Debug test failure for separate I, W execution #138863
Closed
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/138863
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 6 New FailuresAs of commit 5d5f6b5 with merge base failed to retrieve merge base, please contact dev infra: NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Oct 24, 2024
pytorch-bot
bot
added
the
oncall: distributed
Contributor
This PR needs a
|
wconstab
changed the title
This was referenced Oct 25, 2024
wconstab
changed the title
H-Huang
added a commit
that referenced
this pull request
Oct 28, 2024
Ran into issues (#138863) when adding a Schedule with single stage for zero bubble, adding code to support this mostly for test purposes cc awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o [ghstack-poisoned]
H-Huang
added a commit
that referenced
this pull request
Oct 28, 2024
Ran into issues (#138863) when adding a Schedule with single stage for zero bubble, adding code to support this mostly for test purposes cc awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o [ghstack-poisoned]
H-Huang
added a commit
that referenced
this pull request
Oct 28, 2024
Ran into issues (#138863) when adding a Schedule with a single stage, so adding code to support this edge case (mostly for test purposes) cc awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o [ghstack-poisoned]
H-Huang
added a commit
that referenced
this pull request
Oct 28, 2024
Ran into issues (#138863) when adding a Schedule with a single stage, so adding code to support this edge case (mostly for test purposes) cc awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o [ghstack-poisoned]
pytorchmergebot
pushed a commit
that referenced
this pull request
Oct 30, 2024
Ran into issues (#138863) when adding a Schedule with a single stage, so adding code to support this edge case (mostly for test purposes) Pull Request resolved: #138925 Approved by: https://github.com/wconstab
rahulsingh-intel
pushed a commit
to rahulsingh-intel/pytorch
that referenced
this pull request
Nov 5, 2024
Ran into issues (pytorch#138863) when adding a Schedule with a single stage, so adding code to support this edge case (mostly for test purposes) Pull Request resolved: pytorch#138925 Approved by: https://github.com/wconstab
Esquains
pushed a commit
to Esquains/study1
that referenced
this pull request
Dec 15, 2024
Test added in prev PR runs a simple schedule on one rank and using one pp stage. The schedule has 2 microbatches and runs F, I, W ops separately for each microbatch. Then the test compares .grad for the pipelined module to a reference module. The test fails becuase the pipelined module has 'None' grads. The debug logs show that the executor is running the I and W steps, and additional print statements show that the underlying backward_ functions are being called as expected. I identified that the 'param_groups' saved during backward_input were empty-list, and I think that is why there are no gradients computed during backward. This contradicts one comment in the code. There is a special case for stage 0 to execute 'full' backward instead of 'weight' backward when running 'backward_weight_one_chunk. I tried running both the if and else branch of this logic with the same result, 'None' grads. The special case seems to fit my case- I only have 1 stage, so it is stage 0. But running 'full' backward for dW does not seem to help. repro: `TORCH_LOGS=+pp python test/distributed/pipelining/test_schedule.py -k test_grad_with_split_b_w` ghstack-source-id: ef0fc80 Pull Request resolved: pytorch/pytorch#138863
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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):
Test added in prev PR runs a simple schedule on one rank and using one
pp stage. The schedule has 2 microbatches and runs F, I, W ops
separately for each microbatch. Then the test compares .grad for the
pipelined module to a reference module.
The test fails becuase the pipelined module has 'None' grads. The debug
logs show that the executor is running the I and W steps, and additional
print statements show that the underlying backward_ functions are being
called as expected.
I identified that the 'param_groups' saved during backward_input were
empty-list, and I think that is why there are no gradients computed
during backward. This contradicts one comment in the code. There is a
special case for stage 0 to execute 'full' backward instead of 'weight'
backward when running 'backward_weight_one_chunk. I tried running both
the if and else branch of this logic with the same result, 'None' grads.
The special case seems to fit my case- I only have 1 stage, so it is
stage 0. But running 'full' backward for dW does not seem to help.
repro:
TORCH_LOGS=+pp python test/distributed/pipelining/test_schedule.py -k test_grad_with_split_b_w