Fused Cross Entropy Triton - Loss Scaling and Vanishing Grads Bugfix

[sarthak-amd]

Description

The Fused Cross Entropy Triton Kernel currently has 2 bugs

1. if ignore_idx is not None`, the loss should be computed only over valid tokens and not all tokens (new fix)
2. gradient scaling when reduce_loss=False. (This is already fixed in upstream)
   • If reduced loss=False, we should compute per token loss and not reduce it else it would shrink the gradients by 1/N giving wrong (higher) loss.
   • if reduce_loss=False, grad_output is a tensor, not a scalar. We need to load 1 value per row instead of just a scalar.

This fix is validated on Llama3.1 8B model for Pre-training.

Type of change

• Bug fix (non-breaking change which fixes an issue)

[wenchenvincent]

@sarthak-amd Could you post the PR for the upstream fix?

[wenchenvincent]

There is no real change in this file. Let's keep this file intact and then we don't need to add the AMD copyright statement.

[sarthak-amd]

@sarthak-amd Could you post the PR for the upstream fix?

NVIDIA/TransformerEngine@e9a5fa4 @wenchenvincent

[wenchenvincent]

@sarthak-amd Could you post the PR for the upstream fix?

NVIDIA/TransformerEngine@e9a5fa4 @wenchenvincent

Another fix came from the upstream PR NVIDIA/TransformerEngine#1879. Is the change of test in that PR also reflected?

[wenchenvincent]

For the fix for ignore_idx, is there a test for it (without the fix, the test would fail)?

[wenchenvincent]

@sarthak-amd Could you refactor the PR as 3 commits:

• 2 commits would be cherrypicking from the upstream PRs.
• 1 commit for the ignore_idx with a test to cover it.

This way the PR would be very clear and easy to understand.

[wenchenvincent]

@sarthak-amd Could you address the comments? Also, please rebase upon latest dev so that hot fixes for sgpu tests could pass.