Decode · tok/s
72
Prefill
—
tok/100W
30
SGLang + NVFP4 + DFlash2, 72 tok/s greedy median, 1M context.
The number is the repo's own line. Not reproduced on this desk.
Qwen3.8-27B
DGX Spark · NVFP4 · unknown
DGX Spark · NVFP4 · unknown
SGLang + NVFP4 + DFlash2, 72 tok/s greedy median, 1M context.
The number is the repo's own line. Not reproduced on this desk.