Used this tutorial to install pytorch for rocm, however I checked out release 1.5. https://github.com/ROCmSoftwarePlatform/pytorch/wiki/Building-PyTorch-for-ROCm Allennlp was version 0.9.
This used bert-base with a batch size of 8.
Vega FE notes
The vega frontier edition results were obtained from a rented gpueater instance.
A batch size of 16 was also tried for the vega frontier edition to see if it would fit in vram and strangely the time per epoch dropped (01:12) with the larger batch size). This was also with thermal throttling as the vega fe was hitting 87 C and the clocks were down to 1.2 Ghz from 1.6 Ghz. The fans were limited to 40% under load on gpueater.com. It would be interesting to see what the performance is like with better thermals.
|GPU||BERT-base emotion regression||GRU pos-tagger (1-hid)||GRU pos-tagger (2-hid)|
|Vega Frontier (90% fans)||1:09.1||0:02.3||0:03.0|
|Vega frontier (rocm 4.0)||1:07.5||0:02.4||0:02.9|
using rocm apex gave no discernable performance improvement (with use_apex = true) However, it did reduce memory consumption by ~1GB for a batch of 16.
The RTX 3090 was tested with cuda 11, all other nvidia gpus were using cuda 10.2 (the RTX 3090 is not supported in this earlier version of cuda).