Ax

Summary: This diff addresses two issues in the computation of inference trace: 1. The generation strategy is copied inside run_optimization_with_orchestrator --> we retrieve the traces on an unused generation strategy --> get_best_point defaults to the best raw observation on ALL obserations 2. Relevant data not filtered in the fallback option for get_best_parameters_from_model_predictions_with_trial_index Both of these individually lead to the inference trace being incorrect - the first to the best raw value of ALL trials, the second to the best predicted across ALL trials. Changes: Moved copying of generation strategy Added argument use_model_only_if_good to force model-based BPR even if model fit is bad Clearer sequencing in get_best_parameters_from_model_predictions_with_trial_index Differential Revision: D80019803

Summary: As titled, a bit of cleanup and ensuring that we choose the correct GS + experiment types Differential Revision: D69616303

Summary: Adds a arg to that snaps values to a grid anchored at (in ), for both FLOAT and INT parameters. This is the first diff in the step_size unification stack: will subsume both the discrete-grid and limited-resolution () use cases under one knob. Next diff will add storage support. The internal DB has already been updated to include the new column. We will then migrate all current usage off and onto . We will add support for treating low-cardinality float-range parameters as discrete in , so that it is efficiently optimized over the correct grid (rather than having to use continuous optimization + rounding). At this point, we will have proper support for , so we can update the ax/api usage to leverage it, rather than resolving to . We can then deprecate and do any remaining clean-up. In this diff coexists with the existing arg (they are mutually exclusive at construction). Subsequent diffs in the stack migrate storage (JSON + SQA), transforms and utils, and the public API () to , then deprecate in favor of it. Behavior: rounds to the nearest integer and returns . It does NOT clamp to : an out-of-bounds input (e.g. a historical observation recorded outside the current bounds) snaps to the nearest grid point, which may itself be out of bounds. This mirrors the non- , which leaves out-of-bounds values in place rather than silently moving them into range — range validity is enforced by , not . Both bounds must lie on the grid: must be an integer multiple of (within ). Off-grid bounds are rejected at construction. This guarantees is itself a feasible value, so a value near the upper bound snaps to rather than to a grid point short of it. must be strictly positive, and must be integer-valued for INT parameters. accounts for : a grid-valued FLOAT reports the finite number of grid points instead of , and a grid-valued INT counts grid points rather than every integer in . defines a discrete grid but does not, by itself, force discrete acquisition optimization; how the optimizer treats the parameter depends on the grid cardinality and is determined at the generator level. Differential Revision: D107274057