Remove CUDA synchronization in mean_token_accuracy

[cyyever]

What does this PR do?

mean_token_accuracy computation invokes item() on token counts, which triggers CUDA to CPU synchronization. That synchronization becomes a minor performance bottleneck in LLM fine-tuning, as indicated by the following profiling snapshot from v0.15.1:

That bottleneck has been fixed in this PR by accumulating the correct and total token counts in tensors. item() calls are delayed until trainer.log().
The effects of the change are indicated by another profiling that the bottleneck disappears:

Because the metrics are cleared immediately after logging, this change should be safe and backwards-compatible.

Before submitting

• Did you read the contributor guideline,
  Pull Request section?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[qgallouedec]

I don't understand the profiling actually. Where do you get that this line is the bottleneck?
Thank for contributing!

[cyyever]

@qgallouedec It is the sixth line in the first picture. It's not a main bottleneck, however, the GPU utility rose a bit after fixing it.

[qgallouedec]

This?

dequantize_4bit (bitsandbytes/functional. py:1380)

[qgallouedec]

The comparison is not very clear to me tbh, do you have clearer results, like two trainings (one with main, one with your branch) where we can see the speedup in term of steps/sec?

[cyyever]

@qgallouedec Of course, I will provide a comparison ASAP.