Strata - Qwen 3.8 Flash Next - boosting decoding speed Tip
英文标题：Strata - Qwen 3.8 Flash Next - boosting decoding speed Tip
I've got some nice boost by increasing --adapt-every value from calibrated 1 to 6. I tested on 4 consecutive prompts(Slovak language) in chat. Prompt 1-3 are simple questions, 4. is describing source code(20K tokens) Results: [prompt number / ~avg. output length] UD-Q4_K_XL: [1 / 996] 30.3 -> 36.1 t
打开原文相关 2 · 机会 0
中文
I've got some nice boost by increasing --adapt-every value from calibrated 1 to 6. I tested on 4 consecutive prompts(Slovak language) in chat. Prompt 1-3 are simple questions, 4. is describing source code(20K tokens) Results: [prompt number / ~avg. output length] UD-Q4_K_XL: [1 / 996] 30.3 -> 36.1 t
中文全文翻译正在补齐，可先查看下方英文原文。
English / 原文
I've got some nice boost by increasing --adapt-every value from calibrated 1 to 6. I tested on 4 consecutive prompts(Slovak language) in chat. Prompt 1-3 are simple questions, 4. is describing source code(20K tokens) Results: [prompt number / ~avg. output length] UD-Q4_K_XL: [1 / 996] 30.3 -> 36.1 t