PR #16975: Add a few related optimization passes for fp8 gemm custom-calls. #75959
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
![copybara-service[bot]](https://avatars.githubusercontent.com/in/44061?s=60&v=4){kind=small}
copybara-service
bot
force-pushed
the
exported_pr_675755585
branch
9 times, most recently
from
0c0136a
to
52d0410
Compare
…calls. Imported from GitHub PR openxla/xla#16975 This caused convergence issue for fp8 training, tested on GPT3 models: Before: ``` NETWORK BACKEND MATH SDPA XLA_EXTRAS GPUs STEPS/SEC LOSS WALLSECS GPT5B XLA fp8 FA 8 1.064 11.019 1571 [PAX STATUS]: Starting training loop. [PAX STATUS] step_i: 100, training loss: 11.015041 [PAX STATUS] step_i: 200, training loss: 11.016165 [PAX STATUS] step_i: 300, training loss: 11.016386 [PAX STATUS] step_i: 400, training loss: 11.014653 [PAX STATUS] step_i: 500, training loss: 11.014734 [PAX STATUS] step_i: 600, training loss: 11.01613 [PAX STATUS] step_i: 700, training loss: 11.009399 [PAX STATUS] step_i: 800, training loss: 11.017071 [PAX STATUS] step_i: 900, training loss: 11.014582 [PAX STATUS] step_i: 1000, training loss: 11.013434 [PAX STATUS] step_i: 1100, training loss: 11.021271 [PAX STATUS] step_i: 1200, training loss: 11.008364 [PAX STATUS] step_i: 1300, training loss: 11.0198145 [PAX STATUS] step_i: 1400, training loss: 11.01253 [PAX STATUS] step_i: 1500, training loss: 11.019016 ``` After: ``` NETWORK BACKEND MATH SDPA GPUs STEPS/SEC LOSS WALLSECS GPT5B XLA fp8 FA 8 1.020 3.797 1647 [PAX STATUS]: Starting training loop. [PAX STATUS] step_i: 100, training loss: 6.150083 [PAX STATUS] step_i: 200, training loss: 5.8871064 [PAX STATUS] step_i: 300, training loss: 5.4491887 [PAX STATUS] step_i: 400, training loss: 5.6384015 [PAX STATUS] step_i: 500, training loss: 5.273538 [PAX STATUS] step_i: 600, training loss: 5.2011905 [PAX STATUS] step_i: 700, training loss: 4.903013 [PAX STATUS] step_i: 800, training loss: 4.62972 [PAX STATUS] step_i: 900, training loss: 4.507727 [PAX STATUS] step_i: 1000, training loss: 4.625259 [PAX STATUS] step_i: 1100, training loss: 4.428066 [PAX STATUS] step_i: 1200, training loss: 4.252451 [PAX STATUS] step_i: 1300, training loss: 3.8448389 [PAX STATUS] step_i: 1400, training loss: 3.8578327 [PAX STATUS] step_i: 1500, training loss: 3.796958 ``` Copybara import of the project: -- 8bf6d19e5c10024d34a59b889893d203eee6691a by Elfie Guo <[email protected]>: Add a few related optimization pass for fp8 gemm rerwriter. Merging this change closes #16975 FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#16975 from elfiegg:pass 8bf6d19e5c10024d34a59b889893d203eee6691a PiperOrigin-RevId: 675755585
copybara-service
bot
force-pushed
the
exported_pr_675755585
branch
from
52d0410
to
5d7f197
Compare
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.
PR #16975: Add a few related optimization passes for fp8 gemm custom-calls.
Imported from GitHub PR openxla/xla#16975
This caused convergence issue for fp8 training, tested on GPT3 models:
Before:
After:
Copybara import of the project:
--
8bf6d19e5c10024d34a59b889893d203eee6691a by Elfie Guo [email protected]:
Add a few related optimization pass for fp8 gemm rerwriter.
Merging this change closes #16975
FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#16975 from elfiegg:pass 8bf6d19e5c10024d34a59b889893d203eee6691a