nn.Conv1d: input [B, L, C_in], weight [C_out, K, C_in/groups] → [B, L', C_out]. Depthwise (groups == channels) for the gated-DeltaNet causal conv; the model left-pads with the conv-state cache, so padding is 0 here.
nn.Conv1d: input [B, L, C_in], weight [C_out, K, C_in/groups] → [B, L', C_out]. Depthwise (groups == channels) for the gated-DeltaNet causal conv; the model left-pads with the conv-state cache, so padding is 0 here.