need for casting in aten::scatter.value

[linshokaku]

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

Lines 7739 to 7749 in 81f8444

	@torch_op(("aten::scatter.value", "aten::scatter.src"), trace_only=True)
	def aten_scatter(
	self: TReal,
	dim: int, # we have to use int here because ScatterElements() will use this attribute
	index: TInt,
	src: TReal,
	) -> TReal:
	"""scatter_add(Tensor self, int dim, Tensor index, Tensor src) -> Tensor"""
	update = op.Expand(src, op.Shape(index))
	return op.ScatterElements(self, index, update, axis=dim)

The type compatibility between self and value in aten::scatter.value is not required, as implicit casting occurs. Therefore, type compatibility must also be enforced by CastLike in the ops implementation.

>>> data = torch.zeros((3, 3)).float()
>>> indices = torch.tensor([[1, 0, 2], [0, 2, 1]])
>>> torch.ops.aten.scatter.value(data, 0, indices, 1)
tensor([[1., 1., 0.],
        [1., 0., 1.],
        [0., 1., 1.]])
>>> torch.ops.aten.scatter.value(data, 0, indices, int(1))
tensor([[1., 1., 0.],
        [1., 0., 1.],
        [0., 1., 1.]])
>>> torch.ops.aten.scatter.value(data, 0, indices, float(1))
tensor([[1., 1., 0.],
        [1., 0., 1.],
        [0., 1., 1.]])
>>> torch.ops.aten.scatter.value(data, 0, indices, True)
tensor([[1., 1., 0.],
        [1., 0., 1.],
        [0., 1., 1.]])