Hydrogen: Missing environment variables

[dmotz]

Sorry if this has been covered before, can't seem to find a solution.

When I examine os.environ from jupyter or ipython3 or python3 I can see all of my shell environment variables.

However when I run the same using Hydrogen (Python 3 kernel) in the latest version of Atom, many variables are missing. Specifically, I'm having problems with some CUDA variables needed for Python libraries. This is true whether I start Atom via the command line or otherwise.

Anything I might be missing? Is there a way to manually pass environment variables to Hydrogen?

[rgbkrk]

This sounds like we're not passing process.env on to the spawned process. Went ahead and made #405 for this.

[lgeiger]

Thanks for reporting. We just released v0.12.3 to fix this.

[dmotz]

Thanks so much for your help. After finding 0.12.3 didn't solve my issue I did some digging and discovered that OSX 10.11 is known to remove both LD_LIBRARY_PATH and DYLD_LIBRARY_PATH environment variables while SIP is turned on which is why they disappear from Atom's process.env.

While this is certainly not Hydrogen specific, I thought I'd leave this note here for anyone else stuck on trying to use CUDA libraries with it on OSX.

When I set the variables manually in startKernel() of kernel-manager.coffee like so, things work as expected in a Hydrogen session:

process.env.LD_LIBRARY_PATH = process.env.DYLD_LIBRARY_PATH = '/usr/local/cuda/lib'

I attempted the same fix by using Hydrogen's Kernel Specs setting (spec.env), but haven't had any luck. Any issues obvious here?

{
  "kernelspecs": {
    "python3": {
      "resource_dir": "/usr/local/share/jupyter/kernels/python3",
      "spec": {
        "language": "python",
        "display_name": "Python 3",
        "env": {
          "LD_LIBRARY_PATH": "/usr/local/cuda/lib",
          "DYLD_LIBRARY_PATH": "/usr/local/cuda/lib"
        },
        "argv": [
          "/usr/local/opt/python3/bin/python3.5",
          "-m",
          "ipykernel",
          "-f",
          "{connection_file}"
        ]
      }
    }
  }
}

[rgbkrk]

Oooooh, I don't think we've been using the env key in the spec there.

[lgeiger]

Now I know why I couldn't reproduce this issue. I followed this guide to disable SIP.

[lgeiger]

I attempted the same fix by using Hydrogen's Kernel Specs setting (spec.env), but haven't had any luck. Any issues obvious here?

We just released v0.12.4. This should work now.

[dmotz]

Thanks so much for looking into this! Hydrogen works perfectly with CUDA now.

[mscipio]

I think I have the very same problem on a fresh Atom+Hydrogen installation on Ubuntu 16.04.

import os 
os.environ.keys['LD_LIBRARY_PATH']

Running this lines from terminal (after starting conda venv) allows me to see everything I automatically add to the path when starting the vEnv.
Doing the same within Atom gives me a key-not-found error ... I don't think I can use the solution discussed above, as it seems MacOS specific ...

[kylebarron]

The kernelspec isn't mac specific. Run jupyter kernelspec list --json to see the kernelspecs installed

[mscipio]

Yes, I know and I have all my 2 vEnv in there, but the only way I get the PATH and LD_LIBRARY_PATH correctly set is if I first start the vEnv, and inside it I run atom --new-instance

[kylebarron]

Yes, in order for Atom to inherit a shell's environment variables, it has to be started from inside that shell.