Conversation
Exposes Google's Gemma vision-language models (Gemma 4 31B and Gemma 4 26B A4B) as a workflow block, served via OpenRouter with a user-supplied API key. Supports the standard VLM task set: unconstrained, OCR, structured answering, classification, multi-label classification, VQA, captioning, and unprompted object detection. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Mirrors the Llama Vision test suite: manifest validation across image, prompt, model version, api_key, output_structure, temperature, and task-specific required-field paths (prompt, classes, output_structure). Adds object-detection to the classes-required parametrize list since Gemma supports it. Includes mocked execute_gemma_request happy and error paths to cover the OpenRouter choices-None failure mode. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Erol444
requested review from
PawelPeczek-Roboflow,
dkosowski87,
grzegorz-roboflow,
hansent,
probicheaux,
rafel-roboflow and
yeldarby
as code owners
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3dfcc8de9f
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| }, | ||
| }, | ||
| ) | ||
| api_key: Union[Selector(kind=[STRING_KIND]), str] = Field( |
There was a problem hiding this comment.
The api_key field only allows Selector(kind=[STRING_KIND]), which makes this block incompatible with secret-provider outputs (for example environment_secrets_store@v1 emits SECRET_KIND). In practice, workflows that keep OpenRouter keys in the secrets pipeline cannot connect that output to google_gemma@v1, so users are forced to pass the key as a plain string selector instead of a secret-typed value.
Useful? React with 👍 / 👎.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
1 participant
Tested locally. ATM user needs to pass openrouter api key manually, in the future we could udpate this so it uses apiproxy.
Summary
roboflow_core/google_gemma@v1workflow block that exposes Google's Gemma vision-language models (Gemma 4 31B and Gemma 4 26B A4B) through OpenRouter, using the same OpenAI-compatible client pattern as the existing Llama Vision block.apiproxy/openrouterproxy on the platform side yet, sorf_key:accountis intentionally not wired in.test_llama_3_2_vision.py, including mocked happy/error paths for thechoices=Nonefailure mode that OpenRouter occasionally returns.Test plan
pytest tests/workflows/unit_tests/core_steps/models/foundation/test_google_gemma.py— 24 passed/workflows/blocks/describeshowsroboflow_core/google_gemma@v1with all 9 tasks)🤖 Generated with Claude Code