[tivris]

Description

Fixes #1935.

_register_unowned_container_state registered unmapped buffers of every module, but unmapped parameters only of modules with children. The per-layer embedding modules on Gemma 4 E2B/E4B and Gemma 3n and the AltUp projections on Gemma 3n are unmapped leaves, so .half() and the other nn.Module._apply conversions, conversions of a parent module, and parent.parameters() missed their weights.

The children check is gone, so unmapped parameters get the same owner views as unmapped buffers. This also reaches the unmapped multimodal modules: the audio tower on Gemma 4 E2B/E4B, the audio and vision towers on Gemma 3n, the vision tower on T5Gemma2 and the VQ tokenizer on Emu3. bridge.state_dict() keys and hooks are unchanged; parent.state_dict() gains _container_state_owners.* keys for the newly reachable parameters. As noted in #1935, test_parent_and_direct_traversal_have_identical_state no longer flags an adapter that leaves a leaf module unmapped.

test_parent_module_traversal.py gains two Gemma 4 cases (per-layer embeddings, and the audio and vision towers) and test_dtype_conversion_updates_unmapped_leaf_parameters for bridge.half() and parent.to(torch.float16); test_gemma3n_bridge.py gains test_half_converts_unmapped_parameters. All five fail on dev and pass with the fix.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Looks good, merging now!