Ray and python version mismatch

[ciroaceto]

Using the following code,

from ray_on_aml.core import Ray_On_AML

ray_on_aml = Ray_On_AML(ml_client=ml_client, compute_cluster='rl-agents-cluster')

ray = ray_on_aml.getRay(ci_is_head=True,
    num_node=2,
    pip_packages=[
        "tensorflow==2.11"
    ]
)

I am getting this error:

RuntimeError: Version mismatch: The cluster was started with:
    Ray: 2.2.0
    Python: 3.8.10
This process on node 10.2.1.4 was started with:
    Ray: 2.3.0
    Python: 3.8.5

I checked the source code and the python version is assigned from platform.python_version(). Why is it possible that the python version does not match?

On the other hand, ray version, the ray version of the cluster should be the same of the head node, if it is not specified, shouldn't it?

Any thoughts on why this is happening?

Thanks in advance.

[james-tn]

When you specify your own pip_packages, you're supposed to provide ray dependencies as well. Pls add ray dependencies to pip_packages

[ciroaceto]

Even adding "ray[default]==2.3" and "python==3.8.5" to pip_packages the same error comes up.

[james-tn]

@hyssh @chnldw to check

[james-tn]

No need for python. Python is not in pip. Are you sure you did something like this?

pip_packages=["ray[air]==2.3.0","ray[data]==2.3.0"])

[ciroaceto]

Completely sure. This is the code I'm using:

ray = ray_on_aml.getRay(ci_is_head=True,
    num_node=2,
    pip_packages=[
        "ray[rllib]==2.3",
        "ray[air]==2.3",
        "tensorflow==2.11"
    ]
)

Same error of version mismatch, both in ray and python. The kernel used is Python 3.8 - Jupyter. I suppose it isn't really important, but for you to have more information.

[ciroaceto]

Now the ray version is the same, but the python version is different: 3.8.5 and 3.8.10

[jm-nab]

Not exactly related, but this was the top search result for this error.

I ran into this error, when the ray-head and ray-worker versions mismatched. I shelled into the ray head and ran python --version and sure enough, the version on the ray head was different from what was on the workers.

I fixed it by providing the image and repository information to the spec.