New
luce-org/lucebox f3cc3ec — Reducing kernel launch count in Qwen split decoding
Commits adding a late-expert option that reduces per-layer kernel launches on the target GPU from 45.4 to about 36 in two-GPU split decoding, plus additional CUDA·HIP fusions.