[low-bit optim] Upcast everything to FP32 for internal calculations #1068
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1068
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 72ca834 with merge base e7b33bc ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
@msaroufim The failing low-bit optim tests pass now, but now CI timeouts 🤣 |
|
You can extend it to 2h -timeout: 60
+timeout: 120In https://github.com/pytorch/ao/blob/main/.github/workflows/regression_test.yml EDIT: I just made the change myself |
HDCharles
approved these changes
Oct 14, 2024
jainapurva
pushed a commit
that referenced
this pull request
Oct 15, 2024
…1068) * fix dtype * Update regression_test.yml --------- Co-authored-by: Mark Saroufim <[email protected]>
yanbing-j
pushed a commit
to yanbing-j/ao
that referenced
this pull request
Dec 9, 2024
This PR makes torchchat support multi-modality model definition and constructions. To show our power in multi-modality area, we integrate flamingo component into our system. Note that this is only for bare-minimum support for model definition. Please check openai_api_multimodal branch for e2e, and pytorch#1123 (comment) for better structure and llama3.1 support
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Outdated suggestions cannot be applied.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.
Fixes #1067.
Previously, it seems like torch.compile will not check for dtype mismatch when tensor subclass is used (e.g.
tensor_subclass_fp32.lerp(plain_tensor_bf16, weight)). Now it does, raising the error. To fix it, I simply cast everything to FP32.The dtype mismatch comes from the fact that my tensor subclasses for optim state have always used FP32 appearance dtype, even if param is BF16. This results in FP32 calculations, which is correct, though not originally intentional. Now I have made it explicit and intentional. This also means that BF16 param + BF16 optim state combination is now more accurate.
If I have time, I will re-run some some of the benchmarks to make sure things are alright.