Use unique cache locations in backward for pipeline prefetching #2164
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
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D51488959 |
sryap
force-pushed
the
export-D51488959
branch
from
4ba492d to
5148659
Compare
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Dec 14, 2023
…rch#2164) Summary: When pipeline prefetching is enabled (`prefetch_pipeline=True`) for `EmbeddingLocation.MANAGED_CACHING`, TBE has to update `lxu_cache_locations` to ensure cache consistency. Prior to this diff, TBE performs the full cache lookup when updating `lxu_cache_locations` (i.e., looking up all indices although they are duplicated). The time complexity of such the lookup grows with the number of indices. Thus, the lookup can be expensive when the number of indices is large. This diff addresses this problem by looking up only the unique indices (which is normally much smaller than the full indices). The number of unique indices tends to stay more or less the same even when the total number of indices grows. Thus, looking up only unique indices can reduce cost of cache lookup effectively. The unique `lxu_cache_locations` are fed directly to TBE backward to consume. Thus, there is no decompression cost. Reviewed By: ehsanardestani, r-barnes Differential Revision: D51488959
|
This pull request was exported from Phabricator. Differential Revision: D51488959 |
2 similar comments
|
This pull request was exported from Phabricator. Differential Revision: D51488959 |
|
This pull request was exported from Phabricator. Differential Revision: D51488959 |
…rch#2164) Summary: When pipeline prefetching is enabled (`prefetch_pipeline=True`) for `EmbeddingLocation.MANAGED_CACHING`, TBE has to update `lxu_cache_locations` to ensure cache consistency. Prior to this diff, TBE performs the full cache lookup when updating `lxu_cache_locations` (i.e., looking up all indices although they are duplicated). The time complexity of such the lookup grows with the number of indices. Thus, the lookup can be expensive when the number of indices is large. This diff addresses this problem by looking up only the unique indices (which is normally much smaller than the full indices). The number of unique indices tends to stay more or less the same even when the total number of indices grows. Thus, looking up only unique indices can reduce cost of cache lookup effectively. The unique `lxu_cache_locations` are fed directly to TBE backward to consume. Thus, there is no decompression cost. Reviewed By: ehsanardestani, r-barnes Differential Revision: D51488959
sryap
force-pushed
the
export-D51488959
branch
from
5148659 to
1f831b0
Compare
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jan 5, 2024
…rch#2164) Summary: When pipeline prefetching is enabled (`prefetch_pipeline=True`) for `EmbeddingLocation.MANAGED_CACHING`, TBE has to update `lxu_cache_locations` to ensure cache consistency. Prior to this diff, TBE performs the full cache lookup when updating `lxu_cache_locations` (i.e., looking up all indices although they are duplicated). The time complexity of such the lookup grows with the number of indices. Thus, the lookup can be expensive when the number of indices is large. This diff addresses this problem by looking up only the unique indices (which is normally much smaller than the full indices). The number of unique indices tends to stay more or less the same even when the total number of indices grows. Thus, looking up only unique indices can reduce cost of cache lookup effectively. The unique `lxu_cache_locations` are fed directly to TBE backward to consume. Thus, there is no decompression cost. Reviewed By: ehsanardestani, r-barnes Differential Revision: D51488959
|
This pull request was exported from Phabricator. Differential Revision: D51488959 |
sryap
force-pushed
the
export-D51488959
branch
from
1f831b0 to
d4d8e25
Compare
|
This pull request was exported from Phabricator. Differential Revision: D51488959 |
|
This pull request has been merged in e92c30b. |
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:
When pipeline prefetching is enabled (
prefetch_pipeline=True) forEmbeddingLocation.MANAGED_CACHING, TBE has to updatelxu_cache_locationsto ensure cache consistency. Prior to thisdiff, TBE performs the full cache lookup when updating
lxu_cache_locations(i.e., looking up all indices although they areduplicated). The time complexity of such the lookup grows with
the number of indices. Thus, the lookup can be expensive when the
number of indices is large. This diff addresses this problem by
looking up only the unique indices (which is normally much smaller
than the full indices). The number of unique indices tends to stay
more or less the same even when the total number of indices grows.
Thus, looking up only unique indices can reduce cost of cache lookup
effectively. The unique
lxu_cache_locationsare fed directly to TBEbackward to consume. Thus, there is no decompression cost.
Reviewed By: ehsanardestani, r-barnes
Differential Revision: D51488959