ParentFull threadshay_ker·curious that they are doing speculative decoding and not baking MTP into the model, like Nemotronhttps://docs.nvidia.com/megatron-core/developer-guide/0.15.0...View on HN