How to fix model logging in MLFlow?

Question

How to fix model logging in MLFlow?

Gustavo Carmo 30

Hi.

I am getting a 404 when logging models using MLFlow:

...
    os.environ["MLFLOW_TRACKING_URI"] = workspace.mlflow_tracking_uri
...
    mlflow.pytorch.log_model(
        pytorch_model=dummy_model,
        name="simple_pytorch_model",
        signature=signature,
        input_example=dummy_input.numpy()
    )

The error message is: mlflow.exceptions.MlflowException: API request to endpoint /api/2.0/mlflow/logged-models failed with error code 404 != 200. Response body: ''

This is being performed towards the westus3 region and the workspace has been created in May of 2024.

Is this API currently supported/supposed to be working?

Thank you

Manas Mohanty 13,425 Reputation points Moderator

2025-06-30T00:43:30.7633333+00:00

Hi Gustavo Carmo

Hope you found Jerald Felix's input useful. Please let us know if downgrading mlflow <2.8 or skipping the name param in current version fixes your issue.

Hi Jerald , Code and environment details were shared on 2288368

Thank you.
Manas Mohanty 13,425 Reputation points Moderator

2025-07-01T06:55:45.4033333+00:00

Hi Gustavo Carmo

We could not hear back from you.

Hope rolling back to Mlflow<2.8.0 fixed the issue for you. Please let us know if you have any observations to share.

Thank you.

Answer accepted by question author

1 additional answer

Your answer

Manas Mohanty 13,425 Reputation points Moderator

2025-06-30T00:43:30.7633333+00:00

Hi Gustavo Carmo

Hope you found Jerald Felix's input useful. Please let us know if downgrading mlflow <2.8 or skipping the name param in current version fixes your issue.

Hi Jerald , Code and environment details were shared on 2288368

Thank you.
Manas Mohanty 13,425 Reputation points Moderator

2025-07-01T06:55:45.4033333+00:00

Hi Gustavo Carmo

We could not hear back from you.

Hope rolling back to Mlflow<2.8.0 fixed the issue for you. Please let us know if you have any observations to share.

Thank you.

Answer 1

Jerald Felix 9,920

Hi Gustavo Carmo

I ran into the same 404 error last week and it turned out to be a simple client-vs-server version mismatch.

Why it’s happening

MLflow 2.8 (and anything newer) added a “logged-model” API. When you pass a name= argument to mlflow.pytorch.log_model() (or any log_model call), the client tries to hit /api/2.0/mlflow/logged-models/*. Azure ML workspaces are still running an MLflow server ≤ 2.7, which only understands the older /registered-models/* routes. The server doesn’t recognise the new endpoint and sends back a 404.

Two quick ways to unblock your run

Downgrade the clientbash<br>pip install "mlflow<2.8"<br># e.g. mlflow==2.7.1<br>Your script keeps the one-step “log + register” flow and the server responds with 200.

Keep the new client, skip the new API

Drop the name= parameter, then register in a second line: python mlflow.pytorch.log_model( pytorch_model=model, artifact_path="simple_pytorch_model" ) run_uri = f"runs:/{mlflow.active_run().info.run_id}/simple_pytorch_model" mlflow.register_model(run_uri, "simple_pytorch_model")

With no name=, the client falls back to the classic workflow the server already supports.

Tip: Inside an Azure ML job you don’t need to set MLFLOW_TRACKING_URI; the runtime injects the correct azureml://… value automatically. Using the HTTPS URI from the portal can also trigger 404s.

Quick checklist if it still fails

python -m pip show mlflow — confirm the version you expect.

az ml workspace show -n <name> — check the region; rollouts happen region by region.

az ml job create ... in a fresh venv after pinning MLflow < 2.8.

Still 404? Grab the request ID from the error, share a minimal repro notebook to Q and A support here to create suppor ticket —they can confirm whether your workspace has the newer server build yet.

Once the backend upgrades past MLflow 2.8 you can go back to the latest client and the single-line log_model(name=…) call will work again.

Hope that helps—shout if you get stuck!

Best Regards,

Jerald Felix

Gustavo Carmo 30 Reputation points

2025-07-17T20:03:50.0933333+00:00
@Jerald Felix Apologies for the delay, and thanks for the reply.

I tried dropping the name= parameter, but got the same error:

mlflow.pytorch.log_model( pytorch_model=dummy_model, signature=signature, input_example=dummy_input.numpy() )

File "/usr/local/lib/python3.10/dist-packages/mlflow/utils/rest_utils.py", line 310, in verify_rest_response raise MlflowException( mlflow.exceptions.MlflowException: API request to endpoint /api/2.0/mlflow/logged-models failed with error code 404 != 200. Response body: ''

Interestingly, mlflow was not installed directly or as a dependency. I only had azure-mlflow, which I suppose may default to some other mlflow version underneath.

The result of pip3 list | grep mlflow is:

azureml-mlflow 1.60.0 mlflow-skinny 3.1.0

I then forced installed mlflow=2.7.1 with pip3 install mlflow==2.7.1, and it got past that error with this:

Traceback (most recent call last): File "/workspaces/Dexman/src/2025_06_29_dexman_model_logging_test.py", line 71, in <module> mlflow.pytorch.log_model( TypeError: log_model() missing 1 required positional argument: 'artifact_path'

This worked fine:

local_model_path = "simple_pytorch_model.pt" torch.save(dummy_model.state_dict(), local_model_path) mlflow.pytorch.log_model( pytorch_model=dummy_model, signature=signature, input_example=dummy_input.numpy(), artifact_path=local_model_path, ) os.remove(local_model_path)

Thanks!

Gustavo
Jason Koh 6 Reputation points

2025-07-18T20:44:41.4866667+00:00

@Jerald Felix Is there any plan to upgrade mlflow server inside Azure ML to v3?

Because of CVEs, we have to upgrade mlflow client version from v2 to v3, but because AzureML doesn't support mlflow v3, we can't do that, which is a blocker.

https://github.com/mlflow/mlflow/issues/16215#issuecomment-2967218244
Benzshawel, Thomas M 0 Reputation points

2025-12-17T16:58:27.6566667+00:00

Came to this months later, MLflow and azure still have the same issue. I had to specifically install these versions in order to stop getting the error.

azure-ai-ml==1.28.1

azureml-core==1.60.0.post1

azureml-mlflow==1.60.0.post1

mlflow<2.8.0

mlflow-skinny<3.0.0

That got me the mlflow versions below that worked with the azureml-mlflow package properly.

mlflow==2.7.1

mlflow-skinny==2.22.4

Answer 2

Alert101 5

I came across this as well. Fixed it in Azure ML Studio by pinning the mlflow-skinny package to an earlier version in my requirements.yml file. The combination below works. Without pinning mlflow-skinny Azure MLS will install 3.1.4 which doesn't work.

  - pip:
      - azureml-mlflow==1.60.0
      - azure-ai-ml==1.27.1
      - mlflow-skinny==3.1.1

Ref: https://github.com/mlflow/mlflow/issues/16215

Share via

How to fix model logging in MLFlow?

1 additional answer

Your answer