Are you using the default sampler? There may be a slowdown due to Replace `batched_convolution` by PyTensor native implementation by ricardoV94 · Pull Request #1583 · pymc-labs/pymc-marketing · GitHub for small datasets, specially on the default C-backend.
We’re trying to fix the gap in Faster convolve1d in numba backend by ricardoV94 · Pull Request #1378 · pymc-devs/pytensor · GitHub for the Numba backend. The C-backend may require a bit more work