[Feature] Workarounds for ONNX Export Force CPU

[ozen]

Motivation

In the onnx export pipeline, the device argument is disregarded, and the export operation is performed on the CPU.

mmdeploy/mmdeploy/apis/onnx/export.py

Lines 130 to 137 in 2882c64

	# force to export on cpu
	patched_model = patched_model.cpu()
	if isinstance(args, torch.Tensor):
	args = args.cpu()
	elif isinstance(args, (tuple, list)):
	args = tuple([_.cpu() for _ in args])
	else:
	raise RuntimeError(f'Not supported args: {args}')

I can see this was introduce by #2347

In the PR discussion @grimoire raised the issue with the ops without cpu support, and @RunningLeon suggested adding dummy ops as workaround.

I stumbled upon this when trying to export FasterRCNN with GARPNHead, which utilize MaskedConv2d layer that doesn't have CPU support.

So my questions are:

1. How does adding dummy ops not break the model?
2. If adding dummy ops is possible, how can I do it?
3. Why is forcing cpu export required to support panoptic segmentation model?

Thanks.

Related resources

No response

Additional context

No response

[RunningLeon]

@ozen hi

1. mmdeploy does not have op masked_conv from mmcv for any backend, you have to implement it by yourself.
2. you can simply return a dummy results in the beginning of the forward function: https://github.com/open-mmlab/mmcv/blob/b366e923e46321f8e4eb26e79b622029e0eb2683/mmcv/ops/masked_conv.py#L38C21-L38C21

[github-actions]

This issue is marked as stale because it has been marked as invalid or awaiting response for 7 days without any further response. It will be closed in 5 days if the stale label is not removed or if there is no further response.

[github-actions]

This issue is closed because it has been stale for 5 days. Please open a new issue if you have similar issues or you have any new updates now.