Formats the disk, downloads the models, start the llama.cpp server

startup.sh

#!/bin/bash
set -e

echo "🚀 Startup script running..."

NVME_DISK="/dev/nvme1n1"
MOUNT_POINT="/mnt/nvme"
S3_BUCKET="ep-ai-us-east-1"
MODEL_DIR="/mnt/nvme/models"
MODEL_PATH="/mnt/nvme/models/DeepSeek-R1-GGUF/DeepSeek-R1-UD-IQ1_S/DeepSeek-R1-UD-IQ1_S-00001-of-00003.gguf"
LOG_FILE="/var/log/llama_server.log"

# Check if the disk is already mounted
if mount | grep -q "$MOUNT_POINT"; then
    echo "✅ NVMe disk is already mounted. Skipping format & mount steps."
else
    echo "🔹 NVMe disk not found or not mounted. Setting it up..."

    # Format and mount only if the disk is detected but not mounted
    if lsblk | grep -q "nvme1n1"; then
        echo "🔹 Formatting NVMe disk: $NVME_DISK"
        mkfs.ext4 $NVME_DISK
        mkdir -p $MOUNT_POINT
        mount $NVME_DISK $MOUNT_POINT
        echo "$NVME_DISK $MOUNT_POINT ext4 defaults,nofail 0 2" | sudo tee -a /etc/fstab
    else
        echo "❌ NVMe disk not found. Exiting script."
        exit 1
    fi
fi

# Ensure s4cmd is installed
if ! command -v s4cmd &> /dev/null; then
    echo "🔹 Installing s4cmd..."
    sudo apt install -y python3-pip
    pip3 install s4cmd
fi

# Download models only if directory is empty
mkdir -p $MODEL_DIR
if [ -z "$(ls -A $MODEL_DIR)" ]; then
    echo "🔹 Downloading model files from S3..."
    /usr/bin/s4cmd get -r s3://$S3_BUCKET/ $MODEL_DIR/

    chown -R ubuntu:ubuntu $MODEL_DIR
    echo "✅ Model download complete!"
else
    echo "✅ Model directory is not empty, skipping download."
fi

echo "Starting llama server.."

/home/ubuntu/llama.cpp/build/bin/llama-server \
    --model ${MODEL_PATH} \
    --host 0.0.0.0 \
    --port 10000 \
    --cache-type-k q4_0 \
    --n-gpu-layers 15 \
    --threads 16 \
    --ctx-size 2048 \
    --seed 3407 \
    --log-file ${LOG_FILE} --log-prefix --log-timestamps

echo "✅ Startup script completed!"