Use streaming transfers #54
Conversation
1) Added the the write_state_dict and read_state_dict implementations into checkpointing.py 2) Replaced existing torch.save/torch.load with those 3)Added unit tests for write_state_dict/read_state_dict for all the different possible types of torch tensors 4) Added checksum to read/write_state_dict that uses zlib.crc32
|
Hi @Krishn1412! Thank you for your pull request and welcome to our community. Action RequiredIn order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at [email protected]. Thanks! |
facebook-github-bot
added
the
CLA Signed
|
Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks! |
|
@Krishn1412 are you familiar with DTensor? Would you be willing to extend this to support DTensor and other tensor subclasses? |
| from io import BytesIO | ||
| import torch | ||
| from typing import Tuple | ||
| from checkpointing import CheckpointServer, TensorMetadata, write_state_dict, read_state_dict |
There was a problem hiding this comment.
| from checkpointing import CheckpointServer, TensorMetadata, write_state_dict, read_state_dict | |
| from torchft.checkpointing import CheckpointServer, TensorMetadata, write_state_dict, read_state_dict |
There was a problem hiding this comment.
github doesn't seem too happy from the failing unit test.
|
Hey @d4l3k , sure I can try adding support for other tensor subclasses. |
|
@Krishn1412 I'm going to close this PR since I'm adding a different streaming implementation in the core PyTorch repo I've added you as a coauthor to that commit: https://github.com/pytorch/pytorch/pull/146555/commits |
… methods (#146555) Summary: This is intended for use with torchft when we need to do a streaming state dict transfer. This is strictly superior to the prior streaming method in torchft as this supports all tensor subclasses such as DTensor. This supports 100% of the inputs to torch.save/load but is not wire compatible nor intended to have any backwards compatibility. Security wise this fully supports weights_only and defaults to True. It does use pickle for some metadata but uses weights_only for the metadata. Adapted from: pytorch/torchft#101 pytorch/torchft#54 Test Plan: pytest test/distributed/test_serialization.py Pull Request resolved: #146555 Approved by: https://github.com/fegin, https://github.com/mikaylagawarecki Co-authored-by: Krishn Parasar <[email protected]>
… methods (pytorch#146555) Summary: This is intended for use with torchft when we need to do a streaming state dict transfer. This is strictly superior to the prior streaming method in torchft as this supports all tensor subclasses such as DTensor. This supports 100% of the inputs to torch.save/load but is not wire compatible nor intended to have any backwards compatibility. Security wise this fully supports weights_only and defaults to True. It does use pickle for some metadata but uses weights_only for the metadata. Adapted from: pytorch/torchft#101 pytorch/torchft#54 Test Plan: pytest test/distributed/test_serialization.py Pull Request resolved: pytorch#146555 Approved by: https://github.com/fegin, https://github.com/mikaylagawarecki Co-authored-by: Krishn Parasar <[email protected]>
… methods (#146555) Summary: This is intended for use with torchft when we need to do a streaming state dict transfer. This is strictly superior to the prior streaming method in torchft as this supports all tensor subclasses such as DTensor. This supports 100% of the inputs to torch.save/load but is not wire compatible nor intended to have any backwards compatibility. Security wise this fully supports weights_only and defaults to True. It does use pickle for some metadata but uses weights_only for the metadata. Adapted from: pytorch/torchft#101 pytorch/torchft#54 Test Plan: pytest test/distributed/test_serialization.py Pull Request resolved: #146555 Approved by: https://github.com/fegin, https://github.com/mikaylagawarecki Co-authored-by: Krishn Parasar <[email protected]>
Added the the write_state_dict and read_state_dict implementations into checkpointing.py
Replaced existing torch.save/torch.load with those
3)Added unit tests for write_state_dict/read_state_dict for all the different possible types of torch tensors