Is Pytorch still unable to have some smaller version of it for inference or JavaScript solution?

I have been deploying Pytorch model on Web and mobile. However, I often need some other framework, such as ONNX to make it runs in the browser.
I know we have jit or torchserve. However, do we have other suggestion beside them? And what is the file size requirement when installing the dependency? Let’s say I want to deploy a Web application. I can not install all kinds of Pytorch library in the server.
I say torchserve does not work. For example, you can’t use a torchserve if you want to build a NLP transformer model NPC installed at client’s computer because it requires a HTTP backend serve.

There isn’t a PyTorch for the browser or javascript version at the moment. What kind of issues do you run into with onnx.js? Is it the transformation from pytorch to onnx that causes issues or something else? It would also be great to hear what you use ML in the browser for if you can share.