Skye Wanderman-Milne

Would it work to have a (slow) function call that tries to free unused memory? I say "try" because the default allocator allocates large regions, with multiple DeviceArrays possibly occupying...

If you'd like to delete all on-device buffers, you can now reach directly into the underlying C++ internals like this: ```python backend = jax.lib.xla_bridge.get_backend() for buf in backend.live_buffers(): buf.delete() ```...

Are you able to say more about what's not compliant about it, or what kind of error you're getting? I can believe we messed something up here, but it's a...

@hawkinsp do you have any ideas as to where we could (easily) host such a page? I'm gonna mark this contributions welcome for now, since it sounds like anyone can...

I thought GitHub pages were only per-user/organization, but it looks like you can have per-project pages as well. So yes, this does seem like it should be possible. I can...

I haven't tried building with cuda 11.1, only 11.0, so it's possible something's wrong with the build script. Can you post the full output of the build.py command? Alternatively, if...

Hey, sorry for the delay on this. I'm not sure what's going on with your build unfortunately, but I was able to build CUDA 11.1 jaxlib wheels in a docker...

I haven't updated the installation instructions in the README yet but you should be able to use this command: ``` pip install --upgrade jax jaxlib==0.1.56+cuda111 -f https://storage.googleapis.com/jax-releases/jax_releases.html ```

Can you set the environment variable `TF_CPP_MIN_LOG_LOVEL=0`, rerun, and send us the full output? Also can you share the output of `nvidia-smi`?

I filed an internal XLA issue requesting that they expose the device count setting programatically, so we can create a JAX API on top of that. We'll still need to...