2026-11-05 · America/Los_Angeles · 社区动态 · #18

436 tok/s per user on DeepSeek V4.1 Flash 1017. What are you getting?

Disclosure: I work at Sciforium. We run inference on our own AMD GPUs. Our engineers fused MoE, attention, and all reduce into one megakernel for DeepSeek V4.1 Flash 1017. On 4x MI325X, reaching 436 output tok/s per user on MI325X—2.27× faster on our code workload with the same weights and no accuracy loss and 487 output tok/s per user with our MI355X serving stack. Who’s running this model? Curious what you’re seeing: What speeds are you getting on this model, and on what hardware? Anyone else tried megakernels on MoE? Did it hold up with more users? What serving optimizations have made the…

热度 43.9 / 100；排名与评分保留该期记录。