Closed mnauf closed 5 months ago
Hi @mnauf, thank you for pointing out this issue. I pushed a fix to PR #124. Please let us know if this solves your problem.
Thanks @Phil26AT, it works now! Are the following warnings expected?
/home/mnauf/anaconda3/envs/lightglue2/lib/python3.9/site-packages/torch/_inductor/compile_fx.py:140: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
warnings.warn(
Run benchmark for: LightGlue-adaptive-compile
/home/mnauf/Desktop/lightglue/benchmark.py:187: UserWarning: Point pruning is partially disabled for compiled forward.
matcher.compile()
Great! Yes these warning are expected. The first one is from torch. You could change the matmul precision to "high", but this sometimes results in wrong estimates. For safety reasons we do not enable it by default.
The second warning is about point pruning, which is not compatible with the padding and masking used in compiled forward. Newer torch versions support dynamic shapes, which would be compatible with point pruning, but we have not tested this with LightGlue yet.
I have torch >=2, yet I can't run lightglue_full_compile. Logs are attached. Please help.
Environment:
Logs: