causal_conv1d

causal_conv1d(

x,

weight,

bias=None,

activation='identity',

channel_last=False,

)

Depthwise causal 1D conv via the subq_ops CUDA kernel.

Parameters:

• x (Tensor) – Input tensor [B, C, L], or [B, L, C] when channel_last=True.

• weight (Tensor) – Depthwise weight [C, K], or [K, C] when channel_last=True.

• bias (Tensor | None) – Optional per-channel bias [C].

• activation (str) – "identity" (default) or "silu".

• channel_last (bool) – Select the channel-last (NWH) kernel; the input layout and the weight layout follow it. This path is safe under torch.compile.

Returns:

Output tensor with the same shape as x.

Return type:

Tensor