fix: Handle dynamic shapes in where ops

[keehyuna]

Description

Dynamic shape cannot be used in where ops because of exception in torch.broadcast_shapes().
Proposed fix removes expand() for static shape input and only performs prepend ones to have same rank size for static/dynamic shape input. I think broadcast is not required as it is applied to ISelectLayer(addSelect)

Fixes # (issue)

Type of change

Please delete options that are not relevant and/or add your own.

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• My code follows the style guidelines of this project (You can use the linters)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas and hacks
• I have made corresponding changes to the documentation
• I have added tests to verify my fix or my feature
• New and existing unit tests pass locally with my changes
• I have added the relevant labels to my PR in so that relevant reviewers are notified

[keehyuna]

If dtype is torch.bool in input_spec, torch.rand() returns random numbers on [0,1) and casting it to bool value is True

[peri044]

Also please mark torch.where converter in aten_ops_converters.py with supports_dynamic_shapes=True flag. Example : https://github.com/pytorch/TensorRT/blob/release/2.3/py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py#L62

[chohk88]

I have left some comments. It looks like there aren't any major issues with the functionality.

[keehyuna]

default parameter was used to merge below function
get_axes_for_reduce_op = functools.partial(
get_axes_for_reduce_op, has_implicit_batch_dimension=False
)

[chohk88]

Looks good to me!

[zewenli98]

@peri044 I think it's reasonable to deprecate FX converter_utils in Dynamo converter implementation. In my opinion, when copying these helper functions from FX utils, it would be better to change the first arg from network: TRTNetwork to ctx: ConversionContext for two reasons: 1) consistent with other helper functions 2) convenient to call them, just passing in ctx.

[zewenli98]

@peri044 do we still have has_implicit_batch_dimension? Is it possible to remove the arg in the dynamo's converter_utils?

[peri044]

1. We should remove has_implicit_batch_dimension.

2. It would be better to change the first arg from network: TRTNetwork to ctx: ConversionContext.

Yes we should make these changes and not use FX code/ data structures as much as possible. If we use them, we should be consistent with dynamo APIs

[keehyuna]

thanks. Removed has_implicit_batch_dimension flag in get_axes_for_reduce_op()
first arg of boardcast()/prepend_ones() is ctx: ConversionContext, instead of network: TRTNetwork

[keehyuna]

Moved "chore: Better random bool values for example_tensor" change into PR2878 to combine random input changes to one PR.
436e99b

[peri044]

Thanks. added a few more minor changes.

[peri044]

LGTM