start-dev.sh

#!/bin/bash

# Note: This script is designed for Linux/WSL2 environments. Its purpose is to pre-load 2 local models so they are 'warm' and ready to use without the user having to wait for models to load every time a model runs.
# Windows users should use WSL2 or adapt for PowerShell

# Make start-dev.sh executable (one-time)
# chmod +x start-dev.sh
# to run in terminal: ./start-dev.sh

# Step 1: Start Ollama and DB
docker compose up -d ollama db

# Step 2: Wait for Ollama
until docker compose exec ollama ollama list &>/dev/null; do
  echo "Waiting for Ollama..."
  sleep 5
done
echo "Ollama ready!"

# Step 3: Check for models (informational only - no auto-pulling)
echo "Checking available models from your Ollama installation:"
docker compose exec ollama ollama list

# Optional: Uncomment any of these sections if you need to pull models

# # Arctic Text2SQL model (uncomment if needed)
# if ! docker compose exec ollama ollama list | grep -q "arctic-text2sql:latest"; then
#   echo "Pulling Arctic model..."
#   docker compose exec ollama ollama pull hf.co/mradermacher/Arctic-Text2SQL-R1-7B-GGUF:Q4_K_M
#   docker compose exec ollama ollama cp hf.co/mradermacher/Arctic-Text2SQL-R1-7B-GGUF:Q4_K_M arctic-text2sql:latest
#   echo "Arctic model pulled"
# fi

# # Distil-Qwen3 Text2SQL model (uncomment if needed)
# if ! docker compose exec ollama ollama list | grep -q "distil-qwen3-4b:latest"; then
#   echo "Pulling Distil-Qwen3 model..."
#   docker compose exec ollama ollama pull hf.co/distil-labs/distil-qwen3-4b-text2sql-gguf-4bit
#   docker compose exec ollama ollama cp hf.co/distil-labs/distil-qwen3-4b-text2sql-gguf-4bit distil-qwen3-4b:latest
#   echo "Distil-Qwen3 model pulled"
# fi

# # Qwen2.5-Coder-7B SQL generator model (uncomment if needed)
# if ! docker compose exec ollama ollama list | grep -q "qwen2.5-coder-7b:latest"; then
#   echo "Pulling Qwen2.5-Coder-7B model from Hugging Face..."
#   docker compose exec ollama ollama pull hf.co/bartowski/Qwen2.5-Coder-7B-Instruct-GGUF:Q4_K_M
#   # Optional: create a shorter name for convenience
#   docker compose exec ollama ollama cp hf.co/bartowski/Qwen2.5-Coder-7B-Instruct-GGUF:Q4_K_M qwen2.5-coder-7b:latest
#   echo "Qwen2.5-Coder-7B model pulled and available as 'qwen2.5-coder-7b:latest'"
# fi

# # Qwen2.5-Coder-14B judge model (uncomment if needed)
# if ! docker compose exec ollama ollama list | grep -q "qwen2.5-coder:14b"; then
#   echo "Pulling Qwen2.5-Coder-14B model..."
#   docker compose exec ollama ollama pull qwen2.5-coder:14b
#   echo "Qwen2.5-Coder-14B model pulled"
# fi

# Step 4: Pre-load models into VRAM via API
echo "Pre-loading models into VRAM via API..."

# Function to trigger model load
load_model() {
  local model_name=$1
  echo "Loading ${model_name}..."
  curl -s -X POST http://localhost:11434/api/generate \
    -H "Content-Type: application/json" \
    -d "{
      \"model\": \"${model_name}\",
      \"prompt\": \"\",
      \"stream\": false,
      \"options\": {\"num_predict\": 1}
    }" > /dev/null
  if [ $? -eq 0 ]; then
    echo "${model_name} load triggered."
  else
    echo "Failed to trigger load for ${model_name}. It will load on first use."
  fi
}

# Function to warm up model with a real request
warmup_model() {
  local model_name=$1
  echo "Warming up ${model_name} with test request..."
  curl -s -X POST http://localhost:11434/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d "{
      \"model\": \"${model_name}\",
      \"messages\": [{\"role\": \"user\", \"content\": \"Say hello\"}],
      \"max_tokens\": 10,
      \"temperature\": 0.1
    }" > /dev/null
  if [ $? -eq 0 ]; then
    echo "${model_name} warm-up complete."
  else
    echo "Warning: ${model_name} warm-up failed, but model may still work."
  fi
}

# Read models from .env (ignoring commented lines)
TEXT2SQL_MODEL=$(grep -v '^#' .env | grep TEXT2SQL_MODEL | cut -d '=' -f2)
JUDGE_MODEL=$(grep -v '^#' .env | grep JUDGE_MODEL | cut -d '=' -f2)

if [ -z "$TEXT2SQL_MODEL" ] || [ -z "$JUDGE_MODEL" ]; then
  echo "Error: TEXT2SQL_MODEL or JUDGE_MODEL not found in .env file"
  exit 1
fi

# Load ONLY 2 models total (one SQL generator + one AI response generator / judge model).
echo "Loading SQL generator model..."
load_model "$TEXT2SQL_MODEL"

echo "Loading AI response/judge model..."
load_model "$JUDGE_MODEL"

echo "Waiting for models to fully initialize in VRAM..."
sleep 10

# Step 5: Warm up models with actual requests
echo "Warming up models to prevent first-request truncation..."
warmup_model "$TEXT2SQL_MODEL"
warmup_model "$JUDGE_MODEL"

# Step 6: Verify VRAM usage
echo "Current VRAM usage:"
docker compose exec ollama nvidia-smi | grep "MiB /" | head -1

# Step 7: Start the backend (without --watch flag)
echo "Starting backend services..."
docker compose up -d

echo "Setup complete! Models are loaded and warmed up."