Make metrics logging work for pipeline parallelism #383
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
wconstab
added a commit
that referenced
this pull request
Jun 3, 2024
Avoid complicating the ux and leave the status quo of 2 user-selectable behaviors: - log from rank 0 (the default) - log from all ranks (not the default) Modify the meaning of 'log from rank 0' to log from rank 0 in non-pipeline parallel runs, and log from the local rank 0 within the last pipeline-parallel stage group if pp is enabled. (note: earlier pipeline stages still produce some metrics like mfu/memory, but do not compute loss.) ghstack-source-id: 10c6d13f6820642995a09585b12aa5f7da0b2165 Pull Request resolved: #383
facebook-github-bot
added
the
CLA Signed
awgu
reviewed
Jun 4, 2024
train.py
Outdated
| metric_logger = build_metric_logger(job_config) | ||
| if parallel_dims.pp_enabled: | ||
| pp_size = pp_mesh.size() | ||
| metrics_log_rank = int((world_mesh.size() // pp_size) * (pp_size - 1)) |
There was a problem hiding this comment.
Does this rank computation assume that PP is outermost? If so, should we assert/check for that?
There was a problem hiding this comment.
yea, it does. I wonder if there is a better place to do the assert. I'll add it here for now.
I also don't like to do it this way, i might want to propose adding a device-mesh API to help do a calculation like this more robustly.
There was a problem hiding this comment.
i moved this into a util and added an assert.
wconstab
added a commit
that referenced
this pull request
Jun 4, 2024
Avoid complicating the ux and leave the status quo of 2 user-selectable behaviors: - log from rank 0 (the default) - log from all ranks (not the default) Modify the meaning of 'log from rank 0' to log from rank 0 in non-pipeline parallel runs, and log from the local rank 0 within the last pipeline-parallel stage group if pp is enabled. (note: earlier pipeline stages still produce some metrics like mfu/memory, but do not compute loss.) ghstack-source-id: 740e06906e079e366cc6966b549cc3e1fcebcc69 Pull Request resolved: #383
wconstab
added a commit
that referenced
this pull request
Jun 4, 2024
Avoid complicating the ux and leave the status quo of 2 user-selectable behaviors: - log from rank 0 (the default) - log from all ranks (not the default) Modify the meaning of 'log from rank 0' to log from rank 0 in non-pipeline parallel runs, and log from the local rank 0 within the last pipeline-parallel stage group if pp is enabled. (note: earlier pipeline stages still produce some metrics like mfu/memory, but do not compute loss.) ghstack-source-id: 7f60d1045f240327ae41ade3a353aff19d2f289a Pull Request resolved: #383
fegin
approved these changes
Jun 4, 2024
yes, my thoughts exactly. I would like to discuss this offline. I couldn't quickly think of what the best API proposal for devicemesh would be so i went this route instead. |
wconstab
added a commit
that referenced
this pull request
Jun 4, 2024
Avoid complicating the ux and leave the status quo of 2 user-selectable behaviors: - log from rank 0 (the default) - log from all ranks (not the default) Modify the meaning of 'log from rank 0' to log from rank 0 in non-pipeline parallel runs, and log from the local rank 0 within the last pipeline-parallel stage group if pp is enabled. (note: earlier pipeline stages still produce some metrics like mfu/memory, but do not compute loss.) ghstack-source-id: 7f60d1045f240327ae41ade3a353aff19d2f289a Pull Request resolved: #383
tianyu-l
pushed a commit
to tianyu-l/torchtitan_intern24
that referenced
this pull request
Aug 16, 2024
Avoid complicating the ux and leave the status quo of 2 user-selectable behaviors: - log from rank 0 (the default) - log from all ranks (not the default) Modify the meaning of 'log from rank 0' to log from rank 0 in non-pipeline parallel runs, and log from the local rank 0 within the last pipeline-parallel stage group if pp is enabled. (note: earlier pipeline stages still produce some metrics like mfu/memory, but do not compute loss.) ghstack-source-id: 7f60d1045f240327ae41ade3a353aff19d2f289a Pull Request resolved: pytorch#383
philippguevorguian
pushed a commit
to YerevaNN/YNNtitan
that referenced
this pull request
Aug 17, 2024
Avoid complicating the ux and leave the status quo of 2 user-selectable behaviors: - log from rank 0 (the default) - log from all ranks (not the default) Modify the meaning of 'log from rank 0' to log from rank 0 in non-pipeline parallel runs, and log from the local rank 0 within the last pipeline-parallel stage group if pp is enabled. (note: earlier pipeline stages still produce some metrics like mfu/memory, but do not compute loss.) ghstack-source-id: 7f60d1045f240327ae41ade3a353aff19d2f289a Pull Request resolved: pytorch#383
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):
Avoid complicating the ux and leave the status quo of 2 user-selectable
behaviors:
Modify the meaning of 'log from rank 0' to log from rank 0 in
non-pipeline parallel runs, and log from the local rank 0 within the
last pipeline-parallel stage group if pp is enabled. (note: earlier
pipeline stages still produce some metrics like mfu/memory, but do not
compute loss.)