When fixing #24, #26 introduces a new bug. Specifically, _limit returns Float[...] (nominally logits) when callers may be expecting Bool[...]. This leads to OOM, as well as potential calculation errors if the output is interpreted as bool via torch.where and other checks.
When fixing #24, #26 introduces a new bug. Specifically,
_limitreturnsFloat[...](nominally logits) when callers may be expectingBool[...]. This leads to OOM, as well as potential calculation errors if the output is interpreted as bool viatorch.whereand other checks.