DFlash `draft-dflash` acceptance stuck at ~0.15 with official z-lab Qwen3 drafters (8B + 4B, all target precisions, CPU + Vulkan) — net slow

用户在 llama.cpp b10048 (commit 2e1fd76) 上运行 llama-server 或 llama-cli ,使用 --spec-type draft-dflash 模式,配合官方 z-lab 提供的 DFlash 草稿模型(如 Qwen3-8B-DFlash-b16 或
![[Bug] Disabled Weaviate document vectors are not deleted by segment ID](https://www.chat-gpts.plus/wp-content/uploads/2026/07/39174-10c10356-768x403.jpg)







