[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:
aoshen02
2026-10-11 21:15:14 +08:00
committed by GitHub
parent c85616efc6
commit bd5bbce6e4
22 changed files with 181 additions and 14 deletions
@@ -24,6 +24,7 @@ class _RecordingEngine:
self.raise_on_update = raise_on_update
self.started = False
self.finished = False
self.closed = False
self.reset_count = 0
self.supports_draft_weight_update = False
self.update_calls: list[dict] = []
@@ -49,6 +50,9 @@ class _RecordingEngine:
def reset_weight_update_target(self) -> None:
self.reset_count += 1
def shutdown(self) -> None:
self.closed = True
class _RecordingModelRunner:
def __init__(self) -> None:
@@ -82,6 +86,25 @@ def test_reload_weights_sets_current_config():
assert model_runner.seen_config is worker.vllm_config
@pytest.mark.parametrize(
"enabled,active", [(True, False), (True, True), (False, False)]
)
def test_shutdown_weight_transfer_engine_preserves_inference(enabled, active):
engine = _RecordingEngine() if enabled else None
worker = _make_worker(engine)
worker._weight_update_active = active
model_runner = worker.model_runner
if active:
with pytest.raises(RuntimeError, match="during a weight update"):
worker.shutdown_weight_transfer_engine()
else:
worker.shutdown_weight_transfer_engine()
if engine is not None:
assert engine.closed is not active
assert worker.weight_transfer_engine is engine
assert worker.model_runner is model_runner
def test_reload_parameter_lookup_preserves_lora_module_names():
base_layer = nn.Module()
qweight = nn.Parameter(torch.ones(1))