[NVIDIA] Extend the custom fp8 accumulate dtype in non-jit scenarios #3827
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
kaixih
force-pushed
the
fp8_meta_custom_dtype_non_jit
branch
2 times, most recently
from
4b8accf
to
f2070a1
Compare
kaixih
force-pushed
the
fp8_meta_custom_dtype_non_jit
branch
from
f2070a1
to
0b59501
Compare
kaixih
force-pushed
the
fp8_meta_custom_dtype_non_jit
branch
from
0b59501
to
fd9c467
Compare
|
Hi thanks for adding this change. |
|
@IvyZX Thanks for the review. I made one more change to make sure it compatible with the current JAX version that has no earray. PTAL. |
IvyZX
approved these changes
Apr 23, 2024
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Outdated suggestions cannot be applied.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.
This PR extends the custom fp8 dtype in the non-jit scenarios. The custom fp8 dtype is for the grad accumulation where we can do a max op rather than the default add op to accumulate the grads. This is designed for the fp8 parameters whose grads are their new values and we only want to pick the max among the grads of the sub-tensors.
Previously, this custom dtype has to be in the jit scope and this PR relaxes that restriction.
Note, this PR depends on jax-ml/jax#20266
cc. @nluehr @mingxu1067