[Orin Nano] Error when trying to use sparsity

[kyflores]

Hi,

I'm trying to use Apex's sparsity tools on an Orin Nano Super, running the Jetpack 6.1 SDCard image that was released recently with the Super devkit. To access torch, apex, etc, I'm running everything inside the nvcr.io/nvidia/pytorch:24.12-py3-igpu container.

Here's my minimal repro:

# root@4ddd9f30dc2e:/host# cat repro.py 
import torch
import torchvision
from apex.contrib.sparsity import ASP

model = torchvision.models.resnet18(weights=None)
model = model.cuda().train()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4)
ASP.prune_trained_model(model, optimizer)

I would expect this to complete successfully, as this testcase seems to work on and x86_64 + RTX dGPU desktop using the analogous nvcr.io/nvidia/pytorch:24.12-py3 container.

Instead, it fails with this trace: asp_error.log

I'm not sure what causes this, AttributeError: 'ResNet' object has no attribute 'module'. Did you mean: 'modules'?

But I've seen other projects guard the torch.distributed calls with torch.distributed.is_available() to catch cases like the jetson torch build where distributed is disabled.

Thanks.

[ankpatr1]

Hi ,

I checked this issue according to my understanding it happen because of you last line code : ASP.prune_trained_model(model, optimizer)

maybe its taking this model.module value.

add this "try: ASP.prune_trained_model(model_for_pruning, optimizer) print(" successful model pruning") except AttributeError as e: print(f"model pruning error : {e}")