jax-ml/jax · error · ValueError

cov: dtype must be a subclass of float or complex; got {dtyp

Error message

cov: dtype must be a subclass of float or complex; got {dtype=}

What it means

jnp.cov accepts an optional dtype parameter to control the result dtype, but it must be an inexact (float or complex) dtype since covariance is computed via floating-point averaging. If dtype is provided and not a subclass of np.inexact, a ValueError with the offending dtype is raised.

Source

Thrown at jax/_src/numpy/lax_numpy.py:9188

    >>> x = jax.random.normal(key, shape=(3, 100))
    >>> with jnp.printoptions(precision=2):
    ...   print(jnp.cov(x))
    [[0.9  0.03 0.1 ]
     [0.03 1.   0.01]
     [0.1  0.01 0.85]]
  """
  if y is not None:
    m, y = util.promote_args_inexact("cov", m, y)
    if y.ndim > 2:
      raise ValueError("y has more than 2 dimensions")
  else:
    m, = util.promote_args_inexact("cov", m)

  if m.ndim > 2:
    raise ValueError("m has more than 2 dimensions")  # same as numpy error

  if dtype is not None and not dtypes.issubdtype(dtype, np.inexact):
    raise ValueError(f"cov: dtype must be a subclass of float or complex; got {dtype=}")

  X = atleast_2d(m)
  if not rowvar and m.ndim != 1:
    X = X.T
  if X.shape[0] == 0:
    return array([]).reshape(0, 0)

  if y is not None:
    y_arr = atleast_2d(y)
    if not rowvar and y_arr.shape[0] != 1:
      y_arr = y_arr.T
    X = concatenate((X, y_arr), axis=0)
  if X.shape[1] == 0:
    cov_shape = () if X.shape[0] == 1 else (X.shape[0], X.shape[0])
    return array_creation.full(cov_shape, np.nan, dtype=X.dtype)

  if ddof is None:
    ddof = 1 if bias == 0 else 0

View on GitHub (pinned to 1e1c6a8fc0)

Solutions

1. Omit dtype to use the default float promotion
2. Use a float dtype: dtype=jnp.float32 or jnp.float64
3. Validate dtype with jnp.issubdtype(dtype, jnp.inexact) before passing

Example fix

// before
jnp.cov(m, dtype=jnp.int32)  # ValueError
// after
jnp.cov(m, dtype=jnp.float32)

if dtype is not None:
    assert jnp.issubdtype(dtype, jnp.inexact), 'cov dtype must be float/complex'
jnp.cov(m, dtype=dtype)

def is_inexact_dtype(d) -> bool:
    return d is None or jnp.issubdtype(d, jnp.inexact)

When it happens

Trigger: jnp.cov(m, dtype=jnp.int32) or dtype=float-with-integer-numpy-type like np.dtype(int64); any integer or boolean dtype request.

Common situations: Users assuming dtype works like astype on arbitrary arrays; passing a config-driven dtype variable that may be int in some code paths.

Related errors

• fweights must be integer.
• {} does not accept dtype {}. Accepted dtypes are subtypes of
• {name} does not accept dtype {dtype_to_string(aval.dtype)}.
• {} does not accept dtype {} at position {}. Accepted dtypes
• Input type is incompatible with `preferred_element_type`. Th

AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/5836579496f92f59. Report an issue: GitHub.