🔥 On 2× RTX 3060, "MTP didn't help" was a false alarm. Same Qwen3.8-27B Q4_K_M… · @QwenAIFans · TwKit

@QwenAIFans

🔥 On 2× RTX 3060, "MTP didn't help" was a false alarm. Same Qwen3.8-27B Q4_K_M: -sm layer ~15.7 tok/s → -sm tensor + draft-mtp = 42.9 tok/s at 128k (+173%). Catch: MTP only fires if you pass --spec-type draft-mtp. If your "with MTP" numbers match the baseline, the flag never ran. via marjimgu (HF) 👇 https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/discussions/77

曝光 156 · 评论 0 · 点赞 0 · 书签 1