Lemonade Server adds experimental vLLM ROCm backend

This looks like Lemonade widening from “easy local GGUF runtime” into “bring whatever backend fits the workload.” That’s the right move if the team wants AMD users to have a credible alternative when vLLM’s throughput and day-0 model support matter more than simplicity.

• –The self-contained ROCm bundle lowers setup friction, which is the main barrier for backend experimentation on AMD systems
• –Lemonade is clearly testing where vLLM fits versus llama.cpp: better for concurrency and newer transformer support, but still rough around the edges
• –The initial validation focus on gfx1151 and gfx1150 suggests Strix Halo/Strix Point are the first-class targets, with broader GPU coverage still maturing
• –Community feedback matters here because the product decision is bigger than one backend: it is about whether Lemonade becomes an orchestrator for multiple inference engines