Closed joelberkeley closed 8 months ago
Hey @joelberkeley, in this repo we precompile a subset of the XLA library that is necessary specifically for our use case in Elixir (exla). The error you are getting may be because a certain part of XLA is missing. We are happy to support other use cases, as long as it does not increase the precompiled binary too much, however we can't really invest time digging into such issues ourselves.
OK, that's fair, though I might emphasise that this used to work on 0.3.0.
BTW how sure are you that this is the problem? I'm finding it difficult to debug this so want to be as clear as possible
OK, that's fair, though I might emphasise that this used to work on 0.3.0.
Things we includ changed over time, and XLA itself keeps evolving, so the contents of various Bazel packages shifted. At one point XLA was extracted from tensorflow to a separate project.
BTW how sure are you that this is the problem? I'm finding it difficult to debug this so want to be as clear as possible
I am definitely not sure, it could be something with the environment as well. I mostly wanted to emphasise that we don't include the whole XLA and test only the functionality we use in the other project, so things missing are to be expected.
I think there are four options:
(a) wrong API usage; I don't know the internals, perhaps ClientLibrary
expects something else to have been created/registered
(b) something specific to the environment itself
(c) a part of XLA missing in our binaries (expected, not really a bug)
(d) a bug in the XLA source
btw I resolved this by building XLA myself and moving to PJRT
I'm unable to get a local client on recent versions of XLA (I believe after 0.3.0, definitely for 0.6.0)
I've created the following MWE that should reproduce it on mac M1. I'm seeing the same problem in github actions so I am confident you would see it on an Ubuntu machine with the x86_64-linux-gnu-cpu binary
where
mwe.cpp
is