Qwen3.6 27B on a 5090, 6.4k sample tok/s distribution after tuning MTP/cache settings
英文标题：Qwen3.6 27B on a 5090, 6.4k sample tok/s distribution after tuning MTP/cache settings
Spent a while tuning llama.cpp for Qwen3.6 27B on a 9800X3D / 64GB / 5090 box and wanted to share the real distribution instead of just a headline number, since averages hide a lot. Ran with q8 KV cache, 192k context, MTP draft=10, spec-draft-p-min=0.5, batch/ubatch 512. Logged 6,454 samples a
打开原文相关 3 · 机会 0
中文
Spent a while tuning llama.cpp for Qwen3.6 27B on a 9800X3D / 64GB / 5090 box and wanted to share the real distribution instead of just a headline number, since averages hide a lot. Ran with q8 KV cache, 192k context, MTP draft=10, spec-draft-p-min=0.5, batch/ubatch 512. Logged 6,454 samples a
中文全文翻译正在补齐，可先查看下方英文原文。
English / 原文
Spent a while tuning llama.cpp for Qwen3.6 27B on a 9800X3D / 64GB / 5090 box and wanted to share the real distribution instead of just a headline number, since averages hide a lot. Ran with q8 KV cache, 192k context, MTP draft=10, spec-draft-p-min=0.5, batch/ubatch 512. Logged 6,454 samples a