[LiRunGuo]

Summary

Fixes #1594.

liger_cross_entropy_kernel reuses the input tensor as gradient storage. For rows whose target is ignore_index, it was zeroing the input unconditionally, including when the input did not require grad. This mutated caller-owned logits during evaluation or metrics collection even though no backward pass would consume the result.

Gate that store on HAS_GRADIENTS, matching the existing gradient path below it. The actual loss and accuracy outputs are unchanged.

Testing Done

• Hardware Type: NVIDIA GeForce RTX 3090
• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

GPU validation on Yokota:

$ python -m pytest -c pyproject.toml test/transformers/test_cross_entropy.py -q
187 passed, 1 warning in 60.68s

The new regression covers both float32 and bfloat16, verifies that ignored rows remain byte-identical under torch.no_grad(), and compares the non-ignored losses with PyTorch.

$ ruff check --output-format=concise .
All checks passed!

$ ruff format --check --diff .
500 files already formatted

AI assistance was used to investigate, implement, and validate this focused change.

[LiRunGuo]

Full CrossEntropy test file completed on an NVIDIA GeForce RTX 3090:

$ PYTHONPATH=src:. python -m pytest -c pyproject.toml test/transformers/test_cross_entropy.py -q
187 passed, 1 warning in 60.68s (0:01:00)

The two new parameterizations both pass and verify that ignored rows remain byte-identical under torch.no_grad(). The warning is the environment missing pytest-asyncio and is unrelated to this change.

[Tcc0403]

Great catch! Thanks for the contribution!