Adds threading support in torchrec train pipeline. #1694
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
fenghuizhang
force-pushed
the
export-D53453429
branch
from
02b6829 to
b295c12
Compare
|
This pull request was exported from Phabricator. Differential Revision: D53453429 |
fenghuizhang
force-pushed
the
export-D53453429
branch
from
b295c12 to
e7b4fa7
Compare
fenghuizhang
pushed a commit
to fenghuizhang/torchrec
that referenced
this pull request
Feb 8, 2024
Summary: Motivation - In training, we have mostly been focused on optimizing for better GPU utilization. For models that are not GPU bound, we often observe that CPU ops taking nontrivial amount of time. With existing pipelines, these operations are executed in the main thread. Depending on the complexity of the model input, they can take tens of milliseconds in our traces. - There are certain applications that are latency sensitive. While the multi-stage pipelines we have improve throughput greatly, they hurt latency by buffering multiple batches in the pipeline. In this change we add the capability to load data and copy it to gpu in a background thread. This helps reduce the iteration latency for the models we mentioned above, and minimizes the number of batches in the pipeline. In this diff, we are adding a new eval (forwardonly) sparse-data-dist pipeline with threading enabled. Reviewed By: dstaay-fb, leitian Differential Revision: D53453429
|
This pull request was exported from Phabricator. Differential Revision: D53453429 |
fenghuizhang
pushed a commit
to fenghuizhang/torchrec
that referenced
this pull request
Feb 8, 2024
Summary: Motivation - In training, we have mostly been focused on optimizing for better GPU utilization. For models that are not GPU bound, we often observe that CPU ops taking nontrivial amount of time. With existing pipelines, these operations are executed in the main thread. Depending on the complexity of the model input, they can take tens of milliseconds in our traces. - There are certain applications that are latency sensitive. While the multi-stage pipelines we have improve throughput greatly, they hurt latency by buffering multiple batches in the pipeline. In this change we add the capability to load data and copy it to gpu in a background thread. This helps reduce the iteration latency for the models we mentioned above, and minimizes the number of batches in the pipeline. In this diff, we are adding a new eval (forwardonly) sparse-data-dist pipeline with threading enabled. Reviewed By: dstaay-fb, leitian Differential Revision: D53453429
fenghuizhang
force-pushed
the
export-D53453429
branch
from
e7b4fa7 to
277a022
Compare
|
This pull request was exported from Phabricator. Differential Revision: D53453429 |
Summary: Motivation - In training, we have mostly been focused on optimizing for better GPU utilization. For models that are not GPU bound, we often observe that CPU ops taking nontrivial amount of time. With existing pipelines, these operations are executed in the main thread. Depending on the complexity of the model input, they can take tens of milliseconds in our traces. - There are certain applications that are latency sensitive. While the multi-stage pipelines we have improve throughput greatly, they hurt latency by buffering multiple batches in the pipeline. In this change we add the capability to load data and copy it to gpu in a background thread. This helps reduce the iteration latency for the models we mentioned above, and minimizes the number of batches in the pipeline. In this diff, we are adding a new eval (forwardonly) sparse-data-dist pipeline with threading enabled. Reviewed By: dstaay-fb, leitian, joshuadeng Differential Revision: D53453429
fenghuizhang
force-pushed
the
export-D53453429
branch
from
277a022 to
2d7e17d
Compare
|
This pull request was exported from Phabricator. Differential Revision: D53453429 |
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:
Motivation
In this change we add the capability to load data and copy it to gpu in a background thread. This helps reduce the iteration latency for the models we mentioned above, and minimizes the number of batches in the pipeline.
In this diff, we are adding a new eval (forwardonly) sparse-data-dist pipeline with threading enabled.
Reviewed By: dstaay-fb
Differential Revision: D53453429