⚡️ Same Qwen3.8 MTP draft. Three backends. Totally different story. On a 3090 … · @QwenAIFans · TwKit

@QwenAIFans

⚡️ Same Qwen3.8 MTP draft. Three backends. Totally different story. On a 3090 with Q4_K_M, --spec-draft-n-max 1 jumps you from 41.2 → 56.1 tok/s (+36%). Same files, same ~75% draft acceptance on Vulkan: only +3%. On Metal (M1 Max): −11%. If you're on a Mac, leave the draft off. On CUDA, use n-max 1, not the default of 3 (that one leaves a third of the gain sitting there). Caveat: one card / one machine each. Measure yours before you fight about it online. via bowmanslayer 👇 https://huggingface.co/bowmanslayer/Qwen3.8-27B-GGUF

曝光 107 · 评论 0 · 点赞 0 · 书签 0