Yeahh... so I pruned this like an idiot before. I totally removed the MTP module.

We now have num_nextn_predict_layers = 1 and the appended model.layers.61.* MTP module, including enorm, hnorm, eh_proj, embed_tokens, and shared_head tensors.

The routed expert banks have been pruned to experts 0..7, and layer 61 uses the same 8 routed experts as the main MoE layers.