tracel-ai/burn · error

todo!("rfft is not supported for ndarray")

Error message

todo!("rfft is not supported for ndarray")

What it means

The ndarray backend's rfft (crates/burn-ndarray/src/ops/module.rs:406) is a stub that always panics with todo!("rfft is not supported for ndarray"). Real-valued FFT is simply not implemented for the CPU ndarray backend.

Source

Thrown at crates/burn-ndarray/src/ops/module.rs:406

    }

    fn attention(
        query: FloatTensor<Self>,
        key: FloatTensor<Self>,
        value: FloatTensor<Self>,
        mask: Option<burn_backend::tensor::BoolTensor<Self>>,
        attn_bias: Option<FloatTensor<Self>>,
        options: AttentionModuleOptions,
    ) -> FloatTensor<Self> {
        attention_fallback::<Self>(query, key, value, mask, attn_bias, options)
    }

    fn rfft(
        _signal: FloatTensor<Self>,
        _dim: usize,
        _n: Option<usize>,
    ) -> (FloatTensor<Self>, FloatTensor<Self>) {
        todo!("rfft is not supported for ndarray")
    }

    fn irfft(
        _spectrum_re: FloatTensor<Self>,
        _spectrum_im: FloatTensor<Self>,
        _dim: usize,
        _n: Option<usize>,
    ) -> FloatTensor<Self> {
        todo!("irfft is not supported for ndarray")
    }
}

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Switch to a backend that implements rfft (e.g. the candle or cubecl-backed backends).
2. Compute the FFT outside burn (e.g. via the rustfft crate) and load the result back as a tensor.
3. Implement rfft for burn-ndarray using rustfft and contribute it upstream.

Example fix

// before
let (re, im) = signal.rfft(1, None); // panics on ndarray backend
// after: use a supported backend
#[cfg(feature = "candle")]
let (re, im) = signal_candle.rfft(1, None);

// guard before calling rfft on ndarray backend
if std::any::TypeId::of::<B>() != supported_fft_backend() { /* route to rustfft fallback */ }

When it happens

Trigger: Calling Tensor::rfft (any dim/n) on a tensor whose backend is burn-ndarray.

Common situations: Running audio/signal-processing models (spectrogram, STFT) on CPU via ndarray; the same code works on a GPU backend that implements FFT.

Related errors

• todo!("irfft is not supported for ndarray")
• todo!("grid_sample_2d with {:?} mode is not implemented", op
• rfft: unsupported dtype {:?}
• irfft: unsupported dtype {:?}
• Dim not supported {ndims}

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/a181c465d54c1259. Report an issue: GitHub.