Replacing ShardedTensor with DTensor for RW sharding #1991
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D54375878 |
Summary: **This is the first part of migration TorchRec state dict checkpointing from ShardedTensor to DTensor. It sets up the necessary infra to support additional sharding schemes. The general approach is to keep ShardedTensor paths and remove them once all sharding types are supported on DTensor. This includes ShardingPlan and ShardedTensor dataclasses such as ShardedTensorMetadata.** NOTE: This version of LocalShardsWrapper does not support empty shards, that is added in the next diff enabling CW. D57063512 **This diff includes:** + LocalShardsWrapper torch.tensor subclass to be used with DTensor + Changes in TorchRec state_dict load and creation to use DTensor for Row Wise path in both EmbeddingCollection and EmbeddingBagCollection + Changes to DCP to support LocalShardsWrapper for saving and reading (WriteItems and ReadItems) + Added DTensor paths to callsites where ShardedTensors are expected. **LocalShardsWrapper supports the following torch ops:** + torch.ops._c10d_functional.all_gather_into_tensor.default + aten._to_copy.default + aten.view.default + aten.equal.default With extensibility to add more as required by use cases. See https://docs.google.com/document/d/16Ptl50mGFJW2cljdF2HQ6FwsiA0scwbAbjx_4dhabJw/edit?usp=drivesdk for more info regarding design and approach. Differential Revision: D54375878
zainhuda
force-pushed
the
export-D54375878
branch
from
7089269
to
0e61990
Compare
|
This pull request was exported from Phabricator. Differential Revision: D54375878 |
zainhuda
pushed a commit
to zainhuda/pytorch
that referenced
this pull request
May 13, 2024
Summary: X-link: pytorch/torchrec#1991 **This is the first part of migration TorchRec state dict checkpointing from ShardedTensor to DTensor. It sets up the necessary infra to support additional sharding schemes. The general approach is to keep ShardedTensor paths and remove them once all sharding types are supported on DTensor. This includes ShardingPlan and ShardedTensor dataclasses such as ShardedTensorMetadata.** NOTE: This version of LocalShardsWrapper does not support empty shards, that is added in the next diff enabling CW. D57063512 **This diff includes:** + LocalShardsWrapper torch.tensor subclass to be used with DTensor + Changes in TorchRec state_dict load and creation to use DTensor for Row Wise path in both EmbeddingCollection and EmbeddingBagCollection + Changes to DCP to support LocalShardsWrapper for saving and reading (WriteItems and ReadItems) + Added DTensor paths to callsites where ShardedTensors are expected. **LocalShardsWrapper supports the following torch ops:** + torch.ops._c10d_functional.all_gather_into_tensor.default + aten._to_copy.default + aten.view.default + aten.equal.default With extensibility to add more as required by use cases. See https://docs.google.com/document/d/16Ptl50mGFJW2cljdF2HQ6FwsiA0scwbAbjx_4dhabJw/edit?usp=drivesdk for more info regarding design and approach. Test Plan: TODO: add a MAST job using RW and preempt to test checkpointing ``` buck2 test 'fbcode//mode/opt' fbcode//torchrec/distributed/tests:test_model_parallel_nccl ``` ``` buck2 test 'fbcode//mode/opt' fbcode//torchrec/distributed/tests:test_model_parallel_nccl_single_rank -- --exact 'torchrec/distributed/tests:test_model_parallel_nccl_single_rank - torchrec.distributed.tests.test_model_parallel_nccl_single_rank.ModelParallelStateDictTestNccl: test_load_state_dict' ``` Sandcastle Differential Revision: D54375878
|
creating new PR |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
This is the first part of migration TorchRec state dict checkpointing from ShardedTensor to DTensor. It sets up the necessary infra to support additional sharding schemes. The general approach is to keep ShardedTensor paths and remove them once all sharding types are supported on DTensor. This includes ShardingPlan and ShardedTensor dataclasses such as ShardedTensorMetadata.
NOTE: This version of LocalShardsWrapper does not support empty shards, that is added in the next diff enabling CW. D57063512
This diff includes:
LocalShardsWrapper supports the following torch ops:
With extensibility to add more as required by use cases.
See https://docs.google.com/document/d/16Ptl50mGFJW2cljdF2HQ6FwsiA0scwbAbjx_4dhabJw/edit?usp=drivesdk for more info regarding design and approach.
Differential Revision: D54375878