🐛 [Bug] aten.convolution fails when 1D stride is tuple

[gs-olive]

Bug Description

When encountering a Conv1D operator in Dynamo, PyTorch can switch integer components, like dilation, into tuples. This is problematic for the Conv1D operator, since the extend_attr_to_tuple used here, will not work:

TensorRT/py/torch_tensorrt/fx/converters/impl/convolution.py

Lines 111 to 115 in b3089bf

	# Expand parameters manually for Conv1D computations
	if is_conv1d:
	padding = tuple(padding) + (0,)
	stride = extend_attr_to_tuple(stride, 2)
	dilation = extend_attr_to_tuple(dilation, 2)

Specifically, the function extend_attr_to_tuple, shown below, needs to be modified to do the following. It should be able to handle length-1 lists or tuples and extend those to the necessary length specified by the user.

TensorRT/py/torch_tensorrt/fx/converters/converter_utils.py

Lines 128 to 146 in b3089bf

	def extend_attr_to_tuple(
	val: Any,
	num_elem: int,
	) -> Tuple[Any, ...]:
	"""
	If `val` is not a tuple or a list, then we make a tuple of size `num_elem` by
	replicating `val` `num_elem` times.
	Args:
	val (Any): Value that we want to process.
	Returns:
	A tuple.
	"""
	if not isinstance(val, (tuple, list)):
	val = (val,) * num_elem
	if isinstance(val, list):
	val = tuple(val)
	return val

To Reproduce

Compile a model with a Conv1D operator using one of the Dynamo paths.

Expected behavior

Models with Conv1D operators should trace and compile successfully with AOT/Dynamo.

Environment

• Torch-TensorRT Version (e.g. 1.0.0): 8c62fca
• PyTorch Version (e.g. 1.0): 2.1.0.dev20230803+cu121

[zewenli98]

It should be able to handle length-1 lists or tuples and extend those to the necessary length specified by the user.

Sorry I don't get it. Could you give an example for this?

[gs-olive]

@zewenli98 - sure, for instance in the conv1d case, we can get inputs like stride=[1] and dilation=[0]. In those cases, the function extend_attr_to_tuple will not extend those lists because they are not single values. Specifically, extend_attr_to_tuple([1], 2) returns [1], but we want it to return [1, 1]. It would help to have some logic in the extend_attr_to_tuple function which detects singleton lists/tuples and treats them like single elements. For instance:

if isinstance(val, (tuple, list)) and len(val) == 1:
  val = val[0]

[zewenli98]

@gs-olive Thanks for the detailed explanation! It's fixed in the PR #2252