The bear can dance: Qwen 3.8 27B on one 3090 for 3 weeks
英文标题：The bear can dance: Qwen 3.8 27B on one 3090 for 3 weeks
TL;DR: Local agent loop, ~21 days, one RTX 3090. Task was pretty much "build a CUDA inference engine for optimized for yourself on this GPU arch." Got working kernels and benches, not a win over llama.cpp. ~12 human messages. Compaction ate ~83 hours. Old joke: you don’t criticize how well the bear
打开原文相关 5 · 机会 1
中文
TL;DR: Local agent loop, ~21 days, one RTX 3090. Task was pretty much "build a CUDA inference engine for optimized for yourself on this GPU arch." Got working kernels and benches, not a win over llama.cpp. ~12 human messages. Compaction ate ~83 hours. Old joke: you don’t criticize how well the bear
中文全文翻译正在补齐，可先查看下方英文原文。
English / 原文
TL;DR: Local agent loop, ~21 days, one RTX 3090. Task was pretty much "build a CUDA inference engine for optimized for yourself on this GPU arch." Got working kernels and benches, not a win over llama.cpp. ~12 human messages. Compaction ate ~83 hours. Old joke: you don’t criticize how well the bear