Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

Reset Inductor Pool After SGLang Shutdown (#2536)

Sun, Xuehaoafac833efd3d
diff --git a/test/e2e/test_cuda/test_sglang_throughput.py b/test/e2e/test_cuda/test_sglang_throughput.py
index 0d031b934..249f735ee 100644
--- a/test/e2e/test_cuda/test_sglang_throughput.py
+++ b/test/e2e/test_cuda/test_sglang_throughput.py
@@ -96,6 +96,31 @@ def _patched_stop(self, *args, _orig=_orig_stop, **kwargs):
multiprocessing.resource_tracker.ResourceTracker._stop = _patched_stop
+
+def _shutdown_sglang_engine(llm) -> None:
+ """Shut down an in-process sglang engine and reset torch inductor's compile pool.
+
+ ``sgl.Engine.shutdown()`` calls ``kill_process_tree(os.getpid(), include_parent=False)``,
+ which also kills inductor's compile-worker sidecar spawned by the previous
+ ``torch.compile``-enabled quantization. The dead pool stays cached in
+ ``AsyncCompile.process_pool``, so the next quantization that needs a fresh
+ compile fails with ``RuntimeError: Attempting to use a closed pool``.
+ Dropping the cached pool lets inductor spawn a new one on demand.
+ """
+ shutdown = getattr(llm, "shutdown", None)
+ if callable(shutdown):
+ try:
+ shutdown()
+ except Exception:
+ pass
+ try:
+ from torch._inductor.async_compile import shutdown_compile_workers
+
+ shutdown_compile_workers()
+ except Exception:
+ pass
+
+
pytestmark = [
pytest.mark.e2e,
pytest.mark.skipif(
@@ -226,12 +251,7 @@ def _run_sglang_benchmark(
sample_output=outputs[0]["text"],
)
finally:
- shutdown = getattr(llm, "shutdown", None)
- if callable(shutdown):
- try:
- shutdown()
- except Exception:
- pass
+ _shutdown_sglang_engine(llm)
del llm
gc.collect()
if torch.cuda.is_available():
@@ -403,10 +423,7 @@ def test_sglang_awq_format_via_cli(require_cuda):
text = outputs[0]["text"]
assert text.strip() and "!!!" not in text
finally:
- try:
- llm.shutdown()
- except Exception:
- pass
+ _shutdown_sglang_engine(llm)
del llm
gc.collect()
if torch.cuda.is_available():