`--mixed-precision` doesnt work with img transformer 2

Question

`--mixed-precision` doesnt work with img transformer 2

yoinked-h opened this issue 8 months ago · 2 comments

When trying to train with mixed precision (and natten), the pos embedding gets casted to fp32 and not bf16, causing an error later on in the attention.forward call

Answer 1 · 2024-01-24T19:38:34.000Z

I just fixed a similar sounding problem with the dtypes of the tensors being input to natten2dav(), which only occurred using very recent versions of NATTEN (commit: 6ab5146), can you pull and check to see if this fixes your problem?

Answer 2 · 2024-01-24T23:56:26.000Z

this seems to have fixed it, ty!