transformers

ec79b167 - fix SIGABRT in TP/EP tests by forcing gc.collect() after destroy_process_group

fix SIGABRT in TP/EP tests by forcing gc.collect() after destroy_process_group The SIGABRT was a timing race during interpreter shutdown: gloo's pt_gloo_runloop threads were still alive when Python's finalizer called the C++ ProcessGroup destructor (because the Python ProcessGroup object was being GC'd at interpreter exit rather than immediately). Root cause identified via the diagnostic branch (0f562ff2): adding extensive file I/O after destroy_process_group() caused 6 consecutive CI runs to pass, confirming the bug is purely a latency race between gloo thread teardown and interpreter shutdown. The minimal fix is gc.collect() immediately after destroy_process_group() so the C++ destructor fires while the process is still in a clean state, not during interpreter finalization. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>