Really basic question here. I’ve read in a few places that PyTorch seems to be geared towards “development” and not production. Is there some reason why it is not suitable for production?
The claims on the site, “The memory usage in PyTorch is extremely efficient”, “PyTorch is quite fast – whether you run small or large neural networks.” etc. seem to be all great benefits for running in production (and I suppose in development too)
if you dont mind running python in production, then pytorch is ready for production.
However, industrial-grade production deployments prefer C/C++/Java (for example mobile).
Hence our tag of “pytorch is not suited for production”, tied to our deep entanglement with python.
We also tend to always prioritize research flexibility over freezing specs / a more structured static model approach which is also suited for production.
Is PyTorch claiming performance improvements over libraries like Tensorflow even without C or cython?
we dont claim performance improvements over any other library, we leave that exercise to the community.
Forgive me to bring back this discussion again; but, has there been new things in PyTorch that enable production?
Especially that the new pytorch.org headline says:
“PyTorch
FROM
RESEARCH TO
PRODUCTION”
??
It is still Pythonic, so, what does PRODUCTION refer to in the headline?
Coming with the 1.0 preview, you can now use the just-in-time compiler (torch.jit ) to load your model in C++. Have a look at this example.
Also using libtorch you have now a C++ API. Here is an end-to-end example. 
Soumith might have more information for production usage. 