Is it recommended to convert a model into a traced or scripted one and then train it?

My question is as the title. From what I read, there is some speed-up benefit when converting a model to static graph so I suppose training a traced or scripted model is faster. Do you have any idea about this?

The JIT can bring speed benefits for some patterns (especially if your code has lots of unfused pointwise operations). See this post as an example.