I interviewed with the CTO, Dmytro Dzhulgakov, for a Software Engineer position. During the interview, he requested that I draw a diagram to explain my approach. When I asked for a collaborative whiteboard link – a tool I understand other candidates have been provided – he refused to create one.

After I began verbally explaining, he repeatedly mocked me for not having a whiteboard, despite having denied my request for one. At one point, he incorrectly claimed that PyTorch's Fully Sharded Data Parallel (FSDP) cannot have more than one layer on a single GPU rank. I explained that FSDP can support part of a layer, a layer, or several layers per rank and created a diagram to illustrate. Rather than engaging with the explanation after he realized he was wrong, he started to mock the diagram as well when he realized I was correct.

Throughout the interview, he frequently changed the question mid-discussion and put words in my mouth.

Explain FSDP to me with a diagram.