vLLM MonoKernel Slashes AMD Decode Latency

This is a meaningful low-concurrency inference win, but “shipped” overstates the current status: the GLM-5.2 implementation remains a draft PR, while CUDA support is explicitly deferred.

• –Persistent kernels collapse most of a decoder layer into one launch, attacking launch overhead and repeated weight reads.
• –The gain is strongest for latency-sensitive agents and chat workloads; performance improvements shrink as batch size increases.
• –The RFC centralizes routing, cache binding, safety checks, health monitoring, and fail-stop behavior so each new model needs less bespoke runtime code.
• –Unsupported configurations are refused rather than silently falling back, which improves correctness but limits plug-and-play deployment.
• –The ROCm-first design gives AMD a visible software lead in this specific vLLM optimization, while CUDA support still needs a co-author. ([RFC](https://github.com/vllm-project/vllm/issues/60904), [GLM-5.2 PR](https://github.com/vllm-project/vllm/pull/60889))