Releases: oracle/accelerated-data-science
Releases · oracle/accelerated-data-science
ADS 2.10.0
- Improved the progress bar to use the percentage completed of workflow request instead of hardcoded steps.
- Used the service default for
WEB_CONCURRENCY
for model deployment. - Fixed the bug with zipping the model artifacts directory when
TMPRDIR
is provided. - Improved the
watch()
method for model deployment to keep streaming logs when the deployment is finished. - Changed the default log type of watch to both access logs and predict logs.
- Changed the target directory to
artifact_dir
instead of temp directory when saving the model artifacts. - Fixed the mount file system pre-check to check for duplicate
dest
. - Fixed duplicate logs in the model deployment consolidated logs.
- Added support for the optional downloading of artifacts in
GenericModel
using adownload_artifact()
method. - Set the Data Science service endpoint through the environment variable in
OCIDataScienceMixin
. - Made reloading the model to environment as optional at the time of invoking
GenericModel.from_id()
. - Mandated the Python version in
GenericModel.prepare()
when it can't be resolved. - Added a print out of the model deployment OCID in the notebook cell when
deploy()
is called.
ADS 2.9.1
- Added support for deploying LangChain application as OCI Model Deployment.
- Added support for using HuggingFace Evaluation as LLM guardrail.
- Added deployment support for RetrievalQA when using OpenSearchVectorSearch or FAISS vector DB as retriever.
- Added reload parameters in
GenericModel.save()
to provide option to not reload score.py. - Fixed a bug in model deployment progress bar due to fixed number of steps.
- Fixed a bug in
ads opctl build-image job-local
command.
ADS 2.9.0
- Introducing AI Forecast Operator. Learn more about Operators in the "Operators" section of the docs.
- Introducing PII Operator which aims to detect and redact Personal Identifiable Information in data.
- Fixed a bug with the
opctl conda create
andopctl conda publish
commands to ensure functionality on M1 and M2 local machines. - Fixed a bug with failed model deployment return value.
- Fixed a bug when sorting logs for jobs and model deployment.
ADS 2.8.11
- Added support to mount file systems in Data Science notebook sessions and jobs.
- Added support to cancel all job runs in the ADS
api
andopctl
commands. - Updated
ads.set_auth()
to use bothconfig
andsigner
when provided. - Fixed a bug when initializing distributed training artifacts with "Ray" framework.
ADS 2.9.0rc0
We are pleased to announce a release candidate for ADS 2.9.0. If all goes well, we'll release ADS 2.9.0 in few weeks.
The release will be available on PyPI and can be installed with --pre flag:
python -m pip install --pre oracle-ads==2.9.0rc0
Please report any issues with the release candidate on the ADS issue tracker.
ADS 2.8.10
- Improved the
LargeArtifactUploader
class to understand OCI paths to upload model artifacts to the model catalog by reference. - Removed
ADSDataset
runtime dependency ongeopandas
. - Fixed a bug in the progress bar during model registration.
- Fixed a bug where session variable could be referenced before assignment.
- Fixed a bug with model artifact save.
- Fixed a bug with pipelines step.
ADS 2.8.9
- Upgraded the
scikit-learn
dependency to>=1.0
. - Upgraded the
pandas
dependency to>1.2.1,<2.1
to allow you to use ADS with pandas 2.0. - Implemented multi-part upload in the
ArtifactUploader
to upload model artifacts to the model catalog. - Fixed the "Attribute not found" error, when
deploy()
called twice inGenericModel
. - Fixed the fetch of the security token, when the relative path for the
security_token_file
is provided (used in session token-bases authentication).
ADS 2.8.8
- Added
PyTorchDistributed
runtime option for Data Science jobs to add support for training large language models with PyTorch. - Added options to configure flexible shape in
opctl
. - Refactored
deploy()
inGenericModel
to prioritize the parameters. - Fixed the
opctl
commands delete/cancel/watch/activate/deactivate commands to add missing parameter options. - Fixed the
opctl
commands to call run to start an ML job when no YAML is specified. - Deprecated the
DatasetFactory
class, and refactored the code.
ADS 2.8.7
- Added support for leveraging pools in the Data Flow applications.
- Added support for token-based authentication.
- Revised help information for
opctl
commands.
ADS 2.8.6
- Resolved an issue in
ads opctl build-image job-local
when the build ofjob-local
would get stuck. Updated the Python version to 3.8 in the base environment of thejob-local
image. - Fixed a bug that prevented the support of defined tags for Data Science job runs.
- Fixed a bug in the
entryscript.sh
ofads opctl
that attempted to create a temporary folder in the/var/folders
directory. - Added support for defined tags in the Data Flow application and application run.
- Deprecated the old
ModelDeploymentProperties
andModelDeployer
classes, and their corresponding APIs. - Enabled the uploading of large size model artifacts for the
ModelDeployment
class. - Implemented validation for shape name and shape configuration details in Data Science jobs and Data Flow applications.
- Added the capability to create
ADSDataset
using the Pandas accessor. - Provided a prebuilt watch command for monitoring Data Science jobs with
ads opctl
. - Eliminated the legacy
ads.dataflow
package from ADS.