Repository navigation

[Kernels] Add generic FP32-accumulating RMSNorm with optional residual fusion - #167

Draft

bjf-frz wants to merge 5 commits into

InferMatrixLab:mainfrom

bjf-frz:feat/ouro-rmsnorm-cuda

Draft

[Kernels] Add generic FP32-accumulating RMSNorm with optional residual fusion#167
bjf-frz wants to merge 5 commits into
InferMatrixLab:mainfrom
bjf-frz:feat/ouro-rmsnorm-cuda

Commits on Nov 3, 2026
- feat(kernels): add Ouro-compatible fused RMSNorm
  bjf-frz
  committed
  4f3ce8e
- docs: illustrate RMSNorm graph execution
  bjf-frz
  committed
  59f964c
Commits on Nov 10, 2026
- Unify RMSNorm CUDA arithmetic and validate against vLLM
  
  bjf-frz
  and
  codex
  committed
  566aef8
- Use workload-based RMSNorm launch and cache policy
  bjf-frz
  committed
  50a7ce1
- Size RMSNorm input cache from complete vectors
  bjf-frz
  committed
  824cd38