MoE offload on a 3090: prefill 345 -> 466 t/s by tuning cudaMemcpy (Qwen3.6 35B-A3B)
🌐 链接: https://www.reddit.com/r/LocalLLaMA/comments/1xx1owx/moe_offload_on_a_3090_prefill_345_466_ts_by/
🏷️ 分组: Reddit|AI·自托管·VPS
🕒 时间: 2026-11-16 01:43:31
🌐 链接: https://www.reddit.com/r/LocalLLaMA/comments/1xx1owx/moe_offload_on_a_3090_prefill_345_466_ts_by/
🏷️ 分组: Reddit|AI·自托管·VPS
🕒 时间: 2026-11-16 01:43:31