Empty tensor handling

[apbose]

This PR addresses the case of empty tensor in torchTRT based on https://docs.nvidia.com/deeplearning/tensorrt/latest/inference-library/advanced.html#empty-tensors, and also focuses on concat operation edge case involving empty tensor

TODO: Might have to separate the case of concat from this PR, in the case when torch.Tensor([]) and a rank greater tensor are concatenated, which is a valid case in pytorch but not TRT.

This addressed #3865. Corresponding HF transformers issue raised - huggingface/transformers#42027 where empty tensor should not come in the first place

[narendasan]

@apbose this maybe be a case where we would want TRT to properly handle this rather than us doing something hacky. Lets raise it with Yuan Yuo. Dummy inputs do not feel like the right solution

[narendasan]

I strongly want to avoid having nullptr basically anywhere, we should be looking for some sane default

[narendasan]

Its looking good, please add a testcase then should be good to merge

[narendasan]

Why is this needed for empty tensor handling? Also are we near the point we can just remove FX converters?

[apbose]

This is particularly for the case where x is torch.Tensor([]), and y is torch.Tensor torch.randn(3,4) where the validator comes into play.

# makes use of validator
class ConcatEmptyModelEmptyConstantMisMatchDim(nn.Module):
    def __init__(self, dim=0):
        super().__init__()
        self.dim = dim

    def forward(self, x):
        y = torch.tensor([], device="cuda")
        return torch.cat([x, y], dim=self.dim)

This error comes -

"/code/torchTRT/TensorRT/py/torch_tensorrt/fx/converters/aten_ops_converters.py", line 465, in aten_ops_cat                                                                                                                                                                                                                                  
    return acc_ops_converters.acc_ops_cat(network, target, None, kwargs_new, name) 

acc_ops converter is called since validator fails and it resorts to the fx converter implementation which is next in precedence. Yes we should remove the converters IMO, but till that is not done, this would be required I suppose?

[apbose]

Also for the test case, class TestEmptyTensorMemoryLeak(TestCase): should cover the case of no memory leak or memory overload in case of repeated model calls. Let me know if you think any other test case is required