Open shraiysh opened 2 weeks ago
@golechwierowicz
cc @Tixxx for viz.
Hey,
As a mitigation would disabling the pass work for you? It is implemented as such: https://github.com/tensorflow/tensorflow/pull/73606/files.
As for the problems: Re.1. We did not think this would be an issue in practice. But for cases like this, as an escape hatch we wanted to rely on pass disabling mechanism. Re.2. Can you not feed the profile to initialization modules at all? We have tested this with JAX_ENABLE_PGLE feature and it worked well.
As for the solutions: In 3P this proves to be an issue so for the long term plan, I think we can go with solution 1 and issue a relevant warning as a default. A testing methodology is a separate topic and should be pursued in parallel nonetheless.
cc: @frgossen @cheshire
In a recent commit, a hard error was introduced in PGLE when instructions are missing from the profile. The hard error has the following problems:
Possible solutions:
--strict-pgle
. The default behavior should not have a hard error with missing instructions, instead it should use a default cost for them.