Depending on what you mean by "production" you'll probably want to look at "real" serving implementations like HF TGI, vLLM, lmdeploy, Triton Inference Server (tensorrt-llm), etc. There are also more bespoke implementations for things like serving large numbers of LoRA adapters[0].
These are heavily optimized for more efficient memory usage, performance, and responsiveness when serving large numbers of concurrent requests/users in addition to things like model versioning/hot load/reload/etc, Prometheus metrics, things like that.
One major difference is at this level a lot of the more aggressive memory optimization techniques and support for CPU aren't even considered. Generally speaking you get GPTQ and possibly AWQ quantization + their optimizations + CUDA only. Their target users and their use cases are often using A100/H100 and just trying to need fewer of them. Support for lower VRAM cards, older CUDA compute architectures, etc come secondary to that (for the most part).