jax-ml/jax · error · TypeError

Unexpected input type for array

Error message

Unexpected input type for array: {type(object)}

What it means

jnp.array's final fallback: after checking Tracer, np.ndarray, scalar types, lists/tuples (flattened earlier), and buffer-protocol objects, the input type is not convertible to an array and raises TypeError with the offending type name.

Solutions

  1. Extract the numeric data first (e.g. obj.array, np.asarray(obj))
  2. Implement __jax_array__ on your custom class so jnp.array can consume it
  3. For sets/dicts/iterables, convert to a list of leaves first

Example fix

# before
a = jnp.array(my_dataclass)
# after
a = jnp.array(my_dataclass.values)  # or implement __jax_array__ on the class
Defensive patterns

Strategy: type-guard

Validate before calling

def convert_or_fail(obj):
    import numpy as np
    return obj if hasattr(obj, '__jax_array__') else np.asarray(obj)

Type guard

def is_jax_convertible(obj) -> bool:
    import numpy as np
    return hasattr(obj, '__jax_array__') or isinstance(obj, (np.ndarray, list, tuple, int, float, bool, complex, bytes, memoryview))

Try / catch

null

Prevention

When it happens

Trigger: jnp.array(some_arbitrary_object) where object is e.g. a dict, custom class without __jax_array__/__buffer__, file handle, or string under a non-string dtype path.

Common situations: Passing a Python object wrapper (e.g. a dataclass or config object) instead of its .array field; assuming arbitrary iterables like sets or dicts convert like they roughly do in NumPy.

Related errors


AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/9d2bb5737b5e50d7. Report an issue: GitHub.

Appendix: source

Thrown at jax/_src/numpy/array_constructors.py:337

    if object:
      arrs = (array(elt, dtype=dtype, copy=False) for elt in object)
      arrays_out = [lax.expand_dims(arr, [0]) for arr in arrs]
      # lax.concatenate can be slow to compile for wide concatenations, so form a
      # tree of concatenations as a workaround especially for op-by-op mode.
      # (https://github.com/jax-ml/jax/issues/653).
      k = 16
      while len(arrays_out) > k:
        arrays_out = [lax.concatenate(arrays_out[i:i+k], 0)
                      for i in range(0, len(arrays_out), k)]
      out = lax.concatenate(arrays_out, 0)
    else:
      out = np.array([], dtype=dtype)
  elif _supports_buffer_protocol(object):
    object = memoryview(object)
    # TODO(jakevdp): update this once we support NumPy 2.0 semantics for the copy arg.
    out = np.array(object) if copy else np.asarray(object)
  else:
    raise TypeError(f"Unexpected input type for array: {type(object)}")
  out_array: Array = lax._convert_element_type(
      out, dtype, weak_type=weak_type, sharding=sharding)
  if ndmin > np.ndim(out_array):
    out_array = lax.expand_dims(out_array, range(ndmin - np.ndim(out_array)))
  return out_array


def _get_platform(
    device_or_sharding: xc.Device | Sharding | None | str) -> str:
  """Get device_or_sharding platform or look up config.default_device.value."""
  if isinstance(device_or_sharding, xc.Device):
    return device_or_sharding.platform
  elif isinstance(device_or_sharding, Sharding):
    return list(device_or_sharding.device_set)[0].platform
  elif isinstance(device_or_sharding, str):
    return device_or_sharding
  elif device_or_sharding is None:
    if config.default_device.value is None:

View on GitHub (pinned to 1e1c6a8fc0)