[RL] Release weight-transfer communicators without stopping inference (#61060)
Signed-off-by: aoshen02 <aoshen@inferact.ai> Signed-off-by: fh3 <a@b>
This commit is contained in:
@@ -169,6 +169,13 @@ def test_init_weight_transfer_engine_calls_engine(vllm_runner):
|
||||
assert called, "init_transfer_engine should have been called"
|
||||
assert param == "hello", f"Expected 'hello', got {param}"
|
||||
|
||||
llm.shutdown_weight_transfer_engine()
|
||||
assert all(
|
||||
llm.collective_rpc(
|
||||
lambda worker: worker.weight_transfer_engine.shutdown_called
|
||||
)
|
||||
)
|
||||
|
||||
|
||||
@create_new_process_for_each_test()
|
||||
def test_update_weights_calls_engine(vllm_runner):
|
||||
|
||||
Reference in New Issue
Block a user