INSTALLATION_GUIDE.md

PyTorch Python 3.10 Installation Guide

🚀 Quick Start

Option 1: Automated Setup (Recommended)

# 1. Create conda environment
conda create -n pytorch_env python=3.10
conda activate pytorch_env

# 2. Download repository
git clone https://huggingface.co/RDHub/pytorch_python_310
cd pytorch_python_310

# 3. Install all packages
pip install -r lib_wheel/requirements.txt --find-links lib_wheel --no-index

# 4. Set up CUDA libraries
bash setup_cuda_libs.sh

# 5. Test installation
python -c "import torch; print(f'PyTorch {torch.__version__} - CUDA: {torch.cuda.is_available()}')"

Option 2: Manual Setup

# 1. Install packages
pip install -r lib_wheel/requirements.txt --find-links lib_wheel --no-index

# 2. Create CUDA library activation script
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
cat > $CONDA_PREFIX/etc/conda/activate.d/pytorch_cuda_libs.sh << 'EOF'
#!/bin/bash
NVIDIA_LIB_PATH=$(find $CONDA_PREFIX -path "*/nvidia/*/lib" -type d 2>/dev/null | tr '\n' ':')
CUSPARSELT_LIB_PATH=$(find $CONDA_PREFIX -path "*/cusparselt/lib" -type d 2>/dev/null | tr '\n' ':')
export LD_LIBRARY_PATH="${NVIDIA_LIB_PATH}${CUSPARSELT_LIB_PATH}${LD_LIBRARY_PATH}"
EOF
chmod +x $CONDA_PREFIX/etc/conda/activate.d/pytorch_cuda_libs.sh

# 3. Reactivate environment
conda deactivate && conda activate your_env_name

✅ What's Included

• PyTorch 2.7.1 with CUDA 12.6 support
• Transformers 4.52.3 for HuggingFace models
• NumPy 2.0.2 (compatible with OpenCV)
• OpenCV 4.10.0 for computer vision
• 80+ compatible packages tested together
• NVIDIA CUDA libraries (12.6.x series)

🧪 Testing Your Installation

import torch
import transformers
import numpy as np
import cv2

# Test PyTorch CUDA
print(f"PyTorch: {torch.__version__}")
print(f"CUDA Available: {torch.cuda.is_available()}")
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name(0)}")

# Test basic operations
x = torch.randn(100, 100).cuda()
y = torch.matmul(x, x.T)
print("✅ GPU tensor operations working!")

# Test transformers
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
print("✅ Transformers working!")

print(f"NumPy: {np.__version__}")
print(f"OpenCV: {cv2.__version__}")

🔧 Troubleshooting

Issue: "libcufile.so.0: cannot open shared object file"

Solution: Run the CUDA setup script:

bash setup_cuda_libs.sh
conda deactivate && conda activate your_env_name

Issue: "libcusparseLt.so.0: cannot open shared object file"

Solution: Ensure all NVIDIA packages are installed:

pip install --force-reinstall lib_wheel/nvidia_cusparselt_cu12-*.whl
pip install --force-reinstall lib_wheel/nvidia_cufile_cu12-*.whl

Issue: OpenCV + NumPy compatibility errors

Solution: Use the exact versions provided:

pip install --force-reinstall lib_wheel/numpy-2.0.2-*.whl
pip install --force-reinstall lib_wheel/opencv_python-4.10.0.84-*.whl

📋 System Requirements

• OS: Linux x86_64 (Ubuntu 22.04+ recommended)
• Python: 3.10
• CUDA: Compatible with 12.2+ (12.6 optimal)
• Conda: Required for library path management
• Storage: ~2GB for all wheels

🎯 Verified Configurations

✅ Ubuntu 22.04 + Python 3.10 + CUDA 12.2
✅ Ubuntu 22.04 + Python 3.10 + CUDA 12.6
✅ RTX 4090 + CUDA 12.6
✅ Conda environments

🔗 Resources

• Repository: https://huggingface.co/RDHub/pytorch_python_310
• PyTorch Docs: https://pytorch.org/docs/
• CUDA Toolkit: https://developer.nvidia.com/cuda-toolkit