Add fit_tracking_metrics option to configure_generation_strategy (#5238)

[saitcakmak]

Summary:
Implements the feature requested in #4688: a fit_tracking_metrics option on the high-level Client.configure_generation_strategy API, so users can disable modeling of tracking metrics without manually editing the generation strategy.

client.configure_generation_strategy(fit_tracking_metrics=False)

When False, only metrics in the optimization config (objectives and outcome constraints) are modeled. Tracking metrics are still recorded but not modeled by the BoTorch model — useful when many tracking metrics are kept only for book-keeping.

Changes

• ax/api/client.py — added fit_tracking_metrics: bool = True to configure_generation_strategy and _choose_generation_strategy, threading it into the dispatch struct. Docstrings flag that when False, model-dependent analyses (cross-validation, sensitivity analysis, etc.) will not be produced for tracking metrics, that it requires an optimization config, and that it has no effect for random_search.
• ax/api/utils/structs.py — added the fit_tracking_metrics field to GenerationStrategyDispatchStruct.
• ax/api/utils/generation_strategy_dispatch.py — _get_mbm_node injects fit_tracking_metrics into the MBM node's generator_kwargs; only the MBM node receives it (Sobol/random-search nodes don't model outcomes).
• Tests — added test_gs_fit_tracking_metrics (dispatch) and test_configure_generation_strategy_with_fit_tracking_metrics (client); updated existing tests that assert the full MBM generator_kwargs dict.

Implementation note

fit_tracking_metrics is a TorchAdapter.__init__ kwarg. The registry's kwarg-splitter (ax/adapter/registry.py) automatically routes any key matching an adapter constructor param into the adapter, so adding it to the node's generator_kwargs is sufficient — no changes to registry.py, torch.py, or base.py.

Test Plan:

• Dispatch tests: 13 passed. Client tests pass. flake8 + ufmt clean.
• End-to-end verified with a genuine tracking metric m2: fit_tracking_metrics=True → modeled outcomes ['m1', 'm2']; False → ['m1'] (tracking metric excluded from modeling).

Closes #4688

Reviewed By: mpolson64

Differential Revision: D109735998

Pulled By: saitcakmak

[meta-codesync]

@saitcakmak has imported this pull request. If you are a Meta employee, you can view this in D109735998.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.57%. Comparing base (025d7e5) to head (d445a74).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #5238   +/-   ##
=======================================
  Coverage   96.56%   96.57%           
=======================================
  Files         619      619           
  Lines       70273    70296   +23     
=======================================
+ Hits        67861    67889   +28     
+ Misses       2412     2407    -5     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[meta-codesync]

@saitcakmak has exported this pull request. If you are a Meta employee, you can view the originating Diff in D109735998.

[meta-codesync]

@saitcakmak merged this pull request in 5fd8dbf.