perf(lazy_turtle): targeted ancestor+subtree map in materialize_submodule

[Qubitium]

Summary

LazyTurtle.materialize_submodule was building dict(target_model.named_modules()) on every call, which is O(total_modules × modules_per_layer) and dominated per-module load time for large MoE models. This change replaces that full-model scan with a targeted map that contains only the target submodule, its ancestors, and (when recurse=True) its descendants, which is exactly what _resolve_prefer_transposed_hint needs.

# before
modules_by_name = dict(target_model.named_modules())

# after
modules_by_name = {"": target_model}
for i in range(len(parts)):
    modules_by_name[prefix] = target_model.get_submodule(prefix)
modules_by_name[module_path] = target_submodule
if recurse:
    for subname, submod in target_submodule.named_modules():
        modules_by_name[...] = submod

• materialize_submodule now accepts optional module_path, recurse, tie_weights, and show_progress kwargs, builds the targeted map, and passes it to _copy_checkpoint_tensors_into_submodule.
• _copy_checkpoint_tensors_into_submodule now accepts an optional modules_by_name map and falls back to a full named_modules() scan only when none is provided.
• Added tests/test_lazy_turtle_materialize_map.py with regression tests that monkeypatch nn.Module.named_modules to raise if a full model scan is used, while still verifying is_transposed hints on ancestors and descendants and numerical parity of the loaded weights.
• Added scripts/benchmark_lazy_turtle_materialize_map.py and scripts/profile_lazy_turtle_lifecycle.py for micro-benchmarking map construction and measuring the full turtle -> shell -> disk lifecycle on a synthetic MoE model.

Validation:

• ruff check ... clean
• pytest -q tests/test_structure.py tests/test_lazy_turtle_materialize_map.py — 5 passed
• python scripts/benchmark_lazy_turtle_materialize_map.py and LAZY_LIFECYCLE_LAYERS=2 python scripts/profile_lazy_turtle_lifecycle.py run successfully on CPU

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring