Fix ONNX Runtime CPU memory arena configuration (#25003)

* Fix CPU memory arena option on ONNX Runtime sessions

* Limit CPU arena correction to CPU enrichment sessions

* Retain default arena in GPU runner integration regression

* Preserve native detector session-options contract
This commit is contained in:
Rmdev.pro
2026-11-18 15:36:11 -07:00
committed by GitHub
parent 6c09629783
commit 3c122150ae
4 changed files with 18 additions and 16 deletions
+2
View File
@@ -153,6 +153,8 @@ def get_ort_session_options(
sess_options = ort.SessionOptions() sess_options = ort.SessionOptions()
if cpu: if cpu:
# ORT reads the CPU arena flag from session options, not provider options.
sess_options.enable_cpu_mem_arena = False
# idle intra-op threads otherwise spin for ~1s after every run, keeping # idle intra-op threads otherwise spin for ~1s after every run, keeping
# most cores busy while an enrichment runs even once a second # most cores busy while an enrichment runs even once a second
sess_options.add_session_config_entry("session.intra_op.allow_spinning", "0") sess_options.add_session_config_entry("session.intra_op.allow_spinning", "0")
+10 -2
View File
@@ -59,10 +59,18 @@ class TestGetOrtSessionOptions(unittest.TestCase):
ort.GraphOptimizationLevel.ORT_ENABLE_ALL, ort.GraphOptimizationLevel.ORT_ENABLE_ALL,
) )
def test_only_cpu_enrichment_models_disable_the_cpu_arena(self):
for model_type in EnrichmentModelTypeEnum:
for cpu in [False, True]:
with self.subTest(model_type=model_type, cpu=cpu):
options = get_ort_session_options(model_type.value, cpu=cpu)
self.assertEqual(options.enable_cpu_mem_arena, not cpu)
def test_detector_models_use_defaults(self): def test_detector_models_use_defaults(self):
for model_type in [None, ModelTypeEnum.rfdetr.value]: for model_type in [None, ModelTypeEnum.rfdetr.value]:
with self.subTest(model_type=model_type): for cpu in [False, True]:
self.assertIsNone(get_ort_session_options(model_type, cpu=True)) with self.subTest(model_type=model_type, cpu=cpu):
self.assertIsNone(get_ort_session_options(model_type, cpu=cpu))
class TestCudaGraphRunner(unittest.TestCase): class TestCudaGraphRunner(unittest.TestCase):
+4 -1
View File
@@ -140,6 +140,9 @@ class TestLoadRegistry(unittest.TestCase):
spinning = None spinning = None
self.assertEqual(spinning, expected) self.assertEqual(spinning, expected)
self.assertEqual(
options.enable_cpu_mem_arena, provider != "CPUExecutionProvider"
)
class TestLoadedDeviceSnapshot(unittest.TestCase): class TestLoadedDeviceSnapshot(unittest.TestCase):
@@ -560,4 +563,4 @@ class TestMIGraphX(unittest.TestCase):
providers, options = model_util.get_ort_providers() providers, options = model_util.get_ort_providers()
self.assertEqual(providers, ["CPUExecutionProvider"]) self.assertEqual(providers, ["CPUExecutionProvider"])
self.assertEqual(options, [{"enable_cpu_mem_arena": False}]) self.assertEqual(options, [{}])
+2 -13
View File
@@ -320,14 +320,7 @@ def get_ort_providers(
device: str | None = "AUTO", device: str | None = "AUTO",
) -> tuple[list[str], list[dict[str, Any]]]: ) -> tuple[list[str], list[dict[str, Any]]]:
if force_cpu: if force_cpu:
return ( return ["CPUExecutionProvider"], [{}]
["CPUExecutionProvider"],
[
{
"enable_cpu_mem_arena": False,
}
],
)
providers = [] providers = []
options = [] options = []
@@ -389,11 +382,7 @@ def get_ort_providers(
continue continue
elif provider == "CPUExecutionProvider": elif provider == "CPUExecutionProvider":
providers.append(provider) providers.append(provider)
options.append( options.append({})
{
"enable_cpu_mem_arena": False,
}
)
elif provider == "AzureExecutionProvider": elif provider == "AzureExecutionProvider":
# Skip Azure provider - not typically available on local hardware # Skip Azure provider - not typically available on local hardware
# and prevents fallback to OpenVINO when it's the first provider # and prevents fallback to OpenVINO when it's the first provider