Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

updated readme to include async and multinode features #183

Merged
merged 2 commits into from
Nov 8, 2024
Merged

updated readme to include async and multinode features #183

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
67d2148
updated readme to include async and multinode features
jdnurme Nov 8, 2024
6192d87
Address comments
jdnurme Nov 8, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 9 additions & 1 deletion README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,8 @@ The Accelerated Dataloader for PyTorch implements PyTorch’s [dataset primitive

Furthermore, the Accelerated Dataloader for PyTorch provides a checkpointing interface to conveniently save and load checkpoints directly to and from a Google Cloud Storage (GCS) bucket.

All of these features can be used out of the box for performant executions with single-node and multinode ML workflows. Demo code for multinode implementation using FSDP can be found in our [multinode README](https://github.com/GoogleCloudPlatform/dataflux-pytorch/tree/main/demo/lightning/checkpoint/multinode).

## Getting started

### Prerequisites
Expand Down Expand Up @@ -114,7 +116,7 @@ for each_object in dataset:

##### Checkpointing

The Accelerated Dataloader for PyTorch supports fast data loading and allows the user to save and load model checkpoints directly to/from a Google Cloud Storage (GCS) bucket.
The Accelerated Dataloader for PyTorch supports fast data loading and allows the user to save and load model checkpoints directly to/from a Google Cloud Storage (GCS) bucket. The checkpointing implementation leverages multipart file upload to parallelize checkpoint writes to GCS, greatly increasing performance over single-threaded writes.

```python
import torch
Expand Down Expand Up @@ -174,6 +176,12 @@ trainer.save_checkpoint(CKPT)

Note that saving or restoring checkpoint files will stage the checkpoint file in CPU memory during save/restore, requiring additional available CPU memory equal to the size of the checkpoint file.

##### Async Checkpointing

Our lightning checkpointing implementation has built-in support for the experimental [AsyncCheckpointIO](https://lightning.ai/docs/pytorch/stable/api/lightning.pytorch.plugins.io.AsyncCheckpointIO.html#asynccheckpointio) featureset. This is an optimzation that allows for non-blocking `save_checkpoint` calls during a training loop. For more details on our support for this feature please see the [checkpoint README](https://github.com/GoogleCloudPlatform/dataflux-pytorch/tree/main/demo/lightning/checkpoint/singlenode#using-asynccheckpointio). This lightning feature is only supported for single-node executions.

To allow for asyncronous saves with multinode executions we utilize PyTorch's [Async Distributed Save](https://pytorch.org/tutorials/recipes/distributed_async_checkpoint_recipe.html) to allow for similar non-blocking checkpoint operations. This implementation can be found [here](https://github.com/GoogleCloudPlatform/dataflux-pytorch/blob/main/demo/lightning/checkpoint/multinode/strategies.py#L99).

## Performance

### Map-style Dataset
Expand Down