Default installation of JAX + Tensorflow produces inconsistent environment about admin-requests HOT 6 OPEN

I can recreate. Thank you for the detailed information.

I think that:

1. The current jax feedstock should add constraints to the jaxlib dependency and write a repopatch to patch existing builds. cc @conda-forge/jax @conda-forge/jaxlib
2. We are aware that the tensorflow feedstock has become incompatible with the latest packages from conda-forge. A few PRs are trying to alleviate that, but it might require more work due to abseil library mismatches.

If you could:

1. Try to manually specify the versions you want. You will have to downgrade the version of jax you want is my guess.

mamba install "tensorflow>=2.7.0" "jax=x.x.x" "jaxlib=y.y.y."

that would help alleviate things
2. Try to submit repo-data patches to require the appropriate versions of jaxlib for jax.
3. Submit a pull request to jax fixing the dependnecy for future packages.

from admin-requests.

I believe this should be resolved now, no? At least the current packages are compatible. A repodata patch may be worthy tho

from admin-requests.

If I add RUN conda update conda to the Dockerfile I described above then this is indeed fixed.

If I don't do this then the default base image, condaforge/miniforge3:latest is still inconsistent.

from admin-requests.

If I add RUN conda update conda to the Dockerfile I described above then this is indeed fixed.

If I don't do this then the default base image, condaforge/miniforge3:latest is still inconsistent.

Okay, I see.

Btw, is this for a GPU or CPU deployment? I highly recommend you follow hmaarrfk's recommendation above. Our jaxlib>=0.3.10 (thus jax>=0.3.10) and tensorflow>=2.8.1 are somewhat different from previous versions (enabling XLA for tf and adding cuda for jax). Also, both are slightly more advanced in terms of migrations (they're incompatible with things like gdal, etc. as the latter haven't been migrated to the latest icu).

from admin-requests.

I've only tested on CPU.

I was indeed able to work around things with @hmaarrfk's suggestions above, so thanks for that!

from admin-requests.

Who is in charge of condaforge/miniforge3:latest? We can just bump that up?

Re CPU vs GPU, I am asking because our current recommendation is that users specify *cuda* explicitly, e.g. tensorflow>=2.8.1=*cuda* and jaxlib>=0.3.10=*cuda*. In case you need that info in the future :)

from admin-requests.