Hello! Does MS-SWIFT support SFT training of the Draft model for Multi-token-prediction feature? With full model training or Lora? How to quantize the model to FP8-Dynamic (VLLM compatible) preserving Draft model quality? Thanks