Fix Torch2Trt in Savant 0.6 extra

[denisvmedyantsev]

model = resnet.resnet18().eval().cuda()
dummy_input = torch.full((1, 3, 224, 224), 0.5, dtype=torch.float32).cuda()
model_trt = torch2trt(model, [dummy_input])

> AttributeError: 'tensorrt.tensorrt.INetworkDefinition' object has no attribute 'add_convolution'. Did you mean: 'add_convolution_nd'?

We have a sample with torch2trt - nanosam, see #1110

Consider replacing torch2trt with Torch-TensorRT

[IIIllllIlIlllII]

I looked into this. The root cause is that utils/build_torch2trt.sh pins torch2trt to the v0.5.0 tag, which predates TensorRT 10 support. The v0.5.0 TRTModule uses TRT 8 APIs (get_binding_index, execute_async_v2, etc.) that are removed in TRT 10.

TRT 10 support was added on torch2trt master in NVIDIA-AI-IOT/torch2trt#922 (commit 4e820ae), after the v0.5.0 tag. The fix would be to point the build script to that commit instead — the wheel filename stays the same since setup.py still declares version 0.5.0. The collections.Sequence sed patch would also need retargeting from interpolate.py (deleted) to native_converters.py.

The S3-hosted wheel and .so would need to be rebuilt after merging.

I can submit a PR if this approach looks right. I'd also like to tackle #1110 (nanosam) once this is resolved, since it's blocked by this issue.