Bump liger-kernel to 0.4.0

[ByronHsu]

This fixes HF grad acc and other fixes/features/optimizations

[qgallouedec]

As far as I understand, the grad accum thing is only an issue with SFT right?

[kashif]

right i think its more about the updated kernels

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ByronHsu]

Yes grad accum is only used for sft. Beside grad accum, we also have other improvement

[qgallouedec]

I approve, as this is an important issue affecting the most widely used trainer. (Thanks for solving it!)

For the record, generally speaking, I won’t raise the minimum version requirement unless a new feature from the dependency is needed in our codebase.