本地跑 M3.1 Flash 等模型：调优显存与加载省下时间
在两张 AMD Strix Halo 加一张 RTX PRO 6000 上跑 DeepSeek V4.1 Flash、Qwen3.8、GLM-5.3 和 MiniMax M3.1，作者称本周收益多来自配置：显存频率停在 P1 损失约 4.5% 带宽、首 16 token 慢约 1.65 倍，专家加载流水线化后从 109
Same lesson here, many times over. We run DeepSeek V4.1 Flash, Qwen3.8, GLM-5.3 and MiniMax M3.1 on two AMD Strix Halo boxes plus a single RTX PRO 6000. Most of our wins this week were config/plumbing, not new hardware: • GPU memory clock sitting in P1 (13.4 GHz) instead of P0 (14 GHz) during decode: ~4.5% of bandwidth gone, and the first ~16 tokens ran ~1.65x slower in low power states. • Model weights mapped as uncached memory: every tile re-read the same weights from DRAM. Bench looked great, engine didn’t. • One event query per step on Windows WDDM forced early submission of queued copies: 1–2 ms lost per window. • A serial expert loader read 150 GB one expert at a time: load went 109 s → 62 s by pipelining reads (now disk-bound). And many more...