Fix tcp error with pytorch 1.9.1

[sdesrozis]

Fixes #2210

Description:

It seems that using PyTorch 1.9.1, tcp port should have the same value than master_port. Default is 29500.

Check list:

• New tests are added (if a new feature is added)
• New doc strings: description and/or example code are in RST format
• Documentation is updated (if required)

[vfdev-5]

@sdesrozis can you provide please a link to the docs or code about that ?

[sdesrozis]

@sdesrozis can you provide please a link to the docs or code about that ?

That's what I experimented. From now, default env variables are set by torch.distributed.run

'MASTER_ADDR': '127.0.0.1', 'MASTER_PORT': '29500'

[vfdev-5]

OK, I see your point.

[vfdev-5]

LGTM!

Can you send another PR to fix another CI issue ?

By the way, in the fixed test we are using torch.distributed.launch and not torch.distributed.run. Should we update that as well ?

[sdesrozis]

Can you send another PR to fix another CI issue ?

I'm on it.

By the way, in the fixed test we are using torch.distributed.launch and not torch.distributed.run. Should we update that as well ?

I will explore but launch uses run so, it seems more or less the same. I check the differences in details though.

[sdesrozis]

Ouch I tried using PyTorch 1.8.1 1.9.0 and 1.9.1 but nightly does a ValueError assertion. I will fix and add a comment.

[vfdev-5]

We had to also update circle ci but there is no yet 1.9.1 docker images