add mean pooling divisor to cuda stream #1863
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: D55945969 |
|
This pull request was exported from Phabricator. Differential Revision: D55945969 |
zainhuda
pushed a commit
to zainhuda/torchrec
that referenced
this pull request
Apr 10, 2024
Summary: Pull Request resolved: pytorch#1863 Initial mean pooling implementation was not attending to the appropriate CUDA stream properly with respect to train pipeline. We now register the divisor tensor into the CUDA stream in context. The key insight: Tensors used on a different stream than their origin, the memory allocator may reuse the memory unexpectedly. We also split the callback function into two (create divisor, apply mean pooling). Change the context from holding a callable to divisor tensor instead. This is because recording non tensors into a CUDA stream is non trivial, whereas recording a tensor into a CUDA stream is easily supported. This has no perf regressions from the original implementation nor lack of clarity. Differential Revision: D55945969
zainhuda
force-pushed
the
export-D55945969
branch
from
c46a991 to
d091d11
Compare
|
This pull request was exported from Phabricator. Differential Revision: D55945969 |
zainhuda
pushed a commit
to zainhuda/torchrec
that referenced
this pull request
Apr 10, 2024
Summary: Pull Request resolved: pytorch#1863 Initial mean pooling implementation was not attending to the appropriate CUDA stream properly with respect to train pipeline. We now register the divisor tensor into the CUDA stream in context. The key insight: Tensors used on a different stream than their origin, the memory allocator may reuse the memory unexpectedly. We also split the callback function into two (create divisor, apply mean pooling). Change the context from holding a callable to divisor tensor instead. This is because recording non tensors into a CUDA stream is non trivial, whereas recording a tensor into a CUDA stream is easily supported. This has no perf regressions from the original implementation nor lack of clarity. Differential Revision: D55945969
zainhuda
force-pushed
the
export-D55945969
branch
from
d091d11 to
0c19087
Compare
zainhuda
pushed a commit
to zainhuda/torchrec
that referenced
this pull request
Apr 10, 2024
Summary: Initial mean pooling implementation was not attending to the appropriate CUDA stream properly with respect to train pipeline. We now register the divisor tensor into the CUDA stream in context. The key insight: Tensors used on a different stream than their origin, the memory allocator may reuse the memory unexpectedly. We also split the callback function into two (create divisor, apply mean pooling). Change the context from holding a callable to divisor tensor instead. This is because recording non tensors into a CUDA stream is non trivial, whereas recording a tensor into a CUDA stream is easily supported. This has no perf regressions from the original implementation nor lack of clarity. Reviewed By: joshuadeng Differential Revision: D55945969
zainhuda
force-pushed
the
export-D55945969
branch
from
0c19087 to
2ca3a1b
Compare
|
This pull request was exported from Phabricator. Differential Revision: D55945969 |
Summary: Initial mean pooling implementation was not attending to the appropriate CUDA stream properly with respect to train pipeline. We now register the divisor tensor into the CUDA stream in context. The key insight: Tensors used on a different stream than their origin, the memory allocator may reuse the memory unexpectedly. We also split the callback function into two (create divisor, apply mean pooling). Change the context from holding a callable to divisor tensor instead. This is because recording non tensors into a CUDA stream is non trivial, whereas recording a tensor into a CUDA stream is easily supported. This has no perf regressions from the original implementation nor lack of clarity. Reviewed By: joshuadeng Differential Revision: D55945969
zainhuda
force-pushed
the
export-D55945969
branch
from
2ca3a1b to
ba83f32
Compare
|
This pull request was exported from Phabricator. Differential Revision: D55945969 |
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:
Initial mean pooling implementation was not attending to the appropriate CUDA stream properly with respect to train pipeline. We now register the divisor tensor into the CUDA stream in context.
The key insight:
Tensors used on a different stream than their origin, the memory allocator may reuse the memory unexpectedly.
We also split the callback function into two (create divisor, apply mean pooling). Change the context from holding a callable to divisor tensor instead. This is because recording non tensors into a CUDA stream is non trivial, whereas recording a tensor into a CUDA stream is easily supported. This has no perf regressions from the original implementation nor lack of clarity.
Differential Revision: D55945969