Enable gemma3-4b-it in VLM Pipeline

[yangsu2022]

Ticket: CVS-166388

[Copilot]

Pull Request Overview

This PR enables support for the Gemma3-4b-it model in the Visual Language Model (VLM) pipeline. The implementation adds comprehensive support for Gemma3 including configuration classes, input embedders, vision encoders, and the necessary infrastructure for token type IDs which are required for distinguishing between text and image tokens in the multimodal Gemma3 model.

• Addition of Gemma3 model type and configuration parameters for image processing and token handling
• Implementation of specialized embedding classes for Gemma3 with token type ID support
• Integration of Gemma3 image tag <start_of_image> in documentation and API references

Reviewed Changes

Copilot reviewed 35 out of 35 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/python_tests/test_vlm_pipeline.py	Adds test cases for tiny-random-gemma3 model with SDPA backend requirement
src/python/py_vlm_pipeline.cpp	Updates documentation to include gemma-3-4b-it image tag
src/python/openvino_genai/py_openvino_genai.pyi	Adds type stub documentation for Gemma3 image tag
src/cpp/src/visual_language/vlm_config.hpp	Defines GEMMA3 model type and configuration parameters
src/cpp/src/visual_language/vlm_config.cpp	Registers gemma3 model type mapping
src/cpp/src/visual_language/vision_encoder.cpp	Adds VisionEncoderGemma3 creation logic
src/cpp/src/visual_language/processor_config.hpp	Adds Gemma3-specific image size configuration
src/cpp/src/visual_language/processor_config.cpp	Implements Gemma3 size parameter parsing
src/cpp/src/visual_language/pipeline.cpp	Updates attention backend requirements for Gemma3
src/cpp/src/visual_language/llava_next/classes.cpp	Moves merge function to utils namespace
src/cpp/src/visual_language/llava/classes.cpp	Refactors to use shared merge utility function
src/cpp/src/visual_language/inputs_embedder.hpp	Adds token type ID support interface
src/cpp/src/visual_language/inputs_embedder.cpp	Implements token type ID functionality and Gemma3 embedder creation
src/cpp/src/visual_language/gemma3/classes.hpp	Defines Gemma3 vision encoder and input embedder classes
src/cpp/src/visual_language/gemma3/classes.cpp	Implements Gemma3-specific image processing and embedding logic
src/cpp/src/utils.hpp	Declares shared merge function for text/image embeddings
src/cpp/src/utils.cpp	Implements shared merge utility function
src/cpp/src/tokenizer/chat_template_fallback_map.hpp	Adds Gemma3 chat template without trim filter
src/cpp/src/speculative_decoding/speculative_decoding_impl.hpp	Updates interface to support token type IDs
src/cpp/src/speculative_decoding/speculative_decoding_impl.cpp	Implements token type ID passing for speculative decoding
src/cpp/src/sequence_group.hpp	Adds token type ID storage to sequence groups
src/cpp/src/prompt_lookup/prompt_lookup_impl.hpp	Updates interface for token type ID support
src/cpp/src/prompt_lookup/prompt_lookup_impl.cpp	Implements token type ID forwarding
src/cpp/src/lm_encoding.hpp	Updates function signature for token type ID parameter
src/cpp/src/lm_encoding.cpp	Implements token type ID tensor handling in language model
src/cpp/src/llm/pipeline_stateful.cpp	Updates call to pass token type IDs
src/cpp/src/continuous_batching/pipeline_impl.hpp	Adds token type ID support to pipeline interface
src/cpp/src/continuous_batching/pipeline_impl.cpp	Implements token type ID handling in batching pipeline
src/cpp/src/continuous_batching/pipeline_base.hpp	Updates base interface for token type ID support
src/cpp/src/continuous_batching/pipeline_base.cpp	Implements token type ID processing in VLM generation
src/cpp/src/continuous_batching/model_runner.hpp	Adds token type ID tensor and processing logic
src/cpp/include/openvino/genai/visual_language/pipeline.hpp	Updates API documentation with Gemma3 image tag
site/docs/use-cases/image-processing/_sections/_usage_options/index.mdx	Documents Gemma3 image tag usage
site/docs/supported-models/_components/vlm-models-table/models.ts	Adds Gemma3 to supported models table
samples/cpp/visual_language_chat/benchmark_vlm.cpp	Updates comment about continuous batching limitations

Comments suppressed due to low confidence (1)

[Wovchena]

Thank you!