Enable proper optimizer state storing + Test between batches #3053
Conversation
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
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D75565054 |
aporialiao
added a commit
to aporialiao/torchrec
that referenced
this pull request
Jun 6, 2025
…torch#3053) Summary: # Main Changes 1. Enable unit test with an adaptive optimizer `Adagrad` 1. Previously I tested the optimizer state with an optimizer `SGD` that is static throughout training so didn't , instead here I used the `Adagrad` which exposed the previous implementation did not properly store optimziers. 2. Properly store optimizer state in `update_optimizer_state` 2. Append optimizer tensors as inputs to the all2all call, then parse through the output tensors to store the right tensors. 2. Optimizer tensors that did not need to be sent to a new rank are persisted and resaved. 2. After new lookups are created, use `load_state_dict` to load in the saved optimizer state to the current optimizers. 3. Helpers & other small changes 3. Helper to compare optimizer tensors for unit tests 3. Update `DMP` reshard - optimizer saving to match the same fqn Differential Revision: D75565054
aporialiao
force-pushed
the
export-D75565054
branch
from
6822ac4 to
afb8577
Compare
|
This pull request was exported from Phabricator. Differential Revision: D75565054 |
aporialiao
added a commit
to aporialiao/torchrec
that referenced
this pull request
Jun 6, 2025
…torch#3053) Summary: # Main Changes 1. Enable unit test with an adaptive optimizer `Adagrad` 1. Previously I tested the optimizer state with an optimizer `SGD` that is static throughout training so didn't actually test if we stored opt state, instead here I used the `Adagrad` which exposed the previous implementation did not properly store optimziers. 2. Properly store optimizer state in `update_optimizer_state` 2. Append optimizer tensors as inputs to the all2all call, then parse through the output tensors to store the right tensors. 2. Optimizer tensors that did not need to be sent to a new rank are persisted and resaved. 2. After new lookups are created, use `load_state_dict` to load in the saved optimizer state to the current optimizers. 3. Helpers & other small changes 3. Helper to compare optimizer tensors for unit tests 3. Update `DMP` reshard - optimizer saving to match the same fqn Reviewed By: aliafzal Differential Revision: D75565054
aporialiao
force-pushed
the
export-D75565054
branch
from
afb8577 to
fd4f611
Compare
|
This pull request was exported from Phabricator. Differential Revision: D75565054 |
aporialiao
added a commit
to aporialiao/torchrec
that referenced
this pull request
Jun 6, 2025
…torch#3053) Summary: # Main Changes 1. Enable unit test with an adaptive optimizer `Adagrad` 1. Previously I tested the optimizer state with an optimizer `SGD` that is static throughout training so didn't actually test if we stored opt state, instead here I used the `Adagrad` which exposed the previous implementation did not properly store optimziers. 2. Properly store optimizer state in `update_optimizer_state` 2. Append optimizer tensors as inputs to the all2all call, then parse through the output tensors to store the right tensors. 2. Optimizer tensors that did not need to be sent to a new rank are persisted and resaved. 2. After new lookups are created, use `load_state_dict` to load in the saved optimizer state to the current optimizers. 3. Helpers & other small changes 3. Helper to compare optimizer tensors for unit tests 3. Update `DMP` reshard - optimizer saving to match the same fqn Reviewed By: aliafzal Differential Revision: D75565054
aporialiao
force-pushed
the
export-D75565054
branch
from
fd4f611 to
4a7c601
Compare
|
This pull request was exported from Phabricator. Differential Revision: D75565054 |
…torch#3053) Summary: Pull Request resolved: meta-pytorch#3053 # Main Changes 1. Enable unit test with an adaptive optimizer `Adagrad` 1. Previously I tested the optimizer state with an optimizer `SGD` that is static throughout training so didn't actually test if we stored opt state, instead here I used the `Adagrad` which exposed the previous implementation did not properly store optimziers. 2. Properly store optimizer state in `update_optimizer_state` 2. Append optimizer tensors as inputs to the all2all call, then parse through the output tensors to store the right tensors. 2. Optimizer tensors that did not need to be sent to a new rank are persisted and resaved. 2. After new lookups are created, use `load_state_dict` to load in the saved optimizer state to the current optimizers. 3. Helpers & other small changes 3. Helper to compare optimizer tensors for unit tests 3. Update `DMP` reshard - optimizer saving to match the same fqn Reviewed By: aliafzal Differential Revision: D75565054
|
This pull request was exported from Phabricator. Differential Revision: D75565054 |
aporialiao
force-pushed
the
export-D75565054
branch
from
4a7c601 to
f706b80
Compare
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:
Main Changes
AdagradSGDthat is static throughout training so didn't , instead here I used theAdagradwhich exposed the previous implementation did not properly store optimziers.update_optimizer_state2. Append optimizer tensors as inputs to the all2all call, then parse through the output tensors to store the right tensors.
2. Optimizer tensors that did not need to be sent to a new rank are persisted and resaved.
2. After new lookups are created, use
load_state_dictto load in the saved optimizer state to the current optimizers.3. Helper to compare optimizer tensors for unit tests
3. Update
DMPreshard - optimizer saving to match the same fqnDifferential Revision: D75565054