potential torch backend or model compilation issue #1
Description
Spent way too long trying to get rtdetr model to compile. Even after getting it to compile on ubuntu I got stuck on
# Apply compilation and warmup for model
compiled_detector = torch.compile(detector, fullgraph=True, mode="reduce-overhead")
# Compilation happens in the first forward pass, so we need to warmup the model
for _ in range(10):
with torch.no_grad():
compiled_detector(**detector_inputs)
I forgot the exact trace, but your code was passing in inverse or something like that. Then, I just stopped looking through this because of security concern:
Is it safe to compile models from PekingU? Literally there's no wiki available and they have no verifiable social media links or webpages on their bio. This is a huge security concern knowing that pytorch models can contain scripts, not just model weights.