GreenferenceAI/greenference-miner ? reverse-engineered prompt

Build me the miner side app for Greenference. I want a Python service that can accept model workloads, prepare the right runtime for them, and run local inference for a default GPU based text generation setup. It should keep miner specific runtime records on disk, cache published build artifacts so repeated work is faster, and support both normal one shot responses and streaming responses.

When the machine allows it, run model runtimes as local processes with socket access. If the environment is restricted, support a fallback mode that behaves the same way from the workload point of view, but only turns on when an environment flag is explicitly enabled for dev or limited setups.

Please include startup recovery, cleanup of stale runtimes, and a few debug endpoints so I can inspect active runtimes, failed runtimes, and runtime details. I also want tests and basic infrastructure so it is actually deployable and maintainable. Keep it practical and look up current docs online if you need to.