⚡️ Same Qwen3.8-27B GGUF. Same llama.cpp. 54% faster. Stock on one RTX 3090: 2… · @QwenAIFans · TwKit

@QwenAIFans

⚡️ Same Qwen3.8-27B GGUF. Same llama.cpp. 54% faster. Stock on one RTX 3090: 27.6 tok/s at ~16.9 GB. Add flash attention, q8_0 KV, and the built-in MTP head: 42.6 tok/s. Still ~16.9 GB. No new software. Just flags. via alexander-ollman 👇 https://alexander-ollman.github.io/qwen3.8-on-rtx3090/setup.html

曝光 700 · 评论 3 · 点赞 9 · 书签 12