use the already existing MTP module of the model to implement the loss during training via next token prediction - [x] implement in qwen3-vl and qwen3.5 - [x] verify decoding speedup - [ ] test quick training run
use the already existing MTP module of the model to implement the loss during training via next token prediction