Exception Handling

As Ivy is unifying multiple backends, various issues are seen during exception handling:

1. each backend throws its own exceptions

2. exceptions thrown are backend-specific, therefore inconsistent

To unify the handling of exceptions and assertions, Ivy includes a custom exception class and decorator, which are explained further in the following sub-sections.

Ivy Exception Class

Firstly, Ivy’s base exception class is IvyException class, which inherits from the Python Exception class.

# in ivy/utils/exceptions.py
class IvyException(Exception):
    def __init__(self, *messages, include_backend=False):
        self.native_error = (
            messages[0]
            if len(messages) == 1
            and isinstance(messages[0], Exception)
            and not include_backend
            else None
        )
        if self.native_error is None:
            super().__init__(
                _combine_messages(*messages, include_backend=include_backend)
            )
        else:
            super().__init__(str(messages[0]))

In cases where an exception class for a specific purpose is required, we inherit from the IvyException class. For example, the IvyBackendException class is created to unify backend exceptions.

# in ivy/utils/exceptions.py
class IvyBackendException(IvyException):
    def __init__(self, *messages, include_backend=False):
        super().__init__(*messages, include_backend=include_backend)

In some Array API tests, IndexError and ValueError are explicitly tested to ensure that the functions are behaving correctly. Thus, the IvyIndexError and IvyValueError classes unifies these special cases. For a more general case, the IvyError class can be used.

# in ivy/utils/exceptions.py
class IvyError(IvyException):
    def __init__(self, *messages, include_backend=False):
        super().__init__(*messages, include_backend=include_backend)

More Custom Exception classes were created to unify sub-categories of errors. We try our best to ensure that the same type of Exception is raised for the same type of Error regardless of the backend. This will ensure that the exceptions are truly unified for all the different types of errors. The implementations of these custom classes are exactly the same as IvyError class. Currently there are 5 custom exception classes in ivy.

1. IvyIndexError: This Error is raised for anything Indexing related. For Instance, providing out of bound axis in any function.

2. IvyValueError: This is for anything related to providing wrong values. For instance, passing high value

   smaller than low value in ivy.random_uniform.

3. IvyAttributeError: This is raised when an undefined attribute is referenced.

4. IvyBroadcastShapeError: This is raised whenever 2 shapes are expected to be broadcastable but are not.

5. IvyDtypePromotionError: Similar to IvyBroadcastShapeError, this is raised when 2 dtypes are expected to be promotable but are not.

The correct type of Exception class should be used for the corresponding type of error across the backends. This will truly unify all the exceptions raised in Ivy.

Configurable Mode for Stack Trace

Ivy’s transpilation nature allows users to write code in their preferred frontend framework and then execute it with a different backend framework. For example, a user who is comfortable with NumPy can use Ivy’s NumPy frontend to run their code with a JAX backend. However, since they may have no prior experience with JAX or other backend frameworks, they may not want to encounter stack traces that traverse Ivy and JAX functions. In such cases, it may be preferable for the user to avoid encountering stack traces that extend through Ivy and JAX functions.

Therefore, options are made available for the stack traces to either truncate at the frontend or ivy level, or in other cases, no truncation at all.

Let’s look at the 3 different modes with an example of ivy.all below!

1. Full

This is the default mode and keeps the complete stack traces. All numpy frontend, ivy specific, and native jax stack traces are displayed. The format of the error displayed in this mode is Ivy error: backend name: backend function name: native error: error message

>>> ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
  File "/ivy/ivy/utils/exceptions.py", line 198, in _handle_exceptions
    return fn(*args, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 911, in _handle_nestable
    return fn(*args, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 392, in _handle_array_like_without_promotion
    return fn(*args, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 805, in _handle_out_argument
    return fn(*args, out=out, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 432, in _inputs_to_native_arrays
    return fn(*new_args, **new_kwargs)
  File "/ivy/ivy/func_wrapper.py", line 535, in _outputs_to_ivy_arrays
    ret = fn(*args, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 349, in _handle_array_function
    return fn(*args, **kwargs)
  File "/ivy/ivy/functional/backends/jax/utility.py", line 22, in all
    raise ivy.utils.exceptions.IvyIndexError(error)

During the handling of the above exception, another exception occurred:

  File "/ivy/other_test.py", line 22, in <module>
    ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 523, in _handle_numpy_out
    return fn(*args, **kwargs)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 396, in _outputs_to_numpy_arrays
    ret = fn(*args, **kwargs)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 352, in _inputs_to_ivy_arrays_np
    return fn(*ivy_args, **ivy_kwargs)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 453, in _from_zero_dim_arrays_to_scalar
    ret = fn(*args, **kwargs)
  File "/ivy/ivy/functional/frontends/numpy/logic/truth_value_testing.py", line 24, in all
    ret = ivy.all(a, axis=axis, keepdims=keepdims, out=out)
  File "/ivy/ivy/utils/exceptions.py", line 217, in _handle_exceptions
    raise ivy.utils.exceptions.IvyIndexError(

IvyIndexError: jax: all: ValueError: axis 2 is out of bounds for an array of dimension 1

2. Frontend-only

This option displays only frontend-related stack traces. If compared with the stack traces in the full mode above, the jax related traces are pruned. Only the numpy frontend related errors are shown. A message is also displayed to inform that the traces are truncated and the instructions to switch it back to the full mode is included. In this case, the format of the error is Ivy error: backend name: backend function name: error message

>>> ivy.set_exception_trace_mode('frontend')
>>> ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
<stack trace is truncated to frontend specific files, call `ivy.set_exception_trace_mode('full')` to view the full trace>

During the handling of the above exception, another exception occurred:

<stack trace is truncated to frontend specific files, call `ivy.set_exception_trace_mode('full')` to view the full trace>
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 523, in _handle_numpy_out
    return fn(*args, **kwargs)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 396, in _outputs_to_numpy_arrays
    ret = fn(*args, **kwargs)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 352, in _inputs_to_ivy_arrays_np
    return fn(*ivy_args, **ivy_kwargs)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 453, in _from_zero_dim_arrays_to_scalar
    ret = fn(*args, **kwargs)
  File "/ivy/ivy/functional/frontends/numpy/logic/truth_value_testing.py", line 24, in all
    ret = ivy.all(a, axis=axis, keepdims=keepdims, out=out)

IvyIndexError: jax: all: axis 2 is out of bounds for an array of dimension 1

3. Ivy specific

This option displays only ivy-related stack traces. If compared to the different stack traces modes above, the ivy backend jax related traces (which were hidden in the frontend mode) are available again and the ivy frontend numpy related traces remain visible. However, the native jax traces remain hidden because they are not ivy-specific. A message is also displayed to inform that the traces are truncated and the instructions to switch it back to the full mode is included. The format of the error displayed is the same as the frontend mode above.

>>> ivy.set_exception_trace_mode('ivy')
>>> ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
<stack trace is truncated to ivy specific files, call `ivy.set_exception_trace_mode('full')` to view the full trace>
  File "/ivy/ivy/utils/exceptions.py", line 198, in _handle_exceptions
    return fn(*args, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 911, in _handle_nestable
    return fn(*args, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 392, in _handle_array_like_without_promotion
    return fn(*args, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 805, in _handle_out_argument
    return fn(*args, out=out, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 432, in _inputs_to_native_arrays
    return fn(*new_args, **new_kwargs)
  File "/ivy/ivy/func_wrapper.py", line 535, in _outputs_to_ivy_arrays
    ret = fn(*args, **kwargs)
  File "/ivy/ivy/func_wrapper.py", line 349, in _handle_array_function
    return fn(*args, **kwargs)
  File "/ivy/ivy/functional/backends/jax/utility.py", line 22, in all
    raise ivy.utils.exceptions.IvyIndexError(error)

During the handling of the above exception, another exception occurred:

<stack trace is truncated to ivy specific files, call `ivy.set_exception_trace_mode('full')` to view the full trace>
  File "/ivy/other_test.py", line 21, in <module>
    ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 523, in _handle_numpy_out
    return fn(*args, **kwargs)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 396, in _outputs_to_numpy_arrays
    ret = fn(*args, **kwargs)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 352, in _inputs_to_ivy_arrays_np
    return fn(*ivy_args, **ivy_kwargs)
  File "/ivy/ivy/functional/frontends/numpy/func_wrapper.py", line 453, in _from_zero_dim_arrays_to_scalar
    ret = fn(*args, **kwargs)
  File "/ivy/ivy/functional/frontends/numpy/logic/truth_value_testing.py", line 24, in all
    ret = ivy.all(a, axis=axis, keepdims=keepdims, out=out)
  File "/ivy/ivy/utils/exceptions.py", line 217, in _handle_exceptions
    raise ivy.utils.exceptions.IvyIndexError(

IvyIndexError: jax: all: axis 2 is out of bounds for an array of dimension 1

Ivy func_wrapper Pruning

Due to the wrapping operations in Ivy, a long list of less informative func_wrapper traces is often seen in the stack. Including all of these wrapper functions in the stack trace can be very unwieldy, thus they can be prevented entirely by setting ivy.set_show_func_wrapper_trace_mode(False). Examples are shown below to demonstrate the combination of this mode and the 3 different stack traces mode explained above.

1. Full

The func_wrapper related traces have been hidden. All other traces such as ivy-specific, frontend-related and the native traces remain visible. A message is displayed as well to the user so that they are aware of the pruning. The instructions to recover the func_wrapper traces are shown too.

>>> ivy.set_show_func_wrapper_trace_mode(False)
>>> ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
<func_wrapper.py stack trace is squashed, call `ivy.set_show_func_wrapper_trace_mode(True)` in order to view this>
  File "/ivy/ivy/utils/exceptions.py", line 198, in _handle_exceptions
    return fn(*args, **kwargs)
  File "/ivy/ivy/functional/backends/jax/utility.py", line 22, in all
    raise ivy.utils.exceptions.IvyIndexError(error)

During the handling of the above exception, another exception occurred:

<func_wrapper.py stack trace is squashed, call `ivy.set_show_func_wrapper_trace_mode(True)` in order to view this>
  File "/ivy/other_test.py", line 22, in <module>
    ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
  File "/ivy/ivy/functional/frontends/numpy/logic/truth_value_testing.py", line 24, in all
    ret = ivy.all(a, axis=axis, keepdims=keepdims, out=out)
  File "/ivy/ivy/utils/exceptions.py", line 217, in _handle_exceptions
    raise ivy.utils.exceptions.IvyIndexError(

IvyIndexError: jax: all: ValueError: axis 2 is out of bounds for an array of dimension 1

2. Frontend-only

In the frontend-only stack trace mode, the ivy backend wrapping traces were hidden but the frontend wrappers were still visible. By configuring the func wrapper trace mode, the frontend wrappers will also be hidden. This can be observed from the example below.

>>> ivy.set_exception_trace_mode('frontend')
>>> ivy.set_show_func_wrapper_trace_mode(False)
>>> ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
<stack trace is truncated to frontend specific files, call `ivy.set_exception_trace_mode('full')` to view the full trace>
<func_wrapper.py stack trace is squashed, call `ivy.set_show_func_wrapper_trace_mode(True)` in order to view this>

During the handling of the above exception, another exception occurred:

<stack trace is truncated to frontend specific files, call `ivy.set_exception_trace_mode('full')` to view the full trace>
<func_wrapper.py stack trace is squashed, call `ivy.set_show_func_wrapper_trace_mode(True)` in order to view this>
  File "/ivy/ivy/functional/frontends/numpy/logic/truth_value_testing.py", line 24, in all
    ret = ivy.all(a, axis=axis, keepdims=keepdims, out=out)

IvyIndexError: jax: all: axis 2 is out of bounds for an array of dimension 1

3. Ivy specific

As the wrappers occur in ivy itself, all backend and frontend wrappers remain visible in the ivy-specific mode. By hiding the func wrapper traces, the stack becomes cleaner and displays the ivy backend and frontend exception messages only.

>>> ivy.set_exception_trace_mode('frontend')
>>> ivy.set_show_func_wrapper_trace_mode(False)
>>> ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
<stack trace is truncated to ivy specific files, call `ivy.set_exception_trace_mode('full')` to view the full trace>
<func_wrapper.py stack trace is squashed, call `ivy.set_show_func_wrapper_trace_mode(True)` in order to view this>
  File "/ivy/ivy/utils/exceptions.py", line 198, in _handle_exceptions
    return fn(*args, **kwargs)
  File "/ivy/ivy/functional/backends/jax/utility.py", line 22, in all
    raise ivy.utils.exceptions.IvyIndexError(error)

During the handling of the above exception, another exception occurred:

<stack trace is truncated to ivy specific files, call `ivy.set_exception_trace_mode('full')` to view the full trace>
<func_wrapper.py stack trace is squashed, call `ivy.set_show_func_wrapper_trace_mode(True)` in order to view this>
  File "/ivy/other_test.py", line 22, in <module>
    ivy.functional.frontends.numpy.all(ivy.array([1,2,3]), axis=2)
  File "/ivy/ivy/functional/frontends/numpy/logic/truth_value_testing.py", line 24, in all
    ret = ivy.all(a, axis=axis, keepdims=keepdims, out=out)
  File "/ivy/ivy/utils/exceptions.py", line 217, in _handle_exceptions
    raise ivy.utils.exceptions.IvyIndexError(

IvyIndexError: jax: all: axis 2 is out of bounds for an array of dimension 1

@handle_exceptions Decorator

To ensure that all backend exceptions are caught properly, a decorator is used to handle functions in the try/except block.

# in ivy/utils/exceptions.py
def handle_exceptions(fn: Callable) -> Callable:
    @functools.wraps(fn)
    def _handle_exceptions(*args, **kwargs):
        try:
            return fn(*args, **kwargs)
        # Not to rethrow as IvyBackendException
        except IvyNotImplementedException as e:
            raise e
        except IvyError as e:
            _print_traceback_history()
            raise ivy.utils.exceptions.IvyError(fn.__name__, e, include_backend=True)
        except IvyBroadcastShapeError as e:
            _print_traceback_history()
            raise ivy.utils.exceptions.IvyBroadcastShapeError(
                fn.__name__, e, include_backend=True
            )
        except IvyDtypePromotionError as e:
            _print_traceback_history()
            raise ivy.utils.exceptions.IvyDtypePromotionError(
                fn.__name__, e, include_backend=True
            )
        except (IndexError, IvyIndexError) as e:
            _print_traceback_history()
            raise ivy.utils.exceptions.IvyIndexError(
                fn.__name__, e, include_backend=True
            )
        except (AttributeError, IvyAttributeError) as e:
            _print_traceback_history()
            raise ivy.utils.exceptions.IvyAttributeError(
                fn.__name__, e, include_backend=True
            )
        except (ValueError, IvyValueError) as e:
            _print_traceback_history()
            raise ivy.utils.exceptions.IvyValueError(
                fn.__name__, e, include_backend=True
            )
        except (Exception, IvyBackendException) as e:
            _print_traceback_history()
            raise ivy.utils.exceptions.IvyBackendException(
                fn.__name__, e, include_backend=True
            )

    _handle_exceptions.handle_exceptions = True
    return _handle_exceptions

The decorator is then added to each function for wrapping. Let’s look at an example of ivy.all().

# in ivy/functional/ivy/utility.py
@handle_exceptions
def all(
    x: Union[ivy.Array, ivy.NativeArray],
    /,
    *,
    axis: Optional[Union[int, Sequence[int]]] = None,
    keepdims: bool = False,
    out: Optional[ivy.Array] = None,
) -> ivy.Array:
    return ivy.current_backend(x).all(x, axis=axis, keepdims=keepdims, out=out)

When a backend throws an exception, it will be caught in the decorator and then the appropriate Error will be raised. This ensures that all exceptions are consistent.

Let’s look at the comparison of before and after adding the decorator.

without decorator

In NumPy,

>>> x = ivy.array([0,0,1])
>>> ivy.all(x, axis=2)
<error_stack>
numpy.AxisError: axis 2 is out of bounds for an array of dimension 1

In PyTorch,

>>> x = ivy.array([0,0,1])
>>> ivy.all(x, axis=2)
<error_stack>
IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 2)

The errors raised are different across backends, therefore confusing and inconsistent.

with decorator

In NumPy,

>>> x = ivy.array([0,0,1])
>>> ivy.all(x, axis=2)
<error_stack>
IvyIndexError: numpy: all: AxisError: axis 2 is out of bounds for an array of dimension 1

In PyTorch,

>>> x = ivy.array([0,0,1])
>>> ivy.all(x, axis=2)
<error_stack>
IvyIndexError: torch: all: IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 2)

The errors are unified into an IvyIndexError, with the current backend and function stated to provide clearer information. The message string is inherited from the native exception.

Consistency in Errors

For consistency, we make sure that the same type of Exception is raised for the same type of error regardless of the backend set. Let’s take an example of ivy.all() again. In Jax, ValueError is raised when the axis is out of bounds, and for Numpy, AxisError is raised. To unify the behaviour, we raise IvyIndexError for both cases.

In Numpy,

# in ivy/functional/backends/numpy/utility.py
def all(
    x: np.ndarray,
    /,
    *,
    axis: Optional[Union[int, Sequence[int]]] = None,
    keepdims: bool = False,
    out: Optional[np.ndarray] = None,
) -> np.ndarray:
    try:
        return np.asarray(np.all(x, axis=axis, keepdims=keepdims, out=out))
    except np.AxisError as e:
        raise ivy.utils.exceptions.IvyIndexError(error)

In Jax,

# in ivy/functional/backends/jax/utility.py
def all(
    x: JaxArray,
    /,
    *,
    axis: Optional[Union[int, Sequence[int]]] = None,
    keepdims: bool = False,
    out: Optional[JaxArray] = None,
) -> JaxArray:
    x = jnp.array(x, dtype="bool")
    try:
        return jnp.all(x, axis, keepdims=keepdims)
    except ValueError as error:
        raise ivy.utils.exceptions.IvyIndexError(error)

In both cases, IvyIndexError is raised, to make sure the same type of Exception is raised for this specific error.

Assertion Function

There are often conditions or limitations needed to ensure that a function is working correctly.

Inconsistency is observed such as some functions:

1. use assert for checks and throw AssertionError, or

2. use if/elif/else blocks and raise Exception, ValueError, etc.

To unify the behaviours, our policy is to use conditional blocks and raise IvyException whenever a check is required. Moreover, to reduce code redundancy, conditions which are commonly used are collected as helper functions with custom parameters in ivy/assertions.py. This allows them to be reused and promotes cleaner code.

Let’s look at an example!

Helper: check_less

# in ivy/utils/assertions.py
def check_less(x1, x2, allow_equal=False, message=""):
  # less_equal
  if allow_equal and ivy.any(x1 > x2):
      raise ivy.exceptions.IvyException(
          f"{x1} must be lesser than or equal to {x2}"
          if message == ""
          else message
      )
  # less
  elif not allow_equal and ivy.any(x1 >= x2):
      raise ivy.exceptions.IvyException(
          f"{x1} must be lesser than {x2}"
          if message == ""
          else message
      )

ivy.set_split_factor

# in ivy/functional/ivy/device.py
@handle_exceptions
def set_split_factor(
    factor: float,
    device: Union[ivy.Device, ivy.NativeDevice] = None,
    /,
) -> None:
    ivy.assertions.check_less(0, factor, allow_equal=True)
    global split_factors
    device = ivy.default(device, default_device())
    split_factors[device] = factor

Instead of coding a conditional block and raising an exception if the conditions are not met, a helper function is used to simplify the logic and increase code readability.

Round Up

This should have hopefully given you a good feel for how function wrapping is applied to functions in Ivy.

If you have any questions, please feel free to reach out on discord in the exception handling thread!

Video