Ayush Agrawal

This commit introduces **support for specifying hyperparameters** during **distillation tuning** within the `google.genai` library, enabling more granular control over the tuning process. It **adds new fields** like `batch_size`, `learning_rate`, and `tuning_mode` to the relevant type definitions in `google/genai/types.py` and updates the `_CreateTuningJobConfig_to_vertex` function in `google/genai/tunings.py` to correctly map these parameters to the `distillationSpec`. This **new capability** allows users to optimize their distillation jobs by configuring key training parameters, with a new test case ensuring proper functionality.

This commit introduces a **feature enhancement** to the **file upload mechanism** within the `google.genai` library. It implements logic in `google/genai/_api_client.py` to **rewrite upload URLs** based on the `base_url` specified in `http_options` and to merge custom headers for these requests. This ensures that file uploads correctly respect **custom API endpoints** or proxy configurations, enhancing the flexibility of the client. New tests in `google/genai/tests/client/test_upload_errors.py` validate the correct URL rewriting for both synchronous and asynchronous upload operations, confirming the **robustness** of this configuration handling.

This commit performs a **maintenance update** to align the **embedding model tests** with current model availability by replacing the deprecated `text-embedding-004` ID with `gemini-embedding-001`. The changes primarily affect the **test suite** for the `google.genai` library, specifically updating model references in **batch embedding creation tests** for BigQuery and GCS, and various **content embedding test cases**. This **chore** ensures that all embedding-related tests utilize the correct, non-deprecated model, including the addition of output dimensionality configuration in shared test utilities. The update improves the reliability and relevance of the **embedding feature validation** by preventing failures due to outdated model identifiers.

This commit **adds support for multimodal embedding** and **MaaS models** within the **`Models.embed_content()` function**, specifically targeting **Gemini Embedding 2.0** and other **Vertex AI embedding models**. This **new capability** allows users to embed diverse content types, including text, images, and PDFs, by adapting the `embed_content` implementation in the **`google.genai.models` module** to handle different Vertex AI API types and parameter transformations. New types are introduced in **`google.genai.types`** to support the expanded API, and a helper function `t_is_vertex_embed_content_model` aids in model identification. Extensive new test cases validate the multimodal and MaaS model support, significantly enhancing the **embedding functionality** and compatibility with the **Vertex AI API**.

This commit **enhances the documentation** for the **GenAI Python SDK** by adding comprehensive docstrings. Specifically, it introduces detailed explanations and usage examples to the `get` methods within both the synchronous and asynchronous `Operations` classes in `google/genai/operations.py`. This **documentation improvement** aims to provide clearer guidance for developers, making it easier to understand and correctly utilize the **SDK's operations functionality**. The change directly impacts the developer experience by offering better in-line help and examples for key methods.

This commit introduces a **new capability** to the **GenAI SDK**, allowing users to specify an `encryption_spec` when creating **tuning jobs**. This enhancement enables the use of **customer-managed encryption keys (CMEK)**, significantly improving data security for tuning artifacts and models. The change involves adding an `encryption_spec` field to `CreateTuningJobConfig` in `google/genai/types.py`, implementing its mapping for Vertex AI tuning job configurations and adding validation for Gemini API in `google/genai/tunings.py`, and including a new test case in `google/genai/tests/tunings/test_tune.py`.

This commit introduces a **significant internal refactoring** to the **`google.genai` SDK's data conversion logic**. It modifies numerous conversion functions across `models.py` and `tunings.py` to accept a new `root_object` parameter, enabling a more robust internal conversion mechanism. This **architectural update** provides **response-level root object access** to all SDK converters. This change is essential for supporting consistent data handling and cross-SDK compatibility, specifically for the Go SDK.

This commit introduces **support for Open Source Software (OSS) model tuning** within the **GenAI SDK**, enabling users to fine-tune OSS models directly through the SDK. This **new capability** extends the `CreateTuningJobConfig` in `google/genai/types.py` by adding new fields such as `tuning_mode` and `custom_base_model`. The implementation in `google/genai/tunings.py` provides the crucial mapping logic to convert these new OSS tuning parameters, including `batch_size` and `learning_rate`, into the appropriate Vertex AI format. Comprehensive new test cases in `google/genai/tests/tunings/test_tune.py` validate various OSS tuning scenarios, including Supervised Fine-Tuning (SFT) and distillation, ensuring robust functionality for this expanded tuning support.

This commit **introduces a new capability** by adding **support for distillation tuning** within the `google/genai` library. It defines new types such as `DistillationHyperParameters` and `DistillationSpec` in `google/genai/types.py`, and extends the `TuningMethod` enum to include `DISTILLATION`. The **tuning subsystem** in `google/genai/tunings.py` is updated to process these distillation-specific parameters, modifying key conversion functions like `_CreateTuningJobConfig_to_vertex`. A new test case in `google/genai/tests/tunings/test_tune.py` has been added to validate this **new tuning method**, enabling users to configure and execute distillation-based tuning jobs.

This commit performs **dependency management** by **bumping the lower bound** for the `google-auth` library to **2.47.0**, primarily impacting the **GenAI and Vertex SDKs**. This ensures that these SDKs utilize a sufficiently recent version of the authentication library. As part of this **maintenance update**, a minor **type checking refinement** was also applied to `google/genai/live.py` within the **GenAI SDK**. Specifically, a `type ignore` comment was added to the `creds.refresh` call to suppress untyped call warnings, improving code hygiene.

This commit introduces a **new capability** by implementing an empty response for the `tunings.cancel()` method within the **`google.genai.tunings` module**. The `cancel` methods in both `Tunings` and `AsyncTunings` are now modified to process and return a `CancelTuningJobResponse` object, ensuring a consistent return type for cancellation operations. This required defining new `CancelTuningJobResponse` classes and type aliases in the **`google.genai.types` module**. This **API enhancement** provides a clear and predictable response structure for tuning job cancellations, with a related test case updated in `google/genai/tests/tunings/test_cancel.py`.

This commit performs a **maintenance update** to the **tuning tests** within the `google.genai` library. It updates the base models used in various test configurations, specifically changing `gemini-2.0-flash-001` to `gemini-2.5-flash-lite` in `google/genai/tests/tunings/test_end_to_end.py` and `gemini-1.5-pro-002` to `gemini-2.5-flash` in `google/genai/tests/tunings/test_tune.py`. This ensures that the **model tuning functionality** is validated against the latest `gemini-2.5` series models, reflecting current best practices and available model versions. The change primarily affects the **test suite**, ensuring its continued relevance and accuracy without altering core library behavior.

This commit **updates the test suite** for the `generate_content` functionality within the **`google.genai` library**. It primarily migrates various test cases across files like `test_generate_content_config_zero_value.py` and `test_generate_content_part.py` to utilize the newer **`gemini-2.5-flash`** models, replacing older `gemini-1.5` or `gemini-2.0` references. This **maintenance update** ensures the content generation API is validated against the latest model capabilities, incorporating minor adjustments for expected exceptions, media types, and assertion logic to reflect current model behavior. The changes are confined to the **testing infrastructure**, enhancing the reliability of `generate_content` API validation with current model references.

feat: Use pytest-xdist for test parallelization

This commit **updates the test suite** for the **`google.genai` library**, specifically modifying tests for the `generate_content` method. It **migrates existing test cases** in `test_generate_content.py` to utilize the newer **`gemini-2.5` models**, ensuring compatibility and correct behavior with the latest model versions. Additionally, it adjusts stream assertion counts within these tests to align with the expected output characteristics of the updated models. This **maintenance** work improves the **relevance and accuracy of the `generate_content` tests**, validating the library's interaction with cutting-edge Gemini models.

This commit introduces **input validation** to the **`google.genai._transformers` module**, specifically within the `t_model` function. It **sanitizes model name inputs** by preventing the use of problematic characters like '..', '?', or '&', which could lead to unexpected behavior or security vulnerabilities. This **robustness improvement** ensures that the `genai` library handles model parameters more securely and predictably, with new test cases verifying the correct error handling for invalid inputs.

This commit **refactors** the **`Metric` class** within the **`google.genai.types` module** by updating its `validate_name` method. The validation logic for `Metric` names is transitioned from a class-based approach to an **instance-based validator**. This **chore** improves the internal consistency and object-oriented design of the `Metric` validation mechanism. The change is an internal code improvement and does not alter the external API or behavior of the `Metric` class.

This commit **regenerates the project documentation** for version 1.48.0, incorporating significant content and styling updates. It **updates the main documentation page** (`index.rst.txt`) to include new **installation instructions**, **client management** details, **content generation examples** (including image output), **tokenization**, and experimental **MCP/JSON schema support**. Additionally, it **refines the Pygments syntax highlighting CSS** (`pygments.css`) by adjusting background and foreground colors for improved readability. This **documentation update** ensures users have access to the latest information and a more polished reading experience.

This commit **drops official support for Python 3.9** across the project's development and testing infrastructure. Specifically, it **removes Python 3.9** from the supported versions matrix within the `.github/workflows/import.yml` and `.github/workflows/mypy.yml` **CI/CD workflows**. This **maintenance** change reflects Python 3.9 reaching its End-of-Life (EOL), ensuring that automated checks align with currently supported Python versions. As a result, future development and automated testing will no longer include Python 3.9, streamlining the project's continuous integration processes.

chore: Allow google-cloud-storage >=3.0.0