Add an option to use fp8-all-gather only without fp8 computation. #1093
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/pytorch/ao/1093
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 5c5f4f2 with merge base ae77f40 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D63056142 |
weifengpy
approved these changes
Oct 21, 2024
|
CI error looks relevant. maybe take a look before landing? |
y-sq
force-pushed
the
export-D63056142
branch
from
7857bb1 to
2950a2e
Compare
y-sq
added a commit
to y-sq/ao
that referenced
this pull request
Oct 31, 2024
…torch#1093) Summary: The implementation reuses `WeightWithDynamicFloat8CastTensor` class and the `Float8Linear` module. I added an if-else branch in the existing `Float8Linear` module to re-use our existing logics to handle different casting cases, such as pre-/post-forward for delayed scaling, pre-compute amax for fp8-all-gather. Reviewed By: weifengpy Differential Revision: D63056142
|
This pull request was exported from Phabricator. Differential Revision: D63056142 |
y-sq
added a commit
to y-sq/ao
that referenced
this pull request
Oct 31, 2024
…torch#1093) Summary: The implementation reuses `WeightWithDynamicFloat8CastTensor` class and the `Float8Linear` module. I added an if-else branch in the existing `Float8Linear` module to re-use our existing logics to handle different casting cases, such as pre-/post-forward for delayed scaling, pre-compute amax for fp8-all-gather. Reviewed By: weifengpy Differential Revision: D63056142
y-sq
force-pushed
the
export-D63056142
branch
from
2950a2e to
0f3ef23
Compare
|
This pull request was exported from Phabricator. Differential Revision: D63056142 |
…torch#1093) Summary: The implementation reuses `WeightWithDynamicFloat8CastTensor` class and the `Float8Linear` module. I added an if-else branch in the existing `Float8Linear` module to re-use our existing logics to handle different casting cases, such as pre-/post-forward for delayed scaling, pre-compute amax for fp8-all-gather. Reviewed By: weifengpy Differential Revision: D63056142
y-sq
force-pushed
the
export-D63056142
branch
from
0f3ef23 to
5c5f4f2
Compare
|
This pull request was exported from Phabricator. Differential Revision: D63056142 |
yanbing-j
pushed a commit
to yanbing-j/ao
that referenced
this pull request
Dec 9, 2024
vkuzo
added a commit
that referenced
this pull request
Dec 19, 2024
…training Summary: In #1093 we added a config option, off by default, to use only float8 all-gather for training and do the matrix multiply in high precision. This seems generally useful for communication bound workloads, but we can probably think of a cleaner way to add this functionality (such as a weight wrapper tensor subclass). The current implementation adds non-trivial complexity and doesn't jive well with where we want to take this codebase. Since no one is using this internally or externally yet and we haven't talked about it in the release notes, I think we should do a BC-breaking delete as a one-off. However, if people have concerns - let me know and we can talk about less aggressive options. Test Plan: ``` ./test/float8/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags:
vkuzo
added a commit
that referenced
this pull request
Dec 19, 2024
…training Summary: In #1093 we added a config option, off by default, to use only float8 all-gather for training and do the matrix multiply in high precision. This seems generally useful for communication bound workloads, but we can probably think of a cleaner way to add this functionality (such as a weight wrapper tensor subclass). The current implementation adds non-trivial complexity and doesn't jive well with where we want to take this codebase. Since no one is using this internally or externally yet and we haven't talked about it in the release notes, I think we should do a BC-breaking delete as a one-off. However, if people have concerns - let me know and we can talk about less aggressive options. Test Plan: ``` ./test/float8/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags:
vkuzo
added a commit
that referenced
this pull request
Dec 20, 2024
#1451) for now, delete the float8-all-gather-only functionality from float8 training Summary: In #1093 we added a config option, off by default, to use only float8 all-gather for training and do the matrix multiply in high precision. This seems generally useful for communication bound workloads, but we can probably think of a cleaner way to add this functionality (such as a weight wrapper tensor subclass). The current implementation adds non-trivial complexity and doesn't jive well with where we want to take this codebase. Since no one is using this internally or externally yet and we haven't talked about it in the release notes, I think we should do a BC-breaking delete as a one-off. However, if people have concerns - let me know and we can talk about less aggressive options. Test Plan: ``` ./test/float8/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags:
amdfaa
pushed a commit
that referenced
this pull request
Jan 10, 2025
#1451) for now, delete the float8-all-gather-only functionality from float8 training Summary: In #1093 we added a config option, off by default, to use only float8 all-gather for training and do the matrix multiply in high precision. This seems generally useful for communication bound workloads, but we can probably think of a cleaner way to add this functionality (such as a weight wrapper tensor subclass). The current implementation adds non-trivial complexity and doesn't jive well with where we want to take this codebase. Since no one is using this internally or externally yet and we haven't talked about it in the release notes, I think we should do a BC-breaking delete as a one-off. However, if people have concerns - let me know and we can talk about less aggressive options. Test Plan: ``` ./test/float8/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags:
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:
The implementation reuses
WeightWithDynamicFloat8CastTensorclass and theFloat8Linearmodule.I added an if-else branch in the existing
Float8Linearmodule to re-use our existing logics to handle different casting cases, such as pre-/post-forward for delayed scaling, pre-compute amax for fp8-all-gather.Differential Revision: D63056142