Skip to content

RAPIDS_SETUP.md

Latest commit

 

History

History
280 lines (199 loc) · 6.8 KB

RAPIDS_SETUP.md

File metadata and controls

280 lines (199 loc) · 6.8 KB

RAPIDS GPU Setup Guide

This guide explains how to set up NVIDIA RAPIDS cuML for GPU-accelerated demand forecasting.

Prerequisites

Hardware Requirements

  • NVIDIA GPU with CUDA Compute Capability 7.0+ (Volta, Turing, Ampere, Ada, Hopper)
  • 16GB+ GPU memory (recommended for large datasets)
  • 32GB+ system RAM (recommended)
  • CUDA 11.2+ or 12.0+ installed

Software Requirements

  • Python 3.9-3.11 (RAPIDS supports these versions)
  • NVIDIA GPU drivers (latest recommended)
  • CUDA Toolkit (11.2+ or 12.0+)

Installation Methods

Method 1: Pip Installation (Recommended for Virtual Environments)

This is the easiest method for development environments:

# Activate your virtual environment
source env/bin/activate

# Run the installation script
./scripts/setup/install_rapids.sh

Or manually:

# Upgrade pip
pip install --upgrade pip setuptools wheel

# Install RAPIDS from NVIDIA PyPI
pip install --extra-index-url=https://pypi.nvidia.com \
    cudf-cu12 \
    cuml-cu12 \
    cugraph-cu12 \
    cuspatial-cu12 \
    cuproj-cu12 \
    cusignal-cu12 \
    cuxfilter-cu12 \
    cudf-pandas \
    dask-cudf-cu12 \
    dask-cuda

Note: Replace cu12 with cu11 if you have CUDA 11.x installed.

Method 2: Conda Installation (Recommended for Production)

Conda is the recommended method for production deployments:

# Create a new conda environment with RAPIDS
conda create -n rapids-env -c rapidsai -c conda-forge -c nvidia \
    rapids=24.02 python=3.10 cudatoolkit=12.0

# Activate the environment
conda activate rapids-env

# Install additional dependencies
pip install asyncpg psycopg2-binary redis

Method 3: Docker (Recommended for Isolation)

Use the provided Docker Compose configuration:

# Build and run RAPIDS container
docker-compose -f deploy/compose/docker-compose.rapids.yml up -d

# Or build manually
docker build -f Dockerfile.rapids -t warehouse-rapids .
docker run --gpus all -it warehouse-rapids

Verification

1. Check GPU Availability

nvidia-smi

You should see your GPU listed with driver and CUDA version.

2. Test RAPIDS Installation

# Test cuDF (GPU DataFrames)
python -c "import cudf; df = cudf.DataFrame({'a': [1,2,3], 'b': [4,5,6]}); print(df); print('✅ cuDF working')"

# Test cuML (GPU Machine Learning)
python -c "import cuml; from cuml.ensemble import RandomForestRegressor; print('✅ cuML working')"

# Test GPU memory
python -c "import cudf; import cupy as cp; print(f'GPU Memory: {cp.get_default_memory_pool().get_limit() / 1e9:.2f} GB')"

3. Test Forecasting Script

# Run the RAPIDS forecasting script
python scripts/forecasting/rapids_gpu_forecasting.py

You should see:

✅ RAPIDS cuML detected - GPU acceleration enabled

Usage

Running GPU-Accelerated Forecasting

The forecasting system automatically detects RAPIDS and uses GPU acceleration when available:

from scripts.forecasting.rapids_gpu_forecasting import RAPIDSForecastingAgent

# Initialize agent (will use GPU if RAPIDS is available)
agent = RAPIDSForecastingAgent()

# Run forecasting
await agent.initialize_connection()
forecast = await agent.run_batch_forecast(skus=['SKU001', 'SKU002'])

Via API

The forecasting API endpoints automatically use GPU acceleration when RAPIDS is available:

# Start the API server
./scripts/start_server.sh

# Trigger GPU-accelerated training
curl -X POST http://localhost:8001/api/v1/training/start \
  -H "Content-Type: application/json" \
  -d '{"training_type": "advanced"}'

Via UI

  1. Navigate to the Forecasting page: http://localhost:3001/forecasting
  2. Click "Start Training" with "Advanced" mode selected
  3. The system will automatically use GPU acceleration if RAPIDS is available

Performance Benefits

GPU acceleration provides significant performance improvements:

  • Training Speed: 10-100x faster than CPU for large datasets
  • Batch Processing: Process multiple SKUs in parallel
  • Memory Efficiency: Better memory utilization for large feature sets
  • Scalability: Handle larger datasets that would be impractical on CPU

Example Performance

Dataset Size CPU Time GPU Time Speedup
1,000 rows 2.5s 0.8s 3.1x
10,000 rows 25s 1.2s 20.8x
100,000 rows 250s 3.5s 71.4x

Troubleshooting

Issue: "RAPIDS cuML not available - falling back to CPU"

Causes:

  1. RAPIDS not installed
  2. CUDA not available
  3. GPU not detected

Solutions:

# Check GPU
nvidia-smi

# Check CUDA
nvcc --version

# Reinstall RAPIDS
./scripts/setup/install_rapids.sh

Issue: "CUDA out of memory"

Causes:

  • Dataset too large for GPU memory
  • Multiple processes using GPU

Solutions:

  1. Reduce batch size in configuration
  2. Process SKUs in smaller batches
  3. Use CPU fallback for very large datasets
  4. Free GPU memory: python -c "import cupy; cupy.get_default_memory_pool().free_all_blocks()"

Issue: "Driver/library version mismatch"

Causes:

  • NVIDIA driver and CUDA library versions don't match

Solutions:

# Restart NVIDIA driver
sudo systemctl restart nvidia-persistenced

# Or reboot the system
sudo reboot

Issue: Import errors

Causes:

  • Wrong CUDA version package installed
  • Missing dependencies

Solutions:

# Uninstall and reinstall with correct CUDA version
pip uninstall cudf cuml
pip install --extra-index-url=https://pypi.nvidia.com cudf-cu12 cuml-cu12

Configuration

Environment Variables

Set these in your .env file:

# Enable GPU acceleration
USE_GPU=true

# GPU memory fraction (0.0-1.0)
GPU_MEMORY_FRACTION=0.8

# CUDA device ID
CUDA_VISIBLE_DEVICES=0

Code Configuration

# In rapids_gpu_forecasting.py
config = {
    "use_gpu": True,  # Enable GPU acceleration
    "gpu_memory_fraction": 0.8,  # Use 80% of GPU memory
    "batch_size": 1000,  # Process 1000 rows at a time
}

Best Practices

  1. Use Conda for Production: More stable and better dependency management
  2. Monitor GPU Memory: Use nvidia-smi to monitor usage
  3. Batch Processing: Process multiple SKUs in batches for better GPU utilization
  4. Fallback to CPU: Always have CPU fallback for systems without GPU
  5. Memory Management: Free GPU memory between batches if processing large datasets

Additional Resources

Support

For issues specific to RAPIDS: