Add PyTorch 2.x compatibility and modern GPU support (CUDA 12.8, RTX 50 series)

[chansoopark98]

Summary

This PR adds support for modern NVIDIA GPUs (RTX 40/50 series) with CUDA 12.8 and PyTorch 2.8.0.

Changes

1. PyTorch 2.x API Compatibility

Fixed deprecated .type() calls that cause build failures with PyTorch 2.x:

• dpvo/altcorr/correlation_kernel.cu - 4 fixes
• dpvo/lietorch/src/lietorch_gpu.cu - 19 fixes
• dpvo/lietorch/src/lietorch_cpu.cpp - 19 fixes
• dpvo/lietorch/include/dispatch.h - Updated macro

Before (PyTorch 1.x):
AT_DISPATCH_FLOATING_TYPES(tensor.type(), "kernel_name", ([&] { ... }));

After (PyTorch 2.x):
AT_DISPATCH_FLOATING_TYPES(tensor.scalar_type(), "kernel_name", ([&] { ... }));

2. DPViewer Build Fixes

DPViewer/CMakeLists.txt:

• Fixed Python detection for conda environments
• Added C++11 ABI compatibility flag to match PyTorch 2.8.0

DPViewer/setup.py:

• Added explicit Python executable paths for CMake

3. Installation Documentation

Added comprehensive INSTALL.md with:

• Step-by-step installation instructions
• CUDA architecture reference table (RTX 20/30/40/50 series)
• Explanation of --no-build-isolation requirement
• Detailed C++ ABI compatibility guide
• Troubleshooting for common errors

Motivation

The original DPVO was built for PyTorch 1.x and older NVIDIA GPUs. This PR enables:

1. PyTorch 2.x support - PyTorch 2.x deprecated the .type() method
2. Modern GPU support - RTX 40 series (sm_89) and RTX 50 series (sm_120)
3. CUDA 12.x support - Required for latest GPUs
4. Python 3.12 support - Latest Python version

Testing

Tested on:

• GPU: NVIDIA RTX 5090 (Blackwell, compute capability 12.0)
• CUDA: 12.8
• PyTorch: 2.8.0
• Python: 3.12

Verified:

• DPVO imports successfully
• CUDA extensions load correctly
• DPViewer visualization works
• Demo script runs on sample data

Breaking Changes

None. All changes are backward compatible.

(Installation Guide)

INSTALL.md

[Lantianyou]

Modernize this project would be tremendously helpful