ml_switcheroo.generated_tests.runtime
=====================================

.. py:module:: ml_switcheroo.generated_tests.runtime

.. autoapi-nested-parse::

   Runtime helpers for generated verification tests.

   This module contains the reference implementation of ``verify_results``,
   a robust recursive comparison utility for validating equivalence between
   arbitrary data structures (Arrays, Lists, Dicts) across different frameworks.

   It also defines the ``ensure_determinism`` fixture, which is injected into
   generated tests to enforce reproducibility by seeding RNGs for Torch, JAX,
   TensorFlow, Numpy, and Python.



Functions
---------

.. autoapisummary::

   ml_switcheroo.generated_tests.runtime.ensure_determinism
   ml_switcheroo.generated_tests.runtime.verify_results


Module Contents
---------------

.. py:function:: ensure_determinism() -> None

   Auto-injects fixed seeds for reproducibility at the start of every test.

   Covers:
   - Python `random`
   - NumPy `np.random`
   - PyTorch `torch.manual_seed` (CPU & CUDA)
   - TensorFlow `tf.random.set_seed`
   - MLX `mlx.core.random.seed`


.. py:function:: verify_results(ref: Any, val: Any, rtol: float = 0.001, atol: float = 0.0001, exact: bool = False) -> bool

   Cross-framework comparison helper.

   Recursively compares data structures (Lists, Dicts, Tuples, Arrays).

   Modes:
   - **Fuzzy (Default)**: Uses `np.allclose` with tolerances for floats.
   - **Exact**: Enforces strict equality (ids for None, `np.array_equal` for arrays).

   :param ref: The reference value (e.g. from Source Framework).
   :type ref: Any
   :param val: The candidate value (e.g. from Target Framework).
   :type val: Any
   :param rtol: Relative tolerance for floating point comparison.
   :type rtol: float
   :param atol: Absolute tolerance for floating point comparison.
   :type atol: float
   :param exact: If True, disables fuzzy matching.
   :type exact: bool

   :returns: True if values are considered equivalent.
   :rtype: bool