genuine question — how has fireworks or baseten or $reputable_inference_provider worked for your use cases? most production workload probably works fine with one of these and another set as fallback, at least so i think