Recipe is MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks, ring transport is FujitsuPolycom/sparkring, tuned profile from knapcio's TP4 fork.
In deepseek harness I get 70-100 tok/s single decode which is approaching frontier API speeds.
100 tok/s with TP4 deepseek flash v4.1 1017
No switch just ring mesh.
Infinite unlocked opus 4.8 only cost me 20k, bonkers.
Let me know if there's a way to go faster without losing any quality.
Recipe is MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks, ring transport is FujitsuPolycom/sparkring, tuned profile from knapcio's TP4 fork.
No switch
In deepseek harness I get 70-100 tok/s