https://github.com/HawkBearPig/dgpp
I saw this tool mentioned on the NVIDIA forum in the DGX Spark board and decided to try it out.
It seems like it only supports a few models right now, but luckily qwen3.8-flash-next supports both 1 spark and 2 spark, so I set it up with 1 spark.
This is roughly the TPS I got.
It gets around 40 TPS alone, so I expect it to get around 20~26 for actual use.
When I used vllm before,
1 spark got around 20 TPS.
2 spark worked a little better than the new tool.
Theoretically, if I use 2 spark with this tool now, I could get around 70~80????
I'm not sure yet.
