Open rom1504 opened 3 years ago
Another way to get the same error: Wrap your model initialization in deepspeed.zero.Init
. Which is also the reason why #222 is still WIP – it encounters the same issue.
See this comment which explains when this happens in the generation.
I think this is a DeepSpeed issue or us using the API not conforming to their idea. I asked about API usage for our case in this issue but haven't gotten a response sadly. I also tried keeping the VAE completely separate from the DALLE
model, passing it in as a parameter instead but this hasn't helped either.
issue about the fact generate is not possible with fp16 (deepspeed) introduced when fp16 feature was introduced https://github.com/lucidrains/DALLE-pytorch/pull/157 :
that's the exact error. (issue also mentionned in #256 )
I'm looking into it
Any idea on the topic is welcome