[BE] dump compile trace to CI output for debugging, and reduce uploaded folder size in CI #739
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
[ghstack-poisoned]
tianyu-l
added a commit
that referenced
this pull request
Dec 16, 2024
ghstack-source-id: b8ca3dc898a688e89edfc64d0417ebe456d0f394 Pull Request resolved: #739
facebook-github-bot
added
the
CLA Signed
[ghstack-poisoned]
This was referenced Dec 16, 2024
tianyu-l
commented
Dec 16, 2024
test_runner.py
Outdated
Comment on lines
64
to
75
| # TODO: temporarily disabling to let CI be able to run other tests | ||
| # OverrideDefinitions( | ||
| # [ | ||
| # [ | ||
| # "--training.compile", | ||
| # "--activation_checkpoint.mode selective", | ||
| # "--activation_checkpoint.selective_ac_option op", | ||
| # ], | ||
| # ], | ||
| # "1D compile with selective op AC", | ||
| # "1d_compile_sac_op", | ||
| # ), |
There was a problem hiding this comment.
I'll revert this temporary change before merging
tianyu-l
changed the title
fegin
approved these changes
Dec 16, 2024
|
@tianyu-l just curious: the time per PR on CI is significantly impacted by 'upload_artifacts' step even before your change. I was wondering if at some point this is a bottleneck and we need to be more picky about which files we upload. On that note, do you know how large the compile-trace artifacts are and whether it is making a difference to PR CI time? |
as titled, so that we can download the compile-related code after seeing a CI failure [ghstack-poisoned]
tianyu-l
changed the title
…duce uploaded folder size in CI" as titled, so that we can download the compile-related code after seeing a CI failure To answer wconstab's question, I downloaded the artifact with & without compile trace, and here’s my findings: 1. The compile trace is comparable to profiler trace, about <10MB for a job. And there are only two tests with torch.compile. -> so it should be fine to enable uploading compile traces. 2. The biggest overhead is from checkpoint, each job about 300MB or more. I forgot to remove `enable_checkpoint` flag for PP jobs. -> I removed it for PP jobs, which saves a lot. 3. For checkpointing-related tests, we still need `enable_checkpoint`, but I suggest we modify the default interval from 5-steps to 10-steps -> so that we halve the overhead in those tests. 4. I think we can by default disable TB logging and profiler trace logging -> only enable them on 1 job should be fine. results: 1. 4 GPU running time reduced from ~17 minutes to 15m 18s. 2. uploaded folder size dropped from 2.7 GB to 309 MB. [ghstack-poisoned]
tianyu-l
added a commit
that referenced
this pull request
Dec 17, 2024
…kload ghstack-source-id: fe1076b088646dd1d3f3da4bb646c17bd3fa555d Pull Request resolved: #739
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.
Stack from ghstack (oldest at bottom):
as titled, so that we can download the compile-related code after seeing a CI failure
To answer @wconstab's question, I downloaded the artifact with & without compile trace, and here’s my findings:
enable_checkpointflag for PP jobs. -> I removed it for PP jobs, which saves a lot.enable_checkpoint, but I suggest we modify the default interval from 5-steps to 10-steps -> so that we halve the overhead in those tests.results: