jax-ml/jax · error · ValueError

only real valued inputs supported for rfft

Error message

only real valued inputs supported for rfft

What it means

jax.lax.fft with type RFFT requires a real-valued (floating-point) input; it converts to inexact dtype and, if the input is already complex (np.iscomplexobj), raises this ValueError. Use plain 'fft' for complex inputs instead.

Solutions

  1. Use 'fft' (or jnp.fft.fft) for complex inputs
  2. Take the real part first (z.real) if you genuinely want an rfft of real-valued data embedded in complex
  3. Use irfft to go from complex spectrum back to real signal

Example fix

# before
y = lax.fft(jnp.exp(1j * x), 'rfft', (64,))
# after
y = lax.fft(jnp.exp(1j * x), 'fft', (64,))
# or, if real data intended: y = lax.fft(x, 'rfft', (64,))
Defensive patterns

Strategy: type-guard

Validate before calling

if np.iscomplexobj(x):
    fft_type = 'fft'  # rfft requires real input
assert not (fft_type in ('rfft', lax.FftType.RFFT) and np.iscomplexobj(x))

Type guard

def real_input(x) -> bool:
    return not jnp.issubdtype(x.dtype, jnp.complexfloating)

Prevention

When it happens

Trigger: Calling jax.lax.fft(z, 'rfft', (n,)) where z has a complex dtype, or jnp.fft.rfft on complex data via wrappers that hit lax.fft.

Common situations: Data pipelines where an earlier op (e.g. fft then rfft in a spectrogram chain) leaves complex arrays; forgetting that rfft's inverse counterpart is irfft, not rfft.

Related errors


AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/642934d337b7d011. Report an issue: GitHub.

Appendix: source

Thrown at jax/_src/lax/fft.py:78

  elif s in ("rfft", "RFFT"):
    return FftType.RFFT
  elif s in ("irfft", "IRFFT"):
    return FftType.IRFFT
  else:
    raise ValueError(f"Unknown FFT type '{s}'")

@jit(static_argnums=(1, 2))
def fft(x, fft_type: FftType | str, fft_lengths: Sequence[int]):
  if isinstance(fft_type, str):
    typ = _str_to_fft_type(fft_type)
  elif isinstance(fft_type, FftType):
    typ = fft_type
  else:
    raise TypeError(f"Unknown FFT type value '{fft_type}'")

  if typ == FftType.RFFT:
    if np.iscomplexobj(x):
      raise ValueError("only real valued inputs supported for rfft")
    x = lax.convert_element_type(x, dtypes.to_inexact_dtype(dtypes.dtype(x)))
  else:
    x = lax.convert_element_type(x, dtypes.to_complex_dtype(dtypes.dtype(x)))
  if len(fft_lengths) == 0:
    # XLA FFT doesn't support 0-rank.
    return x
  fft_lengths = tuple(fft_lengths)
  return fft_p.bind(x, fft_type=typ, fft_lengths=fft_lengths)

def _fft_impl(x, fft_type, fft_lengths):
  return dispatch.apply_primitive(fft_p, x, fft_type=fft_type, fft_lengths=fft_lengths)

_complex_dtype = lambda dtype: (np.zeros((), dtype) + np.zeros((), np.complex64)).dtype
_real_dtype = lambda dtype: np.finfo(dtype).dtype

def fft_abstract_eval(x, fft_type, fft_lengths):
  if len(fft_lengths) > x.ndim:
    raise ValueError(f"FFT input shape {x.shape} must have at least as many "

View on GitHub (pinned to 1e1c6a8fc0)