unexpected auto-casting / use input references

Hi,

I'm facing two strange issues when I use Apex with fp16:

1. When I use an encoder-decoder architecture, at the end of my encoder forward function the type of my tensor is fp16, as expected. But if I do h = encoder(x) then h is of type fp32, although the type was fp16 in the forward function just before I return h. This is a bit weird and then I have to do something like: output = decoder(h.half()). Is that expected?

2. Sometimes I provide as input a tensor and a dictionary to the encoder: h = encoder(x, dico) and the forward function of my encoder modifies the dictionary. But everything happens like if I was doing h = encoder(x, copy.deepcopy(dico)), i.e. the dictionary is identical when I exit the forward function.

Any ideas?
Thank you

Start by reproducing the two reported cases: fp16 output changing to fp32 after the encoder call, and dictionary changes not persisting when passed alongside the tensor. Compare behavior inside and outside the encoder forward function, then determine whether the observed behavior is expected or should be corrected and documented.