2.9.0

New features:

• GPU enabled Python component can be created.
• Python components with GPU dependency force parallel steps value to 1. A warning message appears if it was not set to 1 beforehand.
• Docker Image iai-sdk-python-runner is updated to be aligned with AI Inference Server 2.9.0.
• Components can be exported individually.
• Package creation fails with error message if package size is above hard limit (20 GB).
• LocalPipelineRunner and ComponentRunner check GPU accessibility for a PythonComponent if GPU was enabled. Warning messages are generated if neither PyTorch nor TensorFlow are installed with the component, or if they cannot access the GPU.
• GPU enabled Python component throws error if PyTorch dependency is requested with incompatible CUDA version.
• GPU enabled Python component throws error if TensorFlow[and-cuda] dependency is requested with incompatible CUDA version.
• Pipeline.export() now saves the edge package directly, without creating an intermediate pipeline configuration package.
• Pipeline.export() automatically creates a delta package, if a previous package exists and its size is larger than the threshold (300 MB by default)
• pipeline.export() compares the previous python environment and creates a lightweight delta automatically if possible.
• Throw an assertion error instead of warning if user intends to add unsupported data type as component input or output

Security upgrades:

• onnx < 1.21.0 - Remote Path Traversal Vulnerability - GHSA-3r9x-f23j-gc73 - 1.21.0
• pip < 26.0.1 - Remote Path Traversal Vulnerability - 26.0.1
• protobuf < 6.33.5 - Remote Denial of Service Vulnerability - GHSA-7gcm-g887-7 - 6.33.5
• wheel < 0.46.2 - Local Arbitrary Code Execution Vulnerability - GHSA-8rrh-rw8j-w5fx - 0.46.2

2.8.0

New features:

• New connection type Realtime_Information_Backbone (RIB) is added to ConnectionTypeAndPayloadFormat.
• The supported_types array now contains the numpy types used by RIB connection.
• Types for pipeline variables now allow array size information in square brackets, such as DoubleArray[30].
• Connections check for compatible pipeline variable types and topic-based pipeline parameter types. Creating a pipeline variable or topic-based pipeline parameter with a non-supported type throws a warning instead of an error. Pipeline packages with non-supported pipeline variables are still created.
• A pipeline with RIB connection logs error level messages (but still constructs the pipeline package) if the pipeline variable type is not supported. Array type pipeline variables must be provided with (non-zero) length.
• Pipeline report files and pipeline execution report files now contain error level log messages.
• RIB Cycle Time can be set for a pipeline.

Fixed issues:

• onnx ≤ 1.18.0 - Remote Path Traversal Vulnerability - CVE-2025-51480
• urllib3 1.0.x < 2.6.0 - Multiple Remote Denial of Service Vulnerabilities - GHSA-gm62-xv2j-4w53, GHSA-2xpw-w6gg-jr37
• urllib3 1.22 < 2.6.3 - Remote Denial of Service Vulnerability - GHSA-38jv-5279-wg99
• pip ≤ 25.2 - Remote Symlink Traversal Vulnerability - CVE-2025-8869

Release 2.7.0

New features:

• For GPU runtime components pipeline_config.yml contains warmup and optimization flags.
• Pipeline parameter types are extended with a new type Secret, which can be added to the pipeline with the add_secret_parameter() method.
• Code that deals with ImageSet payloads can be added to a component directory

Fixed issues:

• Docker Image iai-sdk-python-runner is updated with Python 3.12
• Python 3.10.x ≤ 3.10.18 - Multiple Vulnerabilities - 3.10.19
• Python 3.11.x ≤ 3.11.13 - Multiple Vulnerabilities - 3.11.14
• onnx ≤ 1.18.0 - Remote Path Traversal Vulnerability - CVE-2025-51480
• pip ≤ 25.2 - Remote Symlink Traversal Vulnerability - CVE-2025-8869
• protobuf ≤ 6.31.1 - Remote Path Traversal Vulnerability - 6.32.0
• protobuf 4.x < 4.25.8, 5.x < 5.29.5, 6.x < 6.31.1 - Remote Denial of Service Vulnerability - GHSA-8qvm-5x2c-j2w7
• tensorflow ≤ 2.18.0 - Multiple Vulnerabilities - CVE-2025-55556, CVE-2025-55559
• OpenSSL 3.4.x ≤ 3.4.2 - Multiple Vulnerabilities - 3.4.3
• SQLite ≤ 3.50.1 - Remote Denial of Service Vulnerability - 3.50.2
• Fixed an issue with character encoding incompatibility in PipelineRunnerReportWriter on Windows.
• Fixed: Duplicated pipeline parameter names can not be added.

Release 2.6.0

2.6.0

New features:

• Enabled Python versions for a PythonComponent is 3.11 and 3.12.
• deployment.Component.set_requirements() has an additional optional parameter no_deps (default false), to give the option to download / install Python packages without their dependencies. A warning message appears in the pipeline report file.
• LocalPipelineRunner do not download / install Python package dependencies in components with no_deps parameter. A warning message appears in the execution report file.
• Pipeline variables and parameters can be mapped during packaging, either with tag name or Databus topic.
• Connector name, connector type and connector payload format can be mapped during packaging.
• AI SDK requires the user's consent for collecting telemetry data.

Fixed issues:

• Pipeline.export() checks the final size of the pipeline package correctly.

Release 2.5.1

2.5.1

Fixed issues:

• Invalid schema $id in pipeline.schema.json caused validation error during pipeline.export() when the rfc3986 package (or JupyterLab) is installed.

Release 2.5.0

2.5.0

New features:

• Command line Pipeline execution is enabled.
• Component name is restricted to [-a-zA-Z0-9_] characters.
• The VCAStream class now supports additional image formats when converting a folder of images into an ImageSet input for LocalPipelineRunner.
• The list of image formats VCAStream class can convert a folder of images into ImageSet input for LocalPipelineRunner is extended. The supported image formats are Mono8, RGB8 (same as RGB), BGR8 (same as BGR), BayerRG8, BayerBG8, BayerGB8, BayerGR8, YUV422Packed, and YUV422_YUYV_Packed.
• ImageSet output can be enabled from a component for previewing on AIIS.
• Dependencies can be added from a pyproject.toml file's project.dependencies section.

Fixed issues:

• The execution report lists installed packages related to CPU Python components only.
• The execution report contains the list of installed packages on every run.
• Pipeline.export(...) creates a package with the same package id regardless of pipeline versioning.
• Python 3.10.x ≤ 3.10.16 - Multiple Vulnerabilities - 3.10.17
• Python 3.11.x ≤ 3.11.11 - Multiple Vulnerabilities - 3.11.12
• Python Package: onnx < 1.17.0 - Remote Path Traversal Vulnerability
• Python Package: Google Protocol Buffers (protobuf) - End of Life Notification (EOL)
• Character encoding incompatibility on Windows.

Removed features:

• Running pipelines and components in local execution no longer generates a warning when sqlite3 is imported.

Release 2.4.0

2.4.0

New features:

• New Class ComponentRunner introduced for testing Component object before packaging them.
• _package_component_dependencies in deployment.py returns a set containing (name,version) tuples of all the dependencies.
• Pipeline.export(...) generates a report md file with pipeline information, structure, direct and transitive dependencies, package vulnerabilities and warnings. The report file is placed next to the final generated zip package.
• LocalPipelineRunner generates a report md file with the execution information, zip folder structure, list of python packages installed, input/output payload counts, and warnings. The report file is placed next to the package.
• Error is raised when pipeline input variable names are present also in pipeline outputs.
• VCAStream class can convert a folder of images into ImageSet input for LocalPipelineRunner
• TimeSeriesStream class can feed a csv file as input into LocalPipelineRunner
• Dependencies with inline URLs in requirements.txt are downloaded and the URL is removed at package creation

Deprecated features:

• PythonComponent.add_resources(..) method does NOT exclude hidden files or files in hidden folder when packaging the Pipeline.

Fixed issues:

• Files from hidden folders can be added to PythonComponent resources, which were excluded in previous versions of AI SDK.
• Fixing pip report.json character encoding incompatibility on Windows.
• Python dependencies can be specified with URIs in the requirements.txt file and assigned to a Python Component.

Release 2.3.0

2.3.0

New features:

• Optimizing dependency list to reduce package size
• Pipeline.export(...) checks the size of the final generated zip package, throws error if it exceeds the limit of 2.2 GB
• PythonComponent gives warning when TensorFlow dependency is detected
• Warning about unused dependencies during LocalPipelineRunner execution
• Incomplete input payload, output or metric does not throw AssertionError anymore; instead, a warning message will be issued about the missing variables

Deprecated features:

Fixed issues:

• Python 3.10.x ≤ 3.10.14 - Multiple Vulnerabilities - 3.10.15
• Python 3.11.x ≤ 3.11.9 - Multiple Vulnerabilities - 3.11.10