Source
Ivanov et al. (2020), LIT-066 — ARXIV-2007.00072.
The same finding as SOTA-082, from the other side
TVM’s claim is that a compiler should choose layouts; LIT-066’s is that the layouts frameworks choose by default are suboptimal — a statement about defaults rather than about a missing capability, which is what makes it actionable without new tooling.
Both matter to a reader for different reasons. SOTA-082 says the choice is being made for you and the symptom of a bad one is a transpose in the profile. This says the default is worth doubting: the layout a framework picks was tuned for some hardware, and it may not be yours.
What a reader can actually do
Not much by hand, and the practice should say so rather than implying a knob. The available moves are choosing among the layouts a framework exposes (channels-last, for instance), keeping a model in one layout end to end so no conversion is inserted, and — where it matters enough — letting a compiler retune (SOTA-084).
The cost of getting it wrong is not a crash but a steady tax that looks like the model simply being that expensive. That is why it belongs beside SOTA-091: the layout question is invisible until something is measuring bytes moved rather than time elapsed.