asf0/libstrix provides HIP kernels for AMD Strix Halo (gfx1151, Radeon 8060S) along with a plugin that uses them in upstream vLLM. The kernels are applied through the plugin without modifying vLLM sources, and they do not activate on other graphics devices. The repository description states that in long-context scenarios prefill is up to 31x faster and decode about 2x faster.
Each kernel is compiled for a single operation geometry. The strix.select call is a feature for checking which geometries a kernel supports; if the input geometry or dtype conditions do not match, that kernel is not used and vLLM's default path is used instead. So the design is not to apply kernels to every input, but to select them only when the support conditions are met. The README documents a kernel registry and the selection feature, as well as test code and benchmark configuration. The repository's primary implementation language is Python and its license is MIT.
The performance figures presented have not been independently reproduced using only the repository's own materials, so it is difficult to treat them as conclusive evidence of actual performance. The applicable target is also limited to Strix Halo and the supported geometry and dtype conditions. The performance claims should therefore be treated as repository-side results that need further verification, while also taking into account the narrow scope of applicability.