Add iter singular value into TBE optimizer state#3228
Closed
csmiler wants to merge 1 commit into
Closed
Conversation
❌ Deploy Preview for pytorch-fbgemm-docs failed.
|
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D63909559 |
csmiler
pushed a commit
to csmiler/FBGEMM
that referenced
this pull request
Oct 9, 2024
Summary: X-link: meta-pytorch/torchrec#2474 X-link: facebookresearch/FBGEMM#326 When the optimizer states for sharded embedding tables are tracked in TorchRec, they are assumed to be either point-wise (same shape as the embedding table, for example, Adam's exp_avg), or row-wise (same length as the embedding hashsize, for example, rowwise_adagrad's momentum/sum). However, there may be other formats, a single value for each table. Specifically, for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb and GWD, the `iter` number is a single value tensor, which **cannot be tracked and checkpointed properly** (this also means that there is a bug in Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb usages!) Here we support tracking and checkpointing single-value states, by constructing ShardMetadata with rowwise-sharding and replicating the single-value for each Sharded param (this is similar to how the rowwise state for colume-wise sharded tables are concatenated along row-dim). By doing so, single-value `iter` can be properly checkpointed just like other states, ensuring correct reloading of states and continuous training. This diff checkpoints `iter` for rowwise_adagrad with GWD. The next diff would checkpoint `iter` for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb. Reviewed By: spcyppt Differential Revision: D63909559
csmiler
force-pushed
the
export-D63909559
branch
from
d7fa453 to
2dae0fe
Compare
csmiler
pushed a commit
to csmiler/FBGEMM
that referenced
this pull request
Oct 9, 2024
Summary: X-link: meta-pytorch/torchrec#2474 X-link: facebookresearch/FBGEMM#326 When the optimizer states for sharded embedding tables are tracked in TorchRec, they are assumed to be either point-wise (same shape as the embedding table, for example, Adam's exp_avg), or row-wise (same length as the embedding hashsize, for example, rowwise_adagrad's momentum/sum). However, there may be other formats, a single value for each table. Specifically, for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb and GWD, the `iter` number is a single value tensor, which **cannot be tracked and checkpointed properly** (this also means that there is a bug in Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb usages!) Here we support tracking and checkpointing single-value states, by constructing ShardMetadata with rowwise-sharding and replicating the single-value for each Sharded param (this is similar to how the rowwise state for colume-wise sharded tables are concatenated along row-dim). By doing so, single-value `iter` can be properly checkpointed just like other states, ensuring correct reloading of states and continuous training. This diff checkpoints `iter` for rowwise_adagrad with GWD. The next diff would checkpoint `iter` for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb. Reviewed By: iamzainhuda, spcyppt Differential Revision: D63909559
csmiler
pushed a commit
to csmiler/torchrec
that referenced
this pull request
Oct 9, 2024
Summary: X-link: pytorch/FBGEMM#3228 X-link: facebookresearch/FBGEMM#326 When the optimizer states for sharded embedding tables are tracked in TorchRec, they are assumed to be either point-wise (same shape as the embedding table, for example, Adam's exp_avg), or row-wise (same length as the embedding hashsize, for example, rowwise_adagrad's momentum/sum). However, there may be other formats, a single value for each table. Specifically, for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb and GWD, the `iter` number is a single value tensor, which **cannot be tracked and checkpointed properly** (this also means that there is a bug in Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb usages!) Here we support tracking and checkpointing single-value states, by constructing ShardMetadata with rowwise-sharding and replicating the single-value for each Sharded param (this is similar to how the rowwise state for colume-wise sharded tables are concatenated along row-dim). By doing so, single-value `iter` can be properly checkpointed just like other states, ensuring correct reloading of states and continuous training. This diff checkpoints `iter` for rowwise_adagrad with GWD. The next diff would checkpoint `iter` for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb. Reviewed By: iamzainhuda, spcyppt Differential Revision: D63909559
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D63909559 |
csmiler
pushed a commit
to csmiler/FBGEMM
that referenced
this pull request
Oct 10, 2024
Summary: X-link: meta-pytorch/torchrec#2474 X-link: facebookresearch/FBGEMM#326 When the optimizer states for sharded embedding tables are tracked in TorchRec, they are assumed to be either point-wise (same shape as the embedding table, for example, Adam's exp_avg), or row-wise (same length as the embedding hashsize, for example, rowwise_adagrad's momentum/sum). However, there may be other formats, a single value for each table. Specifically, for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb and GWD, the `iter` number is a single value tensor, which **cannot be tracked and checkpointed properly** (this also means that there is a bug in Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb usages!) Here we support tracking and checkpointing single-value states, by constructing ShardMetadata with rowwise-sharding and replicating the single-value for each Sharded param (this is similar to how the rowwise state for colume-wise sharded tables are concatenated along row-dim). By doing so, single-value `iter` can be properly checkpointed just like other states, ensuring correct reloading of states and continuous training. This diff checkpoints `iter` for rowwise_adagrad with GWD. The next diff would checkpoint `iter` for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb. Reviewed By: iamzainhuda, spcyppt Differential Revision: D63909559
csmiler
force-pushed
the
export-D63909559
branch
from
2dae0fe to
07ab246
Compare
csmiler
pushed a commit
to csmiler/torchrec
that referenced
this pull request
Oct 10, 2024
Summary: X-link: pytorch/FBGEMM#3228 X-link: facebookresearch/FBGEMM#326 When the optimizer states for sharded embedding tables are tracked in TorchRec, they are assumed to be either point-wise (same shape as the embedding table, for example, Adam's exp_avg), or row-wise (same length as the embedding hashsize, for example, rowwise_adagrad's momentum/sum). However, there may be other formats, a single value for each table. Specifically, for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb and GWD, the `iter` number is a single value tensor, which **cannot be tracked and checkpointed properly** (this also means that there is a bug in Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb usages!) Here we support tracking and checkpointing single-value states, by constructing ShardMetadata with rowwise-sharding and replicating the single-value for each Sharded param (this is similar to how the rowwise state for colume-wise sharded tables are concatenated along row-dim). By doing so, single-value `iter` can be properly checkpointed just like other states, ensuring correct reloading of states and continuous training. This diff checkpoints `iter` for rowwise_adagrad with GWD. The next diff would checkpoint `iter` for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb. Reviewed By: iamzainhuda, spcyppt Differential Revision: D63909559
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D63909559 |
Summary: X-link: meta-pytorch/torchrec#2474 X-link: facebookresearch/FBGEMM#326 When the optimizer states for sharded embedding tables are tracked in TorchRec, they are assumed to be either point-wise (same shape as the embedding table, for example, Adam's exp_avg), or row-wise (same length as the embedding hashsize, for example, rowwise_adagrad's momentum/sum). However, there may be other formats, a single value for each table. Specifically, for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb and GWD, the `iter` number is a single value tensor, which **cannot be tracked and checkpointed properly** (this also means that there is a bug in Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb usages!) Here we support tracking and checkpointing single-value states, by constructing ShardMetadata with rowwise-sharding and replicating the single-value for each Sharded param (this is similar to how the rowwise state for colume-wise sharded tables are concatenated along row-dim). By doing so, single-value `iter` can be properly checkpointed just like other states, ensuring correct reloading of states and continuous training. This diff checkpoints `iter` for rowwise_adagrad with GWD. The next diff would checkpoint `iter` for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb. Reviewed By: iamzainhuda, spcyppt Differential Revision: D63909559
csmiler
pushed a commit
to csmiler/torchrec
that referenced
this pull request
Oct 10, 2024
Summary: X-link: pytorch/FBGEMM#3228 X-link: facebookresearch/FBGEMM#326 When the optimizer states for sharded embedding tables are tracked in TorchRec, they are assumed to be either point-wise (same shape as the embedding table, for example, Adam's exp_avg), or row-wise (same length as the embedding hashsize, for example, rowwise_adagrad's momentum/sum). However, there may be other formats, a single value for each table. Specifically, for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb and GWD, the `iter` number is a single value tensor, which **cannot be tracked and checkpointed properly** (this also means that there is a bug in Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb usages!) Here we support tracking and checkpointing single-value states, by constructing ShardMetadata with rowwise-sharding and replicating the single-value for each Sharded param (this is similar to how the rowwise state for colume-wise sharded tables are concatenated along row-dim). By doing so, single-value `iter` can be properly checkpointed just like other states, ensuring correct reloading of states and continuous training. This diff checkpoints `iter` for rowwise_adagrad with GWD. The next diff would checkpoint `iter` for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb. Reviewed By: iamzainhuda, spcyppt Differential Revision: D63909559
csmiler
force-pushed
the
export-D63909559
branch
from
07ab246 to
4c23038
Compare
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D63909559 |
Contributor
|
This pull request has been merged in f9f0600. |
facebook-github-bot
pushed a commit
to meta-pytorch/torchrec
that referenced
this pull request
Oct 11, 2024
Summary: Pull Request resolved: #2474 X-link: pytorch/FBGEMM#3228 X-link: facebookresearch/FBGEMM#326 When the optimizer states for sharded embedding tables are tracked in TorchRec, they are assumed to be either point-wise (same shape as the embedding table, for example, Adam's exp_avg), or row-wise (same length as the embedding hashsize, for example, rowwise_adagrad's momentum/sum). However, there may be other formats, a single value for each table. Specifically, for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb and GWD, the `iter` number is a single value tensor, which **cannot be tracked and checkpointed properly** (this also means that there is a bug in Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb usages!) Here we support tracking and checkpointing single-value states, by constructing ShardMetadata with rowwise-sharding and replicating the single-value for each Sharded param (this is similar to how the rowwise state for colume-wise sharded tables are concatenated along row-dim). By doing so, single-value `iter` can be properly checkpointed just like other states, ensuring correct reloading of states and continuous training. This diff checkpoints `iter` for rowwise_adagrad with GWD. The next diff would checkpoint `iter` for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb. Reviewed By: iamzainhuda, spcyppt Differential Revision: D63909559 fbshipit-source-id: e14c1dc3e8f87bfc4cc95f2321b358526719d88f
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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:
When the optimizer states for sharded embedding tables are tracked in TorchRec, they are assumed to be either point-wise (same shape as the embedding table, for example, Adam's exp_avg), or row-wise (same length as the embedding hashsize, for example, rowwise_adagrad's momentum/sum). However, there may be other formats, a single value for each table. Specifically, for Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb and GWD, the
iternumber is a single value tensor, which cannot be tracked and checkpointed properly (this also means that there is a bug in Adam/Partial_rowwise_adam/Lamb/Partial_rowwise_lamb usages!)Here we support tracking and checkpointing single-value states, by constructing ShardMetadata with rowwise-sharding and replicating the single-value for each Sharded param (this is similar to how the rowwise state for colume-wise sharded tables are concatenated along row-dim).
By doing so, single-value
itercan be properly checkpointed just like other states, ensuring correct reloading of states and continuous training.Differential Revision: D63909559