…332)

[Edit: The issue was encountered while attempting to compile the model
`yolov3`]

Currently the
[`setitem`](https://github.com/CentML/hidet/blob/566f0fe55f441326c3034b7eed44b3fa0b03f38d/python/hidet/graph/frontend/torch/register_functions.py#L280)
function in Hidet will fail on two special scenarios when `setvalue` is
a tensor:

1. When `setvalue` and `x` are of different dtypes, currently there will
be an error that looks like:
> RuntimeError: If-then-else operand 1 and 2 have different types
(hidet.float16 vs hidet.float32) ((((v < 0) || (2 <= v)) ? false :
(((v_1 < 0) || (3 <= v_1)) ? false : (((v_2 < 0) || (3 <= v_2)) ? false
: true))) ? setvalue[v_2, v_1, v] : data[v_2, v_1, v]), occurred when
interpreting operator.setitem with
>   setitem(tensor(...), (Ellipsis, slice(None, 2, None)), tensor(...))

Whereas in PyTorch `setvalue` appears to be casted to the same datatype
as `x` if possible.

2. When `setvalue` and `x` are on different devices, currently this will
result in an error:

> RuntimeError: All inputs of an operator must be on the same device,
occurred when interpreting operator.setitem with
>   setitem(tensor(...), (Ellipsis, slice(None, 2, None)), tensor(...))

Whereas in PyTorch the `setvalue` is moved to the same device as `x`.

…332)

[Edit: The issue was encountered while attempting to compile the model
`yolov3`]

Currently the
[`setitem`](https://github.com/CentML/hidet/blob/566f0fe55f441326c3034b7eed44b3fa0b03f38d/python/hidet/graph/frontend/torch/register_functions.py#L280)
function in Hidet will fail on two special scenarios when `setvalue` is
a tensor:

1. When `setvalue` and `x` are of different dtypes, currently there will
be an error that looks like:
> RuntimeError: If-then-else operand 1 and 2 have different types
(hidet.float16 vs hidet.float32) ((((v < 0) || (2 <= v)) ? false :
(((v_1 < 0) || (3 <= v_1)) ? false : (((v_2 < 0) || (3 <= v_2)) ? false
: true))) ? setvalue[v_2, v_1, v] : data[v_2, v_1, v]), occurred when
interpreting operator.setitem with
>   setitem(tensor(...), (Ellipsis, slice(None, 2, None)), tensor(...))

Whereas in PyTorch `setvalue` appears to be casted to the same datatype
as `x` if possible.

2. When `setvalue` and `x` are on different devices, currently this will
result in an error:

> RuntimeError: All inputs of an operator must be on the same device,
occurred when interpreting operator.setitem with
>   setitem(tensor(...), (Ellipsis, slice(None, 2, None)), tensor(...))

Whereas in PyTorch the `setvalue` is moved to the same device as `x`.

…332)

[Edit: The issue was encountered while attempting to compile the model
`yolov3`]

Currently the
[`setitem`](https://github.com/CentML/hidet/blob/566f0fe55f441326c3034b7eed44b3fa0b03f38d/python/hidet/graph/frontend/torch/register_functions.py#L280)
function in Hidet will fail on two special scenarios when `setvalue` is
a tensor:

1. When `setvalue` and `x` are of different dtypes, currently there will
be an error that looks like:
> RuntimeError: If-then-else operand 1 and 2 have different types
(hidet.float16 vs hidet.float32) ((((v < 0) || (2 <= v)) ? false :
(((v_1 < 0) || (3 <= v_1)) ? false : (((v_2 < 0) || (3 <= v_2)) ? false
: true))) ? setvalue[v_2, v_1, v] : data[v_2, v_1, v]), occurred when
interpreting operator.setitem with
>   setitem(tensor(...), (Ellipsis, slice(None, 2, None)), tensor(...))

Whereas in PyTorch `setvalue` appears to be casted to the same datatype
as `x` if possible.

2. When `setvalue` and `x` are on different devices, currently this will
result in an error:

> RuntimeError: All inputs of an operator must be on the same device,
occurred when interpreting operator.setitem with
>   setitem(tensor(...), (Ellipsis, slice(None, 2, None)), tensor(...))

Whereas in PyTorch the `setvalue` is moved to the same device as `x`.