full and remaining unmasked latent embeddings

[ZhaoLiruii]

Hello, I would like to know how this sentence in the paper is achieved: "In MDT, both the full and remaining unmasked latent embeddings are fed into the diffusion model during training."? I can't find this part in the code