Primary sourceelvis
View sourceEvent ·Multi-source
AMD MI355X 上跑出 1,129 tok/s:sciforium 分享 DeepSeek-V4.1-Flash 推测解码实践
sciforium 在 4 张 AMD MI355X GPU 上将 DeepSeek-V4.1-Flash 的输出速度做到 1,129 tok/s(P90),且使用的是未经改动的原始权重,基于真实生产流量而非测试环境。核心做法在推测解码的 drafter:团队用客户流量对 drafter 做微调,让更长的草拟 token 序列能在验证环节存活,最终让 7-token 草拟长度变得划算。文章还公开了哪些优化手段没有奏效,这类负结果在工程分享中较少见。此前外界曾质疑 AMD GPU 的推理性能是否跟得上,这次纪录是在 AMD 硬件上以生产流量跑出来的,对关心非 NVIDIA 推理栈的工程团队有直接参考价值。
Current read
sciforium 用客户流量微调推测解码的 drafter,让 7-token 草拟通过验证,DeepSeek-V4.1-Flash 在 4 张 AMD MI355X 上跑到 1,129 tok/s,权重还是原版没动过。做推理优化的工程师值得看,他们连失败的尝试都写出来了。
11 sources46° AI heatLast updated 2026/11/17 19:56:16
Evidence
11 sourcesRelated source 1vLLM
View sourceRelated source 2arXiv: DeepSeek
View sourceRelated source 3Nathan Lambert
View sourceRelated source 4NVIDIA AI
View sourceRelated source 5TestingCatalog
View sourceRelated source 6爱范儿
View sourceRelated source 7Detik Inet
View sourceRelated source 8IT之家
View sourceRelated source 9CoreWeave
View sourceRelated source 10Geek
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.