I was trying to write a minimal training code (see #503) - especially when the camera parameters are known (now that we have models like Mast3r, VGGT to estimate the camera parameters). Given the ...

Before my PR, where the profiler hook is removed by the patch, there is one graph break. (pytorch-3.10) [janeyx@devgpu023.odn1 ~/local/pytorch (c0da6b1e)]$ TORCHDYNAMO_VERBOSE=1 ...