[PavelGuzenfeld]

str(f1_s) == "cuda" in Decoder.forward formats the whole feature tensor on every proj call, and that syncs the device. The value never matters: get_autocast_params overwrites enabled on both branches (True for cuda, False otherwise). Passing True gives the same result without the print.

Output is bitwise identical to main. torch.equal on warp and certainty for the toronto and sacre_coeur demo pairs, fused-local-corr on and off.

Timing on an RTX 3060 laptop, torch 2.11, match() on the toronto pair, 30 runs after a 60 s warmup, main and this branch alternated in separate processes three times. With fused-local-corr: 959, 968, 974 ms on main vs 929, 919, 922 ms here. Without it: 1093 vs 1051 ms. So about 3-5%, only measured on that one GPU.

[PavelGuzenfeld]

Also measured on a Jetson Orin (JetPack 6.2, torch 2.4.0, no fused-local-corr there so use_custom_corr=False). main took 4693 and 4699 ms, this branch 3856 and 3856 ms, about 18% less. Same two demo pairs, output bitwise identical again. Main and the branch alternated in separate processes, twice, 10 runs each after a 60 s warmup.