Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

FSDP2 sharded loading gloo test: barrier before a rank tears down its group (#15499)

Daniel Hanebf98cb44f5d
diff --git a/tests/test_fsdp2_native_sharded_load_gloo.py b/tests/test_fsdp2_native_sharded_load_gloo.py
index 58fa11d87f9..a2d66441e7f 100644
--- a/tests/test_fsdp2_native_sharded_load_gloo.py
+++ b/tests/test_fsdp2_native_sharded_load_gloo.py
@@ -328,6 +328,9 @@ def _worker(rank, world, port, case, tmp, kwargs):
try:
res = CASES[case](rank, world, tmp, **kwargs)
torch.save(res, os.path.join(tmp, f"result_{rank}.pt"))
+ # The refusal cases end without a collective. Rank 0 hosts the TCPStore: let it exit while the other
+ # rank is still setting up a mesh subgroup through that store and the other rank dies on a broken pipe.
+ dist.barrier()
finally:
dist.destroy_process_group()