fftconv1d_bhl_w_reshape_chunked#

fftconv1d_bhl_w_reshape_chunked(
x,
kernel,
shortcut=None,
chunk_size=None,
)#

Channel-chunked 1D non-causal convolution via subq_ops for BLH inputs [B, L, H].

Parameters:

  • x (Tensor) – Input tensor [B, L, H].

  • kernel (Tensor) – Kernel tensor [1, K, H]; see fftconv1d_bhl().

  • shortcut (Tensor | None) – Optional per-channel scale [H].

  • chunk_size (int | None) – Channels per kernel call; None means 128.

Returns:

Output tensor [B, L, H] in x.dtype.

Return type:

Tensor