DeepSeek v4 PRO on DwarfStar: Peaks at 50 t/s with Dynamic VRAM/RAM

antirez · x · 2026-08-18

DeepSeek v4 PRO tested on DwarfStar running on a DGX Station, achieving 50 tokens/s peak without DSpark. Key optimizations include dynamic VRAM/RAM allocation for experts based on historical stats, 4k t/s prefill for large chunks, and layer-major streaming. The author notes that mixed RAM/VRAM inference requires careful handling.

Original post →

More from Infra

Infra channel →