[model] support gemma4 padding_free

[Jintao-Huang]

No description provided.

[gemini-code-assist]

Code Review

This pull request refactors the Gemma4 configuration and model initialization to support bidirectional attention in vision contexts, including a null check for the attention mask in the forward pass. A critical issue was identified in the Gemma4TextGPTModel constructor where sliding window attention parameters were being incorrectly cleared for all non-vision models, which would disable a core architectural feature; a code suggestion was provided to ensure these parameters are only reset when the custom bidirectional mask is active with the 'unfused' backend.

[Jintao-Huang]

/gemini review

[gemini-code-assist]

Code Review

This pull request updates the configuration and initialization logic for the gemma4 model. Specifically, it removes a conditional check in the configuration parser to unconditionally set window size parameters and introduces handling for the 'vision' bidirectional attention setting within the Gemma4TextGPTModel, including backend compatibility checks. Feedback was provided regarding the need for defensive programming: specifically, adding null checks for window_size and layer_types in the parser to prevent TypeErrors, and using getattr when accessing configuration attributes to avoid potential AttributeErrors.