Batching for Different Model Architectures

Adapt batching strategies for encoder-decoder models, diffusion models, and other architectures with unique computational patterns.