Reduce gc.collect() frequency in test teardown from every test to every 50

gc.collect() costs ~130-380ms per call scanning all tracked Python objects,
even when it finds zero cyclic garbage. PyTorch tensors are freed by
reference counting, not GC, so per-test collection is unnecessary overhead.
Running it every 50 tests (and at session end) is sufficient to handle any
hypothetical cyclic garbage accumulation while saving ~24% on test runtime.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: TimDettmers

tests/conftest.py

@@ -29,9 +29,15 @@ def pytest_runtest_call(item):
         raise
 
 
+_teardown_counter = 0
+
+
 @pytest.hookimpl(trylast=True)
 def pytest_runtest_teardown(item, nextitem):
-    gc.collect()
+    global _teardown_counter
+    _teardown_counter += 1
+    if _teardown_counter % 50 == 0 or nextitem is None:
+        gc.collect()
     if torch.cuda.is_available():
         torch.cuda.empty_cache()
     elif torch.backends.mps.is_available() and torch.backends.mps.is_built():