after new language finetune, s_pred gives NaN

[DrBrule]

I'm trying to make some new models for other languages, and I'm hitting this error on the generation step.

I've trained on a FT of English that I had, but with a new ASR model for my dataset and the multilingual PL-BERT along with associated configurations.

All the outputs look OK up until this step in the inference code:

    s_pred = sampler(noise = torch.randn((1, 256)).unsqueeze(1).to(device),
                embedding=bert_dur,
                embedding_scale=embedding_scale,
                features=ref_s, # reference from the same speaker as the embedding
                num_steps=diffusion_steps ).squeeze(1)

which only returns NaN using these models. Is there something unworkable about this configuration?