Operator optimization: cross-engine op injection (HyperLoom)

[jiejingzhangamd]

From the 2026 Q3 roadmap.

Today: HyperLoom optimized op injection.
Goal: Cross-engine op/kernel injection using HyperLoom, optimized per new model and user workload.

• Out-of-tree vLLM op-injection plugin (attention + MoE seams, no fork) + op-agnostic optimize-loop scaffold + a first decode MoE kernel: vLLM op-injection plugin: custom Attention + MoE kernels via out-of-tree entry_points (no fork) #40 / PR feat(vllm): op-injection plugin + decode MoE experts kernel (1.28x at small M) #43. Remaining: SGLang/ATOM injection, HyperLoom-driven kernel selection.