Make "RuntimeError: operator torchvision::nms does not exist" error better

[zou3519]

I'm putting this issue in PyTorch repo because I'm not sure where the fix actually goes.
If some shared library is missing, and torchvision._C fails to import, then I get the above error.
Can we do something about this to make the error message clearer?

cc @atalman @NicolasHug

[malfet]

IMO fix should go into torchvision, which since times immemorial had this operational mode when it's partially usable without _C library (probably to support loading resnet thru torchhub), but not sure if this is still relevant

[NicolasHug]

@zou3519 would you have a reproducing example by any chance? I'm trying to understand why we don't hit this better error message first:

vision/torchvision/extension.py

Lines 44 to 54 in 1e53952

	def _assert_has_ops():
	if not _has_ops():
	raise RuntimeError(
	"Couldn't load custom C++ ops. This can happen if your PyTorch and "
	"torchvision versions are incompatible, or if you had errors while compiling "
	"torchvision from source. For further information on the compatible versions, check "
	"https://github.com/pytorch/vision#installation for the compatibility matrix. "
	"Please check your PyTorch version with torch.__version__ and your torchvision "
	"version with torchvision.__version__ and verify if they are compatible, and if not "
	"please reinstall torchvision so that it matches your PyTorch install."
	)

since it should be called before accessing nms:

vision/torchvision/ops/boxes.py

Lines 47 to 48 in 1e53952

	_assert_has_ops()
	return torch.ops.torchvision.nms(boxes, scores, iou_threshold)

---

Regardless, I'd love to raise loudly instead of silencing errors when loading the ops in

vision/torchvision/extension.py

Lines 40 to 41 in 1e53952

	except (ImportError, OSError):
	pass

I think you're right @malfet that this was meant for torchhub, but it might be time to break that.

[umutonuryasar]

Hi, I'd like to take a stab at this if it's still open.

From reading the thread, it seems the root cause is that the ImportError/OSError in extension.py is silently caught, so when a user later calls nms, they get a confusing operator does not exist error instead of something actionable.

I'm thinking the fix could go in one of two directions:

1. Emit a warnings.warn at import time with the original exception message, so users see it immediately.
2. Re-raise a more descriptive error (effectively "breaking" the silent-catch as @NicolasHug suggested), accepting that this may affect torchhub workflows that rely on partial usability.

Before I proceed — is direction 2 the preferred approach, and is the torchhub use case still something we need to preserve?

Happy to put together a draft PR once the approach is confirmed.

[hiSandog]

A clearer diagnostic would help most when _C fails to import but the later fake/meta registration is what users see. One possible path is to cache the original _C import exception and include it when registering ops such as torchvision::nms fails, e.g. pointing at a missing shared library or a torch/torchvision wheel mismatch. That would turn the current symptom into an actionable packaging/runtime message instead of making users debug the operator registry first.

[umutonuryasar]

Thanks @hiSandog — caching the original _C import exception and surfacing it at op-registration time is a cleaner version of what I had in mind for Direction 1. That gives users an actionable packaging/runtime message without breaking the lazy-import behavior.

I'll go ahead and put together a draft PR along these lines unless there's a preference from maintainers for a different approach.

[YeonwooSung]

I'll work on this: cache the extension load error and avoid registering fake torchvision::nms when the C++ ops are missing, so the failure message is actionable.

[YeonwooSung]

Proposed fix: #9649

register_fake for torchvision::nms / qnms is now skipped when compiled ops are missing, so import torchvision no longer fails with operator torchvision::nms does not exist. The original extension load error is cached and included in _assert_has_ops().