In #1320 some of the tests failed due to differences in precision.
We have seen this before in other places.
The log messages included...
I0000 00:00:1774527545.928435 2907 port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
I0000 00:00:1774527545.929005 2907 cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
#1321 attempted to fix this by setting TF_ENABLE_ONEDNN_OPTS to 0 as suggested.
All tests passed locally (Python 3.11.14; tensorflow-2.18.0) but failed under Python 3.10 on ubuntu-latest on the first run. Tests do pass when re-running (e.g. on #1320) but noting as something to keep an eye on and possibly investigate. Mildly annoying to have non-deterministic behaviour creeping in.
On the plus side Sphinx docs appear to build again 🙌
In #1320 some of the tests failed due to differences in precision.
We have seen this before in other places.
The log messages included...
#1321 attempted to fix this by setting
TF_ENABLE_ONEDNN_OPTSto0as suggested.All tests passed locally (Python 3.11.14; tensorflow-2.18.0) but failed under Python 3.10 on
ubuntu-lateston the first run. Tests do pass when re-running (e.g. on #1320) but noting as something to keep an eye on and possibly investigate. Mildly annoying to have non-deterministic behaviour creeping in.On the plus side Sphinx docs appear to build again 🙌