tf.train.experimental.MaxShardSizePolicy | TensorFlow v2.16.1 (original) (raw)

Policy that splits tensors into shards with a max shard size.

View aliases

Compat aliases for migration

tf.compat.v1.train.experimental.MaxShardSizePolicy

tf.train.experimental.MaxShardSizePolicy(
    max_shard_size: int
)

Shards may exceed the max shard size if they contain 1. a single scalar/string tensor that could not be sliced and exceeds the max shard size or 2. the checkpoint object graph, whose size cannot be calculated when saving.

| Attributes | | | ----------- | | | description | |

Methods

`call`

View source

__call__(
    shardable_tensors: Sequence[tf.train.experimental.ShardableTensor]
) -> Sequence[sharding_util.TensorSliceDict]

Callback to split tensors into shards with a max shard size.

Args
shardable_tensors	A list of ShardableTensors.

Returns
List of shard dicts containing tensors. [ {checkpoint key: {slice_spec: tensor} } ]

Except as otherwise noted, the content of this page is licensed under the Creative Commons Attribution 4.0 License, and code samples are licensed under the Apache 2.0 License. For details, see the Google Developers Site Policies. Java is a registered trademark of Oracle and/or its affiliates. Some content is licensed under the numpy license.

Last updated 2024-04-26 UTC.