Enable Gemma4 video input#4114
Draft
yatarkan wants to merge 13 commits into
Draft
Conversation
github-actions
Bot
added
category: visual language
Contributor
There was a problem hiding this comment.
Pull request overview
This PR extends the Gemma4 VLM implementation to accept video inputs by introducing a video placeholder token, adding Gemma4 video frame encoding, and merging video embeddings into the text embeddings flow. It also updates the Python VLM pipeline tests to exercise Gemma4 as a video-capable model.
Changes:
- Add Gemma4 video placeholder token to VLM configuration and Python test prompt generation.
- Implement Gemma4 video encoding (
encode_frames/encode_videos) and video-aware prompt normalization + embedding merge. - Register Gemma4 tiny-random model IDs as video models in the Python test suite.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.
| File | Description |
|---|---|
tests/python_tests/test_vlm_pipeline.py |
Adds Gemma4 model IDs to the video test matrix and defines `< |
src/cpp/src/visual_language/vlm_config.hpp |
Introduces video_token configuration (default `< |
src/cpp/src/visual_language/gemma4/classes.hpp |
Extends Gemma4 vision encoder + inputs embedder interfaces to support video inputs. |
src/cpp/src/visual_language/gemma4/classes.cpp |
Implements video metadata handling, frame encoding/concatenation, video tag expansion in prompts, and video embedding merge. |
…hc offset, add asserts
Comment on lines
+167
to
+181
| if (!video_config.do_sample_frames) { | ||
| video_metadata.frames_indices.resize(total_num_frames); | ||
| std::iota(video_metadata.frames_indices.begin(), video_metadata.frames_indices.end(), 0); | ||
| return; | ||
| } | ||
|
|
||
| size_t num_frames = video_config.num_frames; | ||
|
|
||
| if (num_frames == 0) { | ||
| num_frames = std::min(total_num_frames, video_config.max_frames > 0 ? video_config.max_frames : total_num_frames); | ||
| } | ||
|
|
||
| num_frames = std::min(num_frames, total_num_frames); | ||
| OPENVINO_ASSERT(num_frames > 0, "Number of frames to sample must be positive."); | ||
|
|
Contributor
Author
There was a problem hiding this comment.
Gemma4 video processor config does not have fps and min_frames params and they are not used in HF transformers implementation
Comment on lines
+652
to
660
| void InputsEmbedderGemma4::encode_vision_token_ids() { | ||
| std::call_once(m_vision_token_ids_once_flag, [this]() { | ||
| const auto encoded_vision_tokens = | ||
| m_tokenizer.encode(m_vlm_config.image_token + m_vlm_config.video_token, ov::genai::add_special_tokens(false)).input_ids; | ||
| OPENVINO_ASSERT(encoded_vision_tokens.get_size() == 2, "Encoded vision tokens must contain two tokens"); | ||
| m_image_token_id = encoded_vision_tokens.data<int64_t>()[0]; | ||
| m_video_token_id = encoded_vision_tokens.data<int64_t>()[1]; | ||
| }); | ||
| } |
Contributor
Author
There was a problem hiding this comment.
Such pattern is already used in GenAI to have a single tokenizer.encode call
Comment on lines
+286
to
+297
| std::vector<ov::Tensor> frame_features; | ||
| frame_features.reserve(frames.size()); | ||
|
|
||
| for (const auto& frame : frames) { | ||
| EncodedImage encoded = encode_with_config(frame, m_video_processor_config); | ||
| frame_features.push_back(std::move(encoded.resized_source)); | ||
| } | ||
|
|
||
| // Concatenate all frame features: [1, total_tokens, hidden_size] | ||
| size_t num_soft_tokens_per_frame = frame_features[0].get_shape()[1]; | ||
| const size_t hidden_size = frame_features[0].get_shape()[2]; | ||
| const size_t total_tokens = frames.size() * num_soft_tokens_per_frame; |
Contributor
Author
There was a problem hiding this comment.
Input video frames (and its embeddings) are expected to have the same shape and element type
Comment on lines
+547
to
+551
| for (size_t f = 0; f < encoded_video.frame_num; ++f) { | ||
| ov::Tensor frame_embed(elem_type, {1, tokens_per_frame, hidden_size}); | ||
| std::memcpy(frame_embed.data(), src + f * bytes_per_frame, bytes_per_frame); | ||
| video_embeds.push_back(std::move(frame_embed)); | ||
| } |
… tiny-random-gemma4" This reverts commit 364138b.
Comment on lines
+205
to
210
| std::vector<size_t> original_video_indices(video_frames_num); | ||
| std::iota(original_video_indices.begin(), original_video_indices.end(), 0); | ||
|
|
||
| if (video_metadata.frames_indices.empty() || video_metadata.frames_indices == original_video_indices) { | ||
| return video; | ||
| } |
Comment on lines
6
to
+11
| #include <cmath> | ||
| #include <cstring> | ||
| #include <sstream> | ||
| #include <iomanip> | ||
| #include <numeric> | ||
|
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
This PR adds video input support for Gemma4 model family.
CVS-185850
Checklist: