Hotfix/fused ce triton #409
Open
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
sarthak-amd
requested review from
ipanfilo,
wangye805 and
wenchenvincent
as code owners
ipanfilo
requested changes
Jan 12, 2026
Collaborator
There was a problem hiding this comment.
What is a purpose of this file change?
Collaborator
There was a problem hiding this comment.
Modify copyright date. The file is also maintained in upstream so avoid unnecessary reformattings
| rank (int): The rank of this device in the TP group. | ||
| world_size (int): The size of world involved in this distributed loss calculation. | ||
| ignore_idx (int): Tokens to be ignored for loss and gradient calculation. | ||
| ignore_idx (int): Tokens to be ignored for loss and gradient calculation. (default -100) |
Collaborator
There was a problem hiding this comment.
There is no default here.
|
|
||
| def cross_entropy_backward(_input: torch.Tensor, grad_output: torch.Tensor): | ||
| def cross_entropy_backward( | ||
| _input: torch.Tensor, grad_output: torch.Tensor, is_cg_capturable: bool = False |
Collaborator
There was a problem hiding this comment.
This code interferes and conflicts with upcoming IFU 2.8
wenchenvincent
requested changes
Jan 14, 2026
Collaborator
wenchenvincent
left a comment
wenchenvincent
left a comment
There was a problem hiding this comment.
@sarthak-amd As mentioned in the previous PR, could you refactor the PR as 3 commits:
- 2 commits would be cherrypicking from the upstream PRs. (NVIDIA/TransformerEngine#1879, NVIDIA/TransformerEngine#2139)
- 1 commit for the ignore_idx with a test to cover it.
This way the PR would be very clear and easy to understand.
Collaborator
The one of aforementioned PRs is part of IFU 2.6, i.e. part of ROCm TE already, the other is part of IFU 2.8 |
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.
Kernel fixes:
denom = max(1, (target != ignore_idx).sum())denominstead ofn_rowsgrad_output_strideparameter, compute dynamically:1 if grad_output.numel() > 1 else 0is_cg_capturableflag for CUDA graph compatibilityTest improvements:
torch.nn.CrossEntropyLosstorch.square(loss)for non-trivial backward