
Walk the theta and observation blocks of an MDN's i.i.d. density, calling collect(theta_idx, obs_idx, lp) with the (n_theta_chunk, n_obs_chunk) tensor of log-densities
Source: R/likelihood.R
mdn_iid_blocks.RdThe MLP maps theta to the mixture parameters and never sees x, so
within one theta block the forward pass and the Cholesky assembly run once
and only the quadratic form is chunked over observations. theta itself is
blocked first (mdn_theta_chunk_size()), so max_batch bounds
mdn_mixture()'s output the same way cross_iid() bounds a flow's forward
pass, rather than only chunking the observation side (#240).