float8 training: fix bug with AC + compile #1329
Merged
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
Summary: In #1306 I accidentally broke torchtitan + float8 + AC + compile. I don't have a non-torchtitan repro now, putting up the fix first to ensure torchtitan still works, and we should follow-up later with adding test coverage to torchao to prevent similar breakages in the future. What broke: * in the forward of `Float8Linear`, we were setting an attribute on the module * ^ is not supported with compile + something how torchtitan specifically calls AC The fix: remove this attribute setting altogether. Unfortunately this breaks an edge case feature for ensuring scales are reprensentable in `float16`. Since `float16` training is not commonly used with `float8` and this feature was added during very early testing, removing this for now is fine. If we need to add this feature back in the future, I'd advocate for doing it via explicit configuration such as `config.set_scale_upper_bound` and avoiding the stateful hacks, which are usually not compiler friendly. Test Plan: ``` // this repo ./test/float8/test_everything.sh // torchtitan - broken before this PR, works after this PR with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ``` Reviewers: Subscribers: Tasks: Tags:
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1329
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: ✅ No FailuresAs of commit a8ccff4 with merge base 7489c7d ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
vkuzo
added
topic: bug fix
drisspg
approved these changes
Nov 22, 2024
weifengpy
approved these changes
Nov 22, 2024
facebook-github-bot
added
the
CLA Signed
sunjiweiswift
pushed a commit
to sunjiweiswift/ao
that referenced
this pull request
Nov 25, 2024
Summary: In pytorch#1306 I accidentally broke torchtitan + float8 + AC + compile. I don't have a non-torchtitan repro now, putting up the fix first to ensure torchtitan still works, and we should follow-up later with adding test coverage to torchao to prevent similar breakages in the future. What broke: * in the forward of `Float8Linear`, we were setting an attribute on the module * ^ is not supported with compile + something how torchtitan specifically calls AC The fix: remove this attribute setting altogether. Unfortunately this breaks an edge case feature for ensuring scales are reprensentable in `float16`. Since `float16` training is not commonly used with `float8` and this feature was added during very early testing, removing this for now is fine. If we need to add this feature back in the future, I'd advocate for doing it via explicit configuration such as `config.set_scale_upper_bound` and avoiding the stateful hacks, which are usually not compiler friendly. Test Plan: ``` // this repo ./test/float8/test_everything.sh // torchtitan - broken before this PR, works after this PR with-proxy CONFIG_FILE="./train_configs/llama3_8b.toml" ./run_llama_train.sh --float8.enable_float8_linear --training.compile ``` 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:
In #1306 I accidentally broke torchtitan + float8 + AC + compile.
I don't have a non-torchtitan repro now, putting up the fix first to ensure torchtitan still works, and we should follow-up later with adding test coverage to torchao to prevent similar breakages in the future.
What broke:
Float8Linear, we were setting an attribute on the moduleThe fix: remove this attribute setting altogether. Unfortunately this breaks an edge case feature for ensuring scales are reprensentable in
float16. Sincefloat16training is not commonly used withfloat8and this feature was added during very early testing, removing this for now is fine.If we need to add this feature back in the future, I'd advocate for doing it via explicit configuration such as
config.set_scale_upper_boundand avoiding the stateful hacks, which are usually not compiler friendly.Test Plan:
Reviewers:
Subscribers:
Tasks:
Tags: