SYLIN

Lucebox

by Luce · Runtime · Text & code

C++/ROCm HIP inference server, no Python, vLLM or llama.cpp, builds from source and serves DeepSeek V4 Flash locally on port 8000

Evidence

  1. Jul 27
    Claims 25 tok/s autoregressive vs 32 tok/s with DSpark speculative decoding (verification cap 4), a 26% gain, on the same APU
  2. Jul 27
    Live run measured 22 tok/s on the baseline test, below the 25 tok/s autoregressive figure quoted earlier; the 32 tok/s result was not shown live

Alternatives