Should we turn trace_only default to True in torchlib?

[titaiwangms]

The PR: #2791 removed inliner pass from version converter and broke benchmark.

Some un-traced functions are revealed by benchmarking:

onnxscript/onnxscript/function_libs/torch_lib/ops/core.py

Lines 9071 to 9088 in e6f79e1

	@torch_op(("aten::split", "aten::split.Tensor"))
	def aten_split(self: TTensor, split_size: INT64, dim: int = 0) -> TTensor:
	"""split.Tensor(Tensor(a -> *) self, SymInt split_size, int dim=0) -> Tensor(a)[]"""
	return op.SplitToSequence(self, split_size, axis=dim)
	def aten_split_copy(self: TensorType, split_size: INT64, dim: int = 0) -> TensorType:
	"""split_copy.Tensor(Tensor self, SymInt split_size, int dim=0) -> Tensor[]"""
	raise NotImplementedError()
	@torch_op("aten::split_with_sizes")
	def aten_split_with_sizes(self: TTensor, split_sizes: INT64, dim: int = 0) -> TTensor:
	"""split_with_sizes(Tensor(a -> *) self, SymInt[] split_sizes, int dim=0) -> Tensor(a)[]"""
	return op.SplitToSequence(self, split_sizes, axis=dim)

Although I will add back the inliner into version converter, should we trace all our functions ahead in torchlib? Or there is something blocking us?

cc @justinchuby @xadupre @gramalingam

[justinchuby]

Sure. There are some functions to be rewritten though