LightGlueProject/LightGlue_Deployment/test_tensorrt_integration.py

#!/usr/bin/env python3
"""
测试 TensorRT 集成是否正常工作
"""

import torch

# 测试导入
try:
    import torch_tensorrt
    print("✓ torch_tensorrt imported successfully")
except ImportError as e:
    print(f"✗ Failed to import torch_tensorrt: {e}")
    exit(1)

# 测试基本功能
try:
    # 创建一个简单的模型
    class SimpleModel(torch.nn.Module):
        def __init__(self):
            super().__init__()
            self.conv = torch.nn.Conv2d(1, 64, 3, padding=1)
        
        def forward(self, x):
            return self.conv(x)
    
    model = SimpleModel().eval().cuda()
    example_input = torch.randn(1, 1, 480, 640).cuda()
    
    # 编译为 TensorRT
    print("Compiling model with TensorRT...")
    trt_model = torch_tensorrt.compile(
        model,
        inputs=[example_input],
        enabled_precisions={torch.float, torch.half},
        workspace_size=1 << 30,
    )
    
    # 测试推理
    with torch.no_grad():
        output = trt_model(example_input)
    
    print("✓ TensorRT compilation and inference successful!")
    print(f"  Input shape: {example_input.shape}")
    print(f"  Output shape: {output.shape}")
    
except Exception as e:
    print(f"✗ TensorRT test failed: {e}")
    import traceback
    traceback.print_exc()
    exit(1)

print("\n" + "="*60)
print("TensorRT integration test passed!")
print("You can now use --use_tensorrt flag in your demo script")
print("="*60)