baracuda-cutlass-kernels-sys
============================

The CUTLASS template instantiations under `kernels/` derive from NVIDIA's
CUTLASS examples (https://github.com/NVIDIA/cutlass), licensed under
BSD 3-Clause. CUTLASS itself is not redistributed — its headers are fetched
at build time by `baracuda-cutlass-sys`.

The crate's API shape, plan-based architecture, and v0 SKU set
(`{f16, bf16} × RCR × {sm80, sm90a}` with Identity epilogue, grouped GEMM
as a first-class op with `prepare`/`run` separation and caller-supplied
workspace) were specified by the Fuel ML library team during baracuda's
Phase 4 design review. The Bias epilogue was originally planned but
removed during round-1 review (advertised but unimplemented); per the
team's roadmap it returns "after grouped GEMM lands and a real caller
asks for it." Grouped GEMM landed in 4d as the team's high-priority
deliverable for MoE workloads.

The build-time kernel-compilation infrastructure is provided by
`baracuda-forge`, which itself derives from `cudaforge` by Guoqing Bao.
See `crates/baracuda-forge/NOTICE` for that lineage.
