[UT] Destroy process group after broadcast to resolve port occupation issues in multi-server tests (#12379)

This commit is contained in:
Zeyu Li
2025-11-20 12:51:58 +08:00
committed by GitHub
parent c3c4da71fb
commit 21370ef7b5

View File

@@ -187,6 +187,9 @@ def init_process_hf(
print(f"[hf] {rank=} {broadcast_time=:.3f}s")
param_queue.put(("broadcast_time", broadcast_time))
# Destroy process group and release related resource
torch.distributed.destroy_process_group(group)
# Delete the huggingface models to free up memory.
del hf_instruct_model
del hf_base_model