Fix PyTorch 2.7+ and ROCm/AMD GPU Compatibility

[AKharytonchyk]

Pull Request: Fix PyTorch 2.7+ and ROCm/AMD GPU Compatibility

Summary

This PR fixes critical compatibility issues with PyTorch 2.7+ and adds enhanced support for ROCm/AMD GPUs in ComfyUI Impact Subpack.

Problem Statement

When using PyTorch 2.7.0+rocm6.3 with AMD GPUs, the UltralyticsDetectorProvider node fails to load YOLO models with the following error:

WeightsUnpickler error: Unsupported global: GLOBAL getattr was not an allowed global by default. 
Please use `torch.serialization.add_safe_globals([getattr])` or the `torch.serialization.safe_globals([getattr])` 
context manager to allowlist this global if you trust this class/function.

Root Cause

PyTorch 2.6+ introduced enhanced security features where weights_only=True became the default for torch.load(). YOLO models require the getattr builtin function during deserialization, which is blocked by default for security reasons.

Solution

1. PyTorch 2.7+ Compatibility Fix

• Added getattr to PyTorch safe globals during module initialization
• Maintains PyTorch's security features while allowing trusted YOLO models to load
• Uses torch.serialization.add_safe_globals([getattr]) as recommended by PyTorch documentation

2. ROCm/AMD GPU Auto-Detection

• Enhanced inference_bbox() and inference_segm() functions with automatic device detection
• Automatically uses CUDA/ROCm when available and no device is specified
• Improves performance on AMD GPUs with ROCm support

3. Comprehensive Documentation and Testing

• Added ROCM_FIXES.md with detailed fix documentation
• Included test_rocm_compatibility.py for validation
• Maintains backward compatibility with existing installations

Code Changes

modules/subcore.py

# Add getattr to safe globals for PyTorch 2.6+ compatibility
if hasattr(torch.serialization, 'add_safe_globals'):
    torch.serialization.add_safe_globals([getattr])
    logging.info("[Impact Pack/Subpack] Added getattr to PyTorch safe globals for YOLO model compatibility")

# Enhanced device auto-detection in inference functions
def inference_bbox(..., device: str = ""):
    # If device is empty and CUDA/ROCm is available, use it
    if not device and torch.cuda.is_available():
        device = "cuda"
    ...

def inference_segm(..., device: str = ""):
    # If device is empty and CUDA/ROCm is available, use it
    if not device and torch.cuda.is_available():
        device = "cuda"
    ...

Testing Environment

• OS: Linux
• Python: 3.12.3
• PyTorch: 2.7.0+rocm6.3
• GPU: AMD Radeon RX 7900 XTX
• ComfyUI: 0.3.44

Test Results

✅ All YOLO models (bbox and segm) load successfully
✅ ROCm device detection working correctly
✅ No security warnings or errors
✅ Backward compatibility maintained
✅ Performance improved on AMD GPUs

Impact

• Fixes: Critical model loading failures on PyTorch 2.7+
• Enhances: AMD GPU support with ROCm
• Maintains: Full backward compatibility
• Improves: Performance on systems with GPU acceleration

Related Issues

This addresses the PyTorch 2.6+ security changes and ROCm compatibility issues reported by users upgrading to newer PyTorch versions with AMD GPUs.

Breaking Changes

None. All changes are backward compatible and only activate when:

1. PyTorch 2.6+ is detected (for safe globals)
2. CUDA/ROCm is available (for device auto-detection)

Files Modified

• modules/subcore.py - Core compatibility fixes
• ROCM_FIXES.md - Documentation (new)
• test_rocm_compatibility.py - Test script (new)