Context
PR #1185 fixes the failing Ollama vision tests by splitting them into two tiers:
- Structural payload tests — mocked unit tests that verify mellea correctly embeds
ImageBlock instances into the Ollama conversation payload. These run in CI and locally with no vision model required.
- Dormant live e2e tests — a thin round-trip smoke test that is currently skipped because no current-generation Granite vision model runs on Ollama yet.
What needs to happen
granite-vision-4.1 (ibm-granite/granite-vision-4.1-4b) is the current Granite vision model. It is not yet in the Ollama library (ollama.com/library/granite-vision-4.1 → 404). When it lands, the dormant tests in test/backends/test_vision_ollama.py should be activated.
Activation steps
- Verify
ollama pull granite-vision-4.1 works.
- In
test/backends/test_vision_ollama.py:
- Remove the
pytest.skip(...) call from the vision_session fixture.
- Remove this tracking comment.
- Add
granite-vision-4.1 to the CI model-pull step in .github/workflows/quality.yml so the e2e tests run in CI.
- Optionally add
IBM_GRANITE_VISION_4_1_4B to mellea/backends/model_ids.py and update the _VISION_MODEL constant to use it.
Related
Context
PR #1185 fixes the failing Ollama vision tests by splitting them into two tiers:
ImageBlockinstances into the Ollamaconversationpayload. These run in CI and locally with no vision model required.What needs to happen
granite-vision-4.1(ibm-granite/granite-vision-4.1-4b) is the current Granite vision model. It is not yet in the Ollama library (ollama.com/library/granite-vision-4.1→ 404). When it lands, the dormant tests intest/backends/test_vision_ollama.pyshould be activated.Activation steps
ollama pull granite-vision-4.1works.test/backends/test_vision_ollama.py:pytest.skip(...)call from thevision_sessionfixture.granite-vision-4.1to the CI model-pull step in.github/workflows/quality.ymlso the e2e tests run in CI.IBM_GRANITE_VISION_4_1_4Btomellea/backends/model_ids.pyand update the_VISION_MODELconstant to use it.Related
mellea-websitebranchblog/granite-vision-structured-extraction(see the "Model availability" note in that draft)