execution.

PiperOrigin-RevId: 477578582

This had been implemented as a private function inside tff_task.py.
Moving it to a member function of ModelWeights makes it available for
TFF clients.

PiperOrigin-RevId: 477553269

States explicitly that if `random_seed=None` a nondeterministic seed is used, and verbose description of algorithm is moved from description of `random_seed` arg into pydoc body.

PiperOrigin-RevId: 477484897

change the test precision because there is some compiler-level diff in the new optimizer.

PiperOrigin-RevId: 477286793

PiperOrigin-RevId: 477040301

New optimizer actually works in these tests, but it has some state diff from the old one. In this transition period I am changing to use legacy optimizer for stable tests.

PiperOrigin-RevId: 477006818

PiperOrigin-RevId: 476978932

PiperOrigin-RevId: 476444931

equality assertions in unittests easier.

PiperOrigin-RevId: 476248553

Docker run host port should be 80 instead of 8000, to match both the notebook and the K8s instructions.

PiperOrigin-RevId: 476237802

PiperOrigin-RevId: 476213336

Note: Our golden testing tools do not have the capability to use different golden files for different versions of Python, so for now we can just disable this golden test.
PiperOrigin-RevId: 476148838

having dangling references in the TF cache.

Without this serialization will leak memory.

PiperOrigin-RevId: 476131338

PiperOrigin-RevId: 475895274

PiperOrigin-RevId: 475674293

PiperOrigin-RevId: 475454622

PiperOrigin-RevId: 474860586

* When I updated the logic to unpack/pack `round_number` into the `program_state` (instead of conflating `round_number` with `version`), I did not update all of the the necessary logic. Round number needs to be packed into `structure` so that program state can be loaded as well.

* `tff.program.ProgramStateManager.load` and `tff.program.ProgramStateManager.load_latest` take a nested Python structure, but this example was passing an asyncio Awaitable. This changed when we started using the native TFF platform to execute the computations.

* The loaded `round_number` was off by one.

TESTING: Manually tested.
PiperOrigin-RevId: 474424489

Note that this only provides non-trivial functionality when using TFF optimizers.

PiperOrigin-RevId: 474068574

I've also re-ordered the parameters to be more natural based on how callers should think about setting them. Parameters that typically don't need to be overridden are moved to the end of the parameter list.

PiperOrigin-RevId: 474044845

PiperOrigin-RevId: 474039662

PiperOrigin-RevId: 474034427

PiperOrigin-RevId: 473876655

PiperOrigin-RevId: 473829152

PiperOrigin-RevId: 473806579

Update the `attr.s` attributes of `LearningAlgorithmState` to allow equality checks, but not ordering.

PiperOrigin-RevId: 473780921

PiperOrigin-RevId: 473778862

Improve the error message from raised when selecting more clients than are available from a `tff.program.DatassetDataSourceIterator`.

This adds the number of clients selected and the number of clients that were available at the time the error condition was encountered.

PiperOrigin-RevId: 473771454

PiperOrigin-RevId: 473769877

PiperOrigin-RevId: 473521480

PiperOrigin-RevId: 473307371

PiperOrigin-RevId: 473296886

PiperOrigin-RevId: 473087203

I was tired of occasionally seeing lint errors, so I formatted all the files and manually reverted some of the markdown changes.

PiperOrigin-RevId: 473055143

PiperOrigin-RevId: 473005290

PiperOrigin-RevId: 472972872