Just wanted to share that I was looking for an optimal way to run Ornith 35B in FP8 with E4M3 and MTP with vLLM but there was no out-of-the-box model with MTP drafter support. So I grafted this new model! It's 18% faster than without MTP and the drafter acceptance rate is not bad (70% on avg). It should run on any RTX based setup > 80GB VRAM with full context window 256k. Might also do well on Unified Memory Systems like GB10 (for this - use my script and graft the MTP model into a target N...