A bug for `reduction='none'` in `LigerFusedLinearCrossEntropyLoss`

[yshenaw]

🐛 Describe the bug

I noticed that the fused_linear_cross_entropy function has support for reduction='none', but in the LigerFusedLinearCrossEntropyLoss, there is currently an assertion that restricts reduction to only 'mean' or 'sum'. The current code looks like this:

assert reduction in {
    "mean",
    "sum",
}, f"reduction must be 'mean' or 'sum'. Got: {reduction}"

Here is the proposed change:

assert reduction in {
    "mean",
    "sum",
    "none",
}, f"reduction must be 'mean', 'sum', or 'none'. Got: {reduction}"

I hope this bug can be fixed soon.

Reproduce

No response

Versions

Environment Report:

Operating System: Linux-6.8.0-1031-azure-x86_64-with-glibc2.35
Python version: 3.10.18
Liger Kernel version: 0.6.2
PyTorch version: 2.8.0+cu128
CUDA version: 12.8
HIP(ROCm) version: Not available
Triton version: 3.4.0
Transformers version: 4.56.0
XPU version: XPU Not Available

[shimizust]

Thanks @yshenaw for reporting--that looks like a bug. Are you able to submit that fix?

[yshenaw]

Thanks @yshenaw for reporting--that looks like a bug. Are you able to submit that fix?

I have submit the fix in PR #879. Could you have a review?