Refactor VLM detection in studio

[danielhanchen]

Staging mirror of unslothai#5245

Original PR: unslothai#5245
Author: Datta0

This is a staging copy for review and editing. Once finalized, changes will be pushed back to the original PR.

---

Original description

Fixes : unslothai#5239. This seems to have not been an issue.

• Add gemma4 as VLM to studio models list
• Refactor is VLM checks to single place
• Add a fallback to check vision config for the same.

---

This PR tracks the moving review branch (pr-5245-head). Iteration fix commits land here directly. Review-added tests are in a separate PR.

Changed files:

• studio/backend/utils/models/model_config.py
• studio/backend/tests/test_vision_cache.py

[danielhanchen]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors the Vision Language Model (VLM) detection logic by centralizing it into a new _is_vlm helper function and introducing a fallback mechanism, _raw_config_has_vision_config, which inspects the config.json directly if subprocess checks fail. It also adds support for gemma4 models and includes unit tests for these changes. Feedback was provided to use a more idiomatic Python approach for checking model type prefixes in the detection logic.

[gemini-code-assist]

The startswith method in Python can accept a tuple of strings directly. Using model_type.startswith(tuple(_VLM_MODEL_TYPES)) would be more efficient and idiomatic than using a generator expression with any().

Suggested change

	and any(model_type.startswith(vlm_type) for vlm_type in _VLM_MODEL_TYPES)
	and model_type.startswith(tuple(_VLM_MODEL_TYPES))

[danielhanchen]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors the Vision Language Model (VLM) detection logic by centralizing the identification criteria into a reusable _is_vlm helper function and introducing a fallback mechanism that inspects the raw config.json file when standard methods fail. It also adds support for gemma4 models and includes corresponding unit tests. Review feedback suggests improving the robustness of configuration checks by explicitly verifying that attributes are not None and handling empty configuration dictionaries more effectively.

[gemini-code-assist]

The use of bool(config["vision_config"]) might incorrectly return False if a model uses a default vision configuration represented as an empty dictionary {} in the config.json. A safer check is to verify that the key exists and its value is not None.

Suggested change

	return "vision_config" in config and bool(config["vision_config"])
	return config.get("vision_config") is not None

[gemini-code-assist]

Using hasattr on an AutoConfig object can be misleading because it returns True even if the attribute value is None. While this preserves the previous behavior, it's generally safer to check if the attribute exists and is not None to avoid false positives for models that might have these attributes initialized to None.

Suggested change

	or hasattr(config, "vision_config")
	or hasattr(config, "img_processor")
	or hasattr(config, "image_token_index")
	or getattr(config, "vision_config", None) is not None
	or getattr(config, "img_processor", None) is not None
	or getattr(config, "image_token_index", None) is not None