Repository navigation

fix(ArcLab/ArcQuant): make prefill measurable, then ship the two kernels it proves were switched off - #196

Merged

heydryft merged 3 commits into

release/openrouter-readyfrom

agent/prefill-honest-bench-v2

Aug 21, 2026

Merged

fix(ArcLab/ArcQuant): make prefill measurable, then ship the two kernels it proves were switched off#196
heydryft merged 3 commits into
release/openrouter-readyfrom
agent/prefill-honest-bench-v2

Commits on Aug 21, 2026
- fix(ArcLab/ArcInfer): prefill measured 0.000 tok/s and aborted at teardown
  
  Nirupam Bhowmick
  authored and
  heydryft
  committed
  cb448f6
- test(ArcLab): print the grouped-GEMM launch counters beside the results
  
  Nirupam Bhowmick
  authored and
  heydryft
  committed
  693a6de
- perf(ArcQuant): ship the tuned grouped-GEMM kernel as the default (+19.5% prefill)
  
  Nirupam Bhowmick
  authored and
  heydryft
  committed
  a2937ff