My experiments (the script
) showed that dataset.split_at_indices() with SPREAD tasks have more predictable performance
Concretely: on 10 m5.4xlarge nodes with 5000 iops disk
calling ds.split_at_indices(81) on 200GB dataset with 400 blocks: the split_at_indices without this PR takes 7-19 seconds, split_at_indices with SPREAD takes 7-12 seconds.
Currently, it's not very easy to figure out why a DatasetPipeline may be underperforming. Add some warnings to help guide the user. As a next step, we can try to default to a good pipeline setting based on these constraints.
This is an experimental feature, so the following changes are added only to the WandbLoggerCallback. We are planning to collect feedback about usage and accordingly update or add these changes to the other W&B integration interfaces.
Allow reading the W&B project name and group name from environment variable if not already passed to callback
Add external hooks to fetch W&B API key, and to process any information about W&B run
Signed-off-by: Nikita Vemuri <nikitavemuri@gmail.com>
Make sure the OCR example is tested in documentation after we discovered that example notebooks are not tested in CI.
Signed-off-by: Philipp Moritz <pcmoritz@gmail.com>
The current Dataset.split_at_indices() implementation suffers from O(n^2) memory usage in the small-split case (see issue) due to recursive splitting of the same blocks. This PR implements a split_at_indices() algorithm that minimizes the number of split tasks and data movement while ensuring that at most one block is used in each split task, for the sake of memory stability. Co-authored-by: scv119 <scv119@gmail.com>
The PR adds a new experimental flag to the placement group API to avoid placement group taking all cpus on each node. It is used internally by Air to avoid placement group (created by Tune) is using all CPU resources which are needed for dataset
This PR is to resolve#20888, where users have concern for the dataset-like methods used in dataset pipeline (such as map_batches, random_shuffle_each_window, etc). The reason is currently we define those dataset-like methods implicitly through Python setattr/getattr, to delegate the real work from dataset pipelien to dataset. This does not work very well with external developers/users if they want to navigate to the definition of method, or determine the method's return value data type.
So this PR is to explicitly define every dataset-like APIs in dataset pipeline class. This gives us a view of how much code we need to duplicate in upper bound. If we go with this direction, this means whenever we update or add a new method in Dataset, we need to update or add the same in DatasetPipeline.
When trail is resumed, it is useful for the user to know from which checkpoint it happened.
Signed-off-by: sustr-equi <sustr@equilibretechnologies.com>
Co-authored-by: sustr-equi <sustr@equilibretechnologies.com>
Following up from #26436, this PR adds a distributed benchmark test for Tensorflow FashionMNIST training. It compares training with Ray AIR with training with vanilla PyTorch.
Signed-off-by: Kai Fricke <kai@anyscale.com>
