optimization: move set_metadata to ssd_eviction_stream #5082
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 failed.
|
Contributor
emlin
force-pushed
the
export-D86013406
branch
from
54da6bc to
dfeac9d
Compare
emlin
added a commit
to emlin/FBGEMM
that referenced
this pull request
Nov 4, 2025
Summary: X-link: facebookresearch/FBGEMM#2090 with feature score eviction, tbe will call backend to update feature score metadata separately in forward pass. this process is designed for asynchronous update without blocking forward/backward pass, however the cpu blocking operation blocked the main stream, so after get_cuda, all2all cannot be started immediately. from dummy profile, we can see this trace: {F1983224804} the set metadata operation becomes a blocker in critical path, which took 217ms With this change, we can see the trace is updated to: {F1983224830} where overall prefetch is reduced to less than 70ms, also the get_cuda is followed by all2all immediately without other waiting and stream sync Differential Revision: D86013406
emlin
force-pushed
the
export-D86013406
branch
from
dfeac9d to
5c69428
Compare
emlin
added a commit
to emlin/FBGEMM
that referenced
this pull request
Nov 7, 2025
Summary: X-link: facebookresearch/FBGEMM#2090 with feature score eviction, tbe will call backend to update feature score metadata separately in forward pass. this process is designed for asynchronous update without blocking forward/backward pass, however the cpu blocking operation blocked the main stream, so after get_cuda, all2all cannot be started immediately. from dummy profile, we can see this trace: {F1983224804} the set metadata operation becomes a blocker in critical path, which took 217ms With this change, we can see the trace is updated to: {F1983224830} where overall prefetch is reduced to less than 70ms, also the get_cuda is followed by all2all immediately without other waiting and stream sync Reviewed By: steven1327, kathyxuyy Differential Revision: D86013406
Summary: X-link: facebookresearch/FBGEMM#2090 with feature score eviction, tbe will call backend to update feature score metadata separately in forward pass. this process is designed for asynchronous update without blocking forward/backward pass, however the cpu blocking operation blocked the main stream, so after get_cuda, all2all cannot be started immediately. from dummy profile, we can see this trace: {F1983224804} the set metadata operation becomes a blocker in critical path, which took 217ms With this change, we can see the trace is updated to: {F1983224830} where overall prefetch is reduced to less than 70ms, also the get_cuda is followed by all2all immediately without other waiting and stream sync https://www.internalfb.com/ai_infra/zoomer/profiling-run/overview?profilingRunID=1913270729575721 Reviewed By: steven1327, kathyxuyy Differential Revision: D86013406
emlin
force-pushed
the
export-D86013406
branch
from
5c69428 to
f716372
Compare
Contributor
|
This pull request has been merged in ee26ed4. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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:
X-link: https://github.com/facebookresearch/FBGEMM/pull/2090
with feature score eviction, tbe will call backend to update feature score metadata separately in forward pass.
this process is designed for asynchronous update without blocking forward/backward pass, however the cpu blocking operation blocked the main stream, so after get_cuda, all2all cannot be started immediately.
from dummy profile, we can see this trace:
{F1983224804}
the set metadata operation becomes a blocker in critical path, which took 217ms
With this change, we can see the trace is updated to:
{F1983224830}
where overall prefetch is reduced to less than 70ms, also the get_cuda is followed by all2all immediately without other waiting and stream sync
Differential Revision: D86013406