Fix sample-cache corruption under accelerate data-parallel

[shipbehaves]

What

Fixes #1102.

Under an accelerate data-parallel launch, SampleCache.get_cache_path is not rank-aware, so every rank's @cached wrapper calls cache_samples and writes the same parquet file concurrently. The interleaved writes corrupt the file, and the subsequent get_samples_from_cache / load_dataset call then fails to load it.

Fix

In the cached wrapper (src/lighteval/utils/cache_management.py), only the main process writes the cache, with a wait_for_everyone() barrier so the other ranks wait for that write before they read:

accelerator = getattr(self, "accelerator", None)
if accelerator is None or accelerator.is_main_process:
    cache.cache_samples(...)
if accelerator is not None:
    accelerator.wait_for_everyone()

This is safe with no data loss: by the time results reach the wrapper they are already gathered across ranks (e.g. pad_and_gather in the transformers backend), so the main process holds the full set. This matches @NathanHB's suggestion on the issue. Backends without an accelerator (getattr returns None) are unaffected, and on a single process the guard is a no-op.

Test

Adds test_cache_only_main_process_writes in tests/unit/utils/test_caching.py: with a mocked non-main process the cache file is not written and the barrier is hit; with the main process it is written. It fails on main and passes with this change.

---

AI-assisted: drafted and tested with AI help, written and reviewed by a human who understands and stands behind the change.

[bot-ci-comment]

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.

[NathanHB]

Hey! There seems to be an issue with the tests, will merge when fixed :)