Different kernels within a same batch for a causal convolution

Question

Is there a way in PyTorch (1.6) to compute dynamic causal convolution: where the kernel size is different per sample? One could use the argument group of F.conv1d.