EmbeddingGemma 300M loaded directly
# in Python (CPU), plus a cloudflared tunnel for public HTTPS access.
#
# Usage:
#   ./start-backend.sh                  # default: port 7860
#   ./start-backend.sh --port 7861      # custom FastAPI port
#   ./start-backend.sh --skip-download  # skip LFS data download
#
# Prerequisites:
#   - Python 3.10+ with pip
#   - cloudflared binary at ../cloudflared
#   - ~20 GB disk for search indices
#   - ~20 GB RAM for loaded indices + embedder
# ============================================================================

set -euo pipefail

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

# ── Configuration ────────────────────────────────────────────────────────────
FASTAPI_PORT="${FASTAPI_PORT:-7860}"
VENV_DIR="$SCRIPT_DIR/.venv"
CLOUDFLARED="$SCRIPT_DIR/../cloudflared"
LOG_DIR="$SCRIPT_DIR/logs"
SKIP_DOWNLOAD=false
NO_WHISPER=false  # Whisper encoder works on CPU too

# Parse args
while [[ $# -gt 0 ]]; do
    case "$1" in
        --port) FASTAPI_PORT="$2"; shift 2 ;;
        --skip-download) SKIP_DOWNLOAD=true; shift ;;
        --with-whisper) NO_WHISPER=false; shift ;;
        *) echo "Unknown arg: $1"; exit 1 ;;
    esac
done

mkdir -p "$LOG_DIR"

echo "============================================================"
echo " LAION-Tunes Backend Startup"
echo "============================================================"
echo " FastAPI port:  $FASTAPI_PORT"
echo " Venv:          $VENV_DIR"
echo " Skip download: $SKIP_DOWNLOAD"
echo " No whisper:    $NO_WHISPER"
echo "============================================================"

# ── Step 1: Download LFS data (search indices) if needed ─────────────────────
if [ "$SKIP_DOWNLOAD" = false ]; then
    echo ""
    echo "[1/4] Checking search index data..."
    mkdir -p "$SCRIPT_DIR/search_index"

    # Check if metadata.db exists and is a real file (not LFS pointer)
    if [ -f "$SCRIPT_DIR/search_index/metadata.db" ] && \
       [ "$(stat -c%s "$SCRIPT_DIR/search_index/metadata.db" 2>/dev/null || echo 0)" -gt 1000 ]; then
        echo "  Search indices already downloaded."
    else
        echo "  Downloading search indices from HuggingFace (this may take a while ~16 GB)..."

        if [ -d "$SCRIPT_DIR/repo-source/.git" ]; then
            echo "  Using git-lfs pull in cloned repo..."
            cd "$SCRIPT_DIR/repo-source"
            git lfs pull --include="search_index/*" 2>&1 | tail -5
            # Copy real files over
            cp -f search_index/* "$SCRIPT_DIR/search_index/" 2>/dev/null || true
            cd "$SCRIPT_DIR"
        elif command -v huggingface-cli &>/dev/null; then
            echo "  Using huggingface-cli..."
            huggingface-cli download laion/laion-tunes \
                --repo-type dataset \
                --include "search_index/*" \
                --local-dir "$SCRIPT_DIR" \
                2>&1 | tail -5
        else
            echo "  ERROR: No method to download LFS data."
            echo "  Either clone the repo first:"
            echo "    GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/laion/laion-tunes repo-source"
            echo "  Or install huggingface-cli:"
            echo "    pip install huggingface-hub"
            exit 1
        fi
        echo "  Download complete."
    fi
else
    echo ""
    echo "[1/4] Skipping data download (--skip-download)"
fi

# ── Step 2: Set up Python virtual environment ────────────────────────────────
echo ""
echo "[2/4] Setting up Python environment..."

if [ ! -d "$VENV_DIR" ]; then
    echo "  Creating virtual environment..."
    python3 -m venv "$VENV_DIR"
fi

source "$VENV_DIR/bin/activate"

# Install dependencies
echo "  Installing Python dependencies..."
pip install --quiet --upgrade pip

# Set HF token for gated model access (google/embeddinggemma-300m is gated)
if [ -z "${HF_TOKEN:-}" ]; then
    if [ -f "$HOME/.cache/huggingface/token" ]; then
        export HF_TOKEN="$(cat "$HOME/.cache/huggingface/token")"
        echo "  HF token loaded from ~/.cache/huggingface/token"
    elif [ -f "$HOME/.huggingface/token" ]; then
        export HF_TOKEN="$(cat "$HOME/.huggingface/token")"
        echo "  HF token loaded from ~/.huggingface/token"
    else
        echo "  WARNING: No HF_TOKEN set. google/embeddinggemma-300m is a gated model."
        echo "  Set HF_TOKEN env var or run: huggingface-cli login"
    fi
fi

# Install CPU-only PyTorch first (smaller, no CUDA)
pip install --quiet torch --index-url https://download.pytorch.org/whl/cpu

# Install everything else
pip install --quiet \
    fastapi \
    'uvicorn[standard]' \
    faiss-cpu \
    numpy \
    pandas \
    scipy \
    tqdm \
    requests \
    pydantic \
    python-multipart \
    sentence-transformers \
    transformers \
    optimum[onnxruntime] \
    onnxruntime \
    librosa \
    soundfile

echo "  Python environment ready."

# ── Step 3: Start FastAPI server (with direct Python EmbeddingGemma) ─────────
echo ""
echo "[3/4] Starting FastAPI search server on port $FASTAPI_PORT..."
echo "  EmbeddingGemma 300M will be loaded directly in Python (CPU)."
echo "  First query may take ~500ms for embedding; subsequent ones ~430ms."

WHISPER_FLAG=""
if [ "$NO_WHISPER" = true ]; then
    WHISPER_FLAG="--no-whisper"
fi

# Kill any existing server on this port
fuser -k "${FASTAPI_PORT}/tcp" 2>/dev/null || true
sleep 1

# Run WITHOUT --tei-url so server.py loads EmbeddingGemma via SentenceTransformer
# HF_TOKEN is inherited from the environment for gated model access
nohup env HF_TOKEN="${HF_TOKEN:-}" "$VENV_DIR/bin/python" server.py \
    --port "$FASTAPI_PORT" \
    --host 0.0.0.0 \
    --gpu 0 \
    $WHISPER_FLAG \
    > "$LOG_DIR/fastapi.log" 2>&1 &

FASTAPI_PID=$!
echo "  FastAPI PID: $FASTAPI_PID"
echo "$FASTAPI_PID" > "$LOG_DIR/fastapi.pid"

# Wait for FastAPI to be ready
echo "  Waiting for FastAPI to start (loading indices ~20 GB RAM + downloading EmbeddingGemma)..."
FASTAPI_READY=false
for i in $(seq 1 600); do
    if curl -sf "http://localhost:${FASTAPI_PORT}/api/stats" >/dev/null 2>&1; then
        FASTAPI_READY=true
        echo "  FastAPI ready after ${i}s"
        break
    fi
    # Check if process is still alive
    if ! kill -0 "$FASTAPI_PID" 2>/dev/null; then
        echo "  ERROR: FastAPI process died. Check logs:"
        tail -50 "$LOG_DIR/fastapi.log"
        exit 1
    fi
    # Show progress every 30s
    if (( i % 30 == 0 )); then
        echo "  Still loading... (${i}s elapsed)"
        tail -1 "$LOG_DIR/fastapi.log" 2>/dev/null | sed 's/^/    /'
    fi
    sleep 1
done

if [ "$FASTAPI_READY" = false ]; then
    echo "  WARNING: FastAPI not ready after 600s. Check: tail -f $LOG_DIR/fastapi.log"
fi

# ── Step 4: Start cloudflared tunnel for public HTTPS ────────────────────────
echo ""
echo "[4/4] Starting cloudflared tunnel for public HTTPS..."

# Kill existing tunnel if running
if [ -f "$LOG_DIR/cloudflared-backend.pid" ]; then
    kill "$(cat "$LOG_DIR/cloudflared-backend.pid")" 2>/dev/null || true
fi

nohup "$CLOUDFLARED" tunnel --url "http://localhost:${FASTAPI_PORT}" \
    > "$LOG_DIR/cloudflared-backend.log" 2>&1 &

CF_PID=$!
echo "$CF_PID" > "$LOG_DIR/cloudflared-backend.pid"

# Wait for tunnel URL
sleep 5
TUNNEL_URL=""
for i in $(seq 1 30); do
    TUNNEL_URL=$(grep -oP 'https://[a-z0-9-]+\.trycloudflare\.com' "$LOG_DIR/cloudflared-backend.log" 2>/dev/null | head -1)
    if [ -n "$TUNNEL_URL" ]; then
        break
    fi
    sleep 1
done

echo ""
echo "============================================================"
echo " LAION-Tunes Backend is running!"
echo "============================================================"
echo ""
echo " Local:    http://localhost:${FASTAPI_PORT}"
if [ -n "$TUNNEL_URL" ]; then
    echo " Public:   $TUNNEL_URL"
else
    echo " Public:   (tunnel URL pending, check: grep trycloudflare $LOG_DIR/cloudflared-backend.log)"
fi
echo ""
echo " Logs:"
echo "   FastAPI:      tail -f $LOG_DIR/fastapi.log"
echo "   Cloudflared:  tail -f $LOG_DIR/cloudflared-backend.log"
echo ""
echo " Stop everything:"
echo "   kill $FASTAPI_PID $CF_PID"
echo "   # or: ./stop-all.sh"
echo "============================================================"