Corey James- 3kernel libraries in the release built on it
- 2attention needs the stock kernels did not meet
- 2candidate changes rejected and kept as negative results
Extending the vendor's templates
ActiveKV needed two things from attention on the Intel Arc Pro GPU that stock kernels do not provide, and I decided to extend Intel's CUTLASS and CuTe port for SYCL to get them. The library comes from the CUTLASS 4.2.1 lineage.
The vendor's templates already encode the tile, copy and MMA structure for Xe2, so I added only what ActiveKV needed.
The one statistic exact merging needs
The first need was the log-sum-exp statistic from paged decode. It is the one statistic that makes exact merging across memory tiers possible, so I had it added to the paged-attention path and nothing more.
The second need was attention operators that read history from the staging working set and from the compressed representation. Those were derived from the vendor's templates as specialised operators.
Context beyond VRAM covers why the exact merge matters.
Gated on bit-identical output
Every kernel change was built inside the vendor's own image toolchain, so the ABI matched what vLLM loads.
No kernel patch ships unless its outputs match the unpatched path bit for bit.
Bringing a serving stack up on new silicon covers the wider kernel programme. CUTLASS itself is NVIDIA's, and the SYCL port is Intel's.