Reproducibility

If an operation depends on randomness or on non-deterministic kernels (for example some parallel reductions, or GPU kernels), two runs may not produce bit-identical results. TensorPlay provides a set of helpers to request deterministic algorithms and to detect when one is in use.

import tensorplay as tp

print(tp.are_deterministic_algorithms_enabled())   # False by default
tp.use_deterministic_algorithms(True)
print(tp.are_deterministic_algorithms_enabled())   # True

use_deterministic_algorithms is the main switch. When enabled, TensorPlay raises an error for operations that cannot be run deterministically, so you can find them instead of silently getting a different answer.

tp.use_deterministic_algorithms(True, warn_only=True)

Passing warn_only=True downgrades that error to a warning, which is useful for tracking down culprits without stopping mid-run. tp.is_deterministic_algorithms_warn_only_enabled() reports whether you are in that mode.

There is also a debug mode that tells you which operation is at fault:

tp.set_deterministic_debug_mode(True)   # report the offending op
tp.get_deterministic_debug_mode()       # read the mode back