Experiencing slower running speed using HGTConv from v2.5.0 compared to v2.3.0

[junhongmit]

Hi, I noticed HGTConv in v2.5.0 has been improved over the previous version, v2.3.0, with notably changes like using HeteroDictLinear and slicing node features into single tensor. However, I am experiencing roughly 4 times slower training speed (20 secs -> 80 secs) and 2 times slower inference speed (37 secs -> 66 secs) when simply replacing the HGTConv with the new convolution operator in v2.5.0. See the following speed comparison. I am pretty sure this is originated from HGTConv itself only. Because, after I replace the newer hgt_conv.py with the old hgt_conv.py from v2.3.0, the speed becomes normal again. Is it because the tensor slicing costs more time in the newer version?

[rusty1s]

Do you mind sharing a small example that reproduces this? Happy to take a look.