Inductor's on-disk caches default to one directory per user, not per process, so the four workers under pytest -n 4 share /tmp/torchinductor_<user> and its fxgraph, aotautograd and Triton subtrees. The upstream recipe is explicit that a common TORCHINDUCTOR_CACHE_DIR is what makes processes share compiled artifacts, so a different value per worker is how they are kept apart.

Sharing is not obviously wrong — the entries are content-addressed. It is still a write-write interaction between processes that nothing here controls, and a cache is exactly the sort of thing that turns a deterministic suite into an intermittent one. This removes the interaction for the price of some recompilation.

What it buys, measured rather than assumed

A full unsloth_zoo run against an empty, dedicated cache directory:

COLD  578s   cache entries=0   size=0
WARM  503s   cache entries=0   size=0

That suite never populates the on-disk cache at all, so for the largest step in the Core legs this is neither a saving nor a cost — and the 75s cold/warm gap is page cache, not compilation, since there is nothing in the directory either time. Worth recording, because it also rules out persisting TORCHINDUCTOR_CACHE_DIR across CI runs for that suite: there is nothing to persist.

So this is insurance for the suites that do compile, applied wherever tests run in parallel rather than only where a problem has already been seen.

Details that matter

• It runs as a module-level import in the two conftests, not a fixture, because it must take effect before torch is imported.
• TRITON_CACHE_DIR is set explicitly rather than left to follow: it only derives from the inductor directory when unset, and an environment that exports it would keep all four workers on one Triton cache.
• An explicit TORCHINDUCTOR_CACHE_DIR is split underneath rather than replaced, so a CI path chosen on purpose is respected.
• A single-process run is left alone entirely — it already has the default to itself.

Verification

Under real xdist, not only unit tests: four workers report four distinct directories, and the parallel suites still pass. The guard also asserts that both conftests still reach the helper, because a helper nobody imports is the silent failure here.