How to Speed Up Qwen3.8-27B in llama.cpp (NVIDIA's Draft Model May Lose to MTP)
DEV Community
How to Speed Up Qwen3.8-27B in llama.cpp (NVIDIA's Draft Model May Lose to MTP)
The 4B draft accepts 46% more tokens than MTP but is 10x bigger. VRAM math for 16 and 24 GB cards, plus the llama.cpp and vLLM commands.
0 comments
No comments yet.