Enabling Optimizer checkpointing for KeyValueEmbeddingFusedOptimizer #3248
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
meta-cla
bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D78131693 |
Raahul46
force-pushed
the
export-D78131693
branch
from
67cbbc2 to
5736cb1
Compare
Raahul46
pushed a commit
to Raahul46/torchrec
that referenced
this pull request
Aug 4, 2025
…ytorch#3248) Summary: **Context:** 1. We Introduced KeyValueEmbeddingFusedOptimizer for SSD Optimizer offloading https://www.internalfb.com/code/fbsource/[a43d796a4169]/fbcode/torchrec/distributed/batched_embedding_kernel.py?lines=341-376 But, currently the optimizer weights for SSD use-cases are not offloaded and are still on HBM Refer optimizer state dict CP: https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/minimal_viable_ai/core/model_family_api/optimizer.py?lines=1019-1028 Due to this, we want to initialize the optimizer class for SSD that allows us to get the latest optimizer weights values during checkpointing (get_optimizer_state call): https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/deeplearning/fbgemm/fbgemm_gpu/fbgemm_gpu/tbe/ssd/training.py?lines=2540-2545 **Hence, In this Diff:** We have made the following changes: 1. Loop through every embedding table a. Change the table placement to CPU b. Create a ShardedTensor for embedding weight c. Create a ShardedTensor for optimizer weight --> There are three cases for optimizers --> Single Optimizer Value per Shard --> Row-wise Optimizer value per Shard --> Point-wise Optimizer value per Shard and then initialize the optimizer class with the appropriate parameters Differential Revision: D78131693
Raahul46
pushed a commit
to Raahul46/torchrec
that referenced
this pull request
Aug 4, 2025
…ytorch#3248) Summary: Pull Request resolved: pytorch#3248 **Context:** 1. We Introduced KeyValueEmbeddingFusedOptimizer for SSD Optimizer offloading https://www.internalfb.com/code/fbsource/[a43d796a4169]/fbcode/torchrec/distributed/batched_embedding_kernel.py?lines=341-376 But, currently the optimizer weights for SSD use-cases are not offloaded and are still on HBM Refer optimizer state dict CP: https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/minimal_viable_ai/core/model_family_api/optimizer.py?lines=1019-1028 Due to this, we want to initialize the optimizer class for SSD that allows us to get the latest optimizer weights values during checkpointing (get_optimizer_state call): https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/deeplearning/fbgemm/fbgemm_gpu/fbgemm_gpu/tbe/ssd/training.py?lines=2540-2545 **Hence, In this Diff:** We have made the following changes: 1. Loop through every embedding table a. Change the table placement to CPU b. Create a ShardedTensor for embedding weight c. Create a ShardedTensor for optimizer weight --> There are three cases for optimizers --> Single Optimizer Value per Shard --> Row-wise Optimizer value per Shard --> Point-wise Optimizer value per Shard and then initialize the optimizer class with the appropriate parameters Differential Revision: D78131693
Raahul46
force-pushed
the
export-D78131693
branch
from
5736cb1 to
cc4260b
Compare
|
This pull request was exported from Phabricator. Differential Revision: D78131693 |
Raahul46
force-pushed
the
export-D78131693
branch
from
cc4260b to
797fa10
Compare
Raahul46
pushed a commit
to Raahul46/torchrec
that referenced
this pull request
Aug 5, 2025
…ytorch#3248) Summary: **Context:** 1. We Introduced KeyValueEmbeddingFusedOptimizer for SSD Optimizer offloading https://www.internalfb.com/code/fbsource/[a43d796a4169]/fbcode/torchrec/distributed/batched_embedding_kernel.py?lines=341-376 But, currently the optimizer weights for SSD use-cases are not offloaded and are still on HBM Refer optimizer state dict CP: https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/minimal_viable_ai/core/model_family_api/optimizer.py?lines=1019-1028 Due to this, we want to initialize the optimizer class for SSD that allows us to get the latest optimizer weights values during checkpointing (get_optimizer_state call): https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/deeplearning/fbgemm/fbgemm_gpu/fbgemm_gpu/tbe/ssd/training.py?lines=2540-2545 **Hence, In this Diff:** We have made the following changes: 1. Loop through every embedding table a. Change the table placement to CPU b. Create a ShardedTensor for embedding weight c. Create a ShardedTensor for optimizer weight --> There are three cases for optimizers --> Single Optimizer Value per Shard --> Row-wise Optimizer value per Shard --> Point-wise Optimizer value per Shard and then initialize the optimizer class with the appropriate parameters Differential Revision: D78131693
…ytorch#3248) Summary: Pull Request resolved: pytorch#3248 **Context:** 1. We Introduced KeyValueEmbeddingFusedOptimizer for SSD Optimizer offloading https://www.internalfb.com/code/fbsource/[a43d796a4169]/fbcode/torchrec/distributed/batched_embedding_kernel.py?lines=341-376 But, currently the optimizer weights for SSD use-cases are not offloaded and are still on HBM Refer optimizer state dict CP: https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/minimal_viable_ai/core/model_family_api/optimizer.py?lines=1019-1028 Due to this, we want to initialize the optimizer class for SSD that allows us to get the latest optimizer weights values during checkpointing (get_optimizer_state call): https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/deeplearning/fbgemm/fbgemm_gpu/fbgemm_gpu/tbe/ssd/training.py?lines=2540-2545 **Hence, In this Diff:** We have made the following changes: 1. Loop through every embedding table a. Change the table placement to CPU b. Create a ShardedTensor for embedding weight c. Create a ShardedTensor for optimizer weight --> There are three cases for optimizers --> Single Optimizer Value per Shard --> Row-wise Optimizer value per Shard --> Point-wise Optimizer value per Shard and then initialize the optimizer class with the appropriate parameters Differential Revision: D78131693
|
This pull request was exported from Phabricator. Differential Revision: D78131693 |
Raahul46
force-pushed
the
export-D78131693
branch
from
797fa10 to
e7480cb
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:
Context:
https://www.internalfb.com/code/fbsource/[a43d796a4169]/fbcode/torchrec/distributed/batched_embedding_kernel.py?lines=341-376
But, currently the optimizer weights for SSD use-cases are not offloaded and are still on HBM
Refer optimizer state dict CP:
https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/minimal_viable_ai/core/model_family_api/optimizer.py?lines=1019-1028
Due to this, we want to initialize the optimizer class for SSD that allows us to get the latest optimizer weights values during checkpointing (get_optimizer_state call):
https://www.internalfb.com/code/fbsource/[6303aefbae20]/fbcode/deeplearning/fbgemm/fbgemm_gpu/fbgemm_gpu/tbe/ssd/training.py?lines=2540-2545
Hence, In this Diff:
We have made the following changes:
a. Change the table placement to CPU
b. Create a ShardedTensor for embedding weight
c. Create a ShardedTensor for optimizer weight
--> There are three cases for optimizers
--> Single Optimizer Value per Shard
--> Row-wise Optimizer value per Shard
--> Point-wise Optimizer value per Shard
and then initialize the optimizer class with the appropriate parameters
Differential Revision: D78131693