[Common][PyTorch] Add z_loss_weight and log_sum_exp output to parallel_cross_entropy#2707
Draft
bassoy wants to merge 2 commits intoNVIDIA:mainfrom
Draft
[Common][PyTorch] Add z_loss_weight and log_sum_exp output to parallel_cross_entropy#2707bassoy wants to merge 2 commits intoNVIDIA:mainfrom
bassoy wants to merge 2 commits intoNVIDIA:mainfrom
Conversation
…l_cross_entropy - Triton kernel: adds z_loss_weight (tl.constexpr) and a log_sum_exp output buffer; lse = m + log(d) reuses values already in registers — zero extra compute; z_loss blocks are dead-code-eliminated by Triton when z_loss_weight=0.0 - API: new z_loss_weight=0.0 and return_log_sum_exp=False parameters; default return is a single Tensor (backward compatible); return_log_sum_exp=True returns (loss, log_sum_exp) tuple; return type: Union[Tensor, Tuple[Tensor, Tensor]] matching TE convention - Tests: 15 tests covering z_loss correctness, BF16, non-uniform backward gradients (loss masking), log_sum_exp semantics, and backward compatibility Signed-off-by: Cem Bassoy <cem.bassoy@deepl.com>
for more information, see https://pre-commit.ci
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Description
Adds z-loss regularization (see https://arxiv.org/abs/2202.08906) and
log_sum_expoutput toparallel_cross_entropy. The Triton kernel already computeslse = m + log(d)as part of the online softmax. We store it to a new output buffer and optionally addz_loss_weight * lse^2to the loss. Atz_loss_weight=0.0all z-loss logic is dead-code-eliminated by Triton (tl.constexpr), so there is no overhead.Type of change
Changes
z_loss_weight(tl.constexpr) andlog_sum_expoutput buffer to Triton CE kernelparallel_cross_entropy (z_loss_weight=0.0, return_log_sum_exp=False)-- backward compatible; default return unchanged (singleTensor)Checklist: