fftconv1d_fp32_bhl

fftconv1d_fp32_bhl(x, kernel, shortcut=None)

1D FFT convolution with optional shortcut, for inputs with layout (batch, hidden, length).

When shortcut provided, then the output is given by shortcut(x) + conv(x, kernel).

Accepts any input dtype. Internally casts x and kernel to float32 for numerical stability and returns the result in the original dtype of x.

Parameters:

• x (torch.Tensor) – Input tensor of shape (batch_size, hidden_dim, seq_len).

• kernel (torch.Tensor) – Kernel tensor of shape (1, hidden_dim, kernel_len).

• shortcut (torch.Tensor | None, optional) – Optional shortcut tensor of shape (hidden_dim). Defaults to None.

Returns:

Output tensor of shape (batch_size, hidden_dim, seq_len), in the original dtype of x.

Return type:

torch.Tensor