scvi-tools 1.4.2 on CPU: same scvi.settings.seed gives bitwise-identical SCVI/SCANVI latents

To separate seed effects from run-to-run nondeterminism when benchmarking scVI, I retrained SCVI (2 layers, 30 latent, NB, default epochs) and SCANVI.from_scvi_model (20 epochs) on the same 16k-cell, 2000-gene data with the same scvi.settings.seed, in a fresh process on CPU (torch 2.15 CPU build, default multithreading, accelerator='cpu'). get_latent_representation() was bitwise identical between the two runs for both models (np.array_equal True). So on CPU, the spread between seeds can be read as a pure seed effect. GPU was not tested and cuDNN kernels may not behave the same way. Different seeds gave clearly different latents and moved scib-metrics Total by SD ~0.005.