mlflow-torchserve
mlflow-torchserve copied to clipboard
mlflow
[FR] Compatibility with MLflow 2.0
Thank you for submitting a feature request. Before proceeding, please review MLflow's Issue Policy for feature requests and the MLflow Contributing Guide.
Please fill in this feature request template to ensure a timely and thorough response.
Willingness to contribute
The MLflow Community encourages new feature contributions. Would you or another member of your organization be willing to contribute an implementation of this feature (as an enhancement to the MLflow TorchServe Deployment plugin code base)?
- [ ] Yes. I can contribute this feature independently.
- [ ] Yes. I would be willing to contribute this feature with guidance from the MLflow community.
- [X] No. I cannot contribute this feature at this time.
Proposal Summary
In MLflow 2.0 (scheduled for release on Nov. 14), we will be making small modifications to the MLflow Model Server's RESTful scoring protocol (documented here: https://output.circle-artifacts.com/output/job/bb07270e-1101-421c-901c-01e72bc7b6df/artifacts/0/docs/build/html/models.html#deploy-mlflow-models) and the MLflow Deployment Client predict() API (documented here: https://output.circle-artifacts.com/output/job/bb07270e-1101-421c-901c-01e72bc7b6df/artifacts/0/docs/build/html/python_api/mlflow.deployments.html#mlflow.deployments.BaseDeploymentClient.predict).
For compatibility with MLflow 2.0, the mlflow-torchserve plugin will need to be updated to conform to the new scoring protocol and Deployment Client interface. The MLflow maintainers are happy to assist with this process, and we apologize for the short notice.
Motivation
- What is the use case for this feature? Provide a richer, more extensible scoring protocol and broaden the deployment client prediction interface beyond dataframe inputs.
- Why is this use case valuable to support for MLflow TorchServe Deployment plugin users in general? Necessary for compatibility for MLflow 2.0
- Why is this use case valuable to support for your project(s) or organization? ^
- Why is it currently difficult to achieve this use case? Without these changes, the
mlflow-torchserveplugin will break in MLflow 2.0.
What component(s) does this feature affect?
Components
- [X]
area/deploy: Main deployment plugin logic - [ ]
area/build: Build and test infrastructure for MLflow TorchServe Deployment Plugin - [ ]
area/docs: MLflow TorchServe Deployment Plugin documentation pages - [ ]
area/examples: Example code
@shrinath-suresh @chauhang Do you have bandwidth on your end to migrate the mlflow-torchserve plugin to the updated scoring protocol and adjust the Deployment Client predict() API? By my estimates, it should only take a few hours of work at most. Apologies for the short notice.
@dbczumar I am testing the changes with MLflow 2.0.
From MLflow 2.0, is it mandatory to return predict response of type pd.DataFrame or pd.Series ? . Getting error while converting text/json due to result.to_json() convertion done here
Till MLflow 1.30.0, we are returning the result as a string (code link) and mlflow cli converts it to json (code link).
Most of the ML examples (bert, mnist, iris classification), the output of the model will be either json or text. If the changes are intentional, we need to change the mlflow_torchserve plugin's predict method to convert the user response to dataframe and return it to mlflow cli
@shrinath-suresh So sorry for the delay here. The deployment client predict() method always needs to return a PredictionsResponse object (https://mlflow.org/docs/latest/python_api/mlflow.deployments.html#mlflow.deployments.PredictionsResponse). For your use case, feel free to subclass PredictionsResponse and define to_json() as desired :)
@dbczumar Can you please review this PR when you find time ? https://github.com/mlflow/mlflow-torchserve/pull/97
Has this been merged/solved and we are good to go with the latest version? Looks like it by #97 but just wanted to confirm if there/s anything outstanding since Issue is still open
