From 049f0268969e561a0165eaed459d16a47c46194f Mon Sep 17 00:00:00 2001
From: Yue Sun <yuesu@microsoft.com>
Date: Fri, 31 Jul 2026 15:11:25 +0800
Subject: [PATCH] fix(perf): avoid native stderr redirect for VitisAI

---
 src/winml/modelkit/commands/perf.py  | 12 +++++++--
 tests/unit/commands/test_perf_cli.py | 38 ++++++++++++++++++++++++++++
 2 files changed, 48 insertions(+), 2 deletions(-)

diff --git a/src/winml/modelkit/commands/perf.py b/src/winml/modelkit/commands/perf.py
index b23a13842..917b9273b 100644
--- a/src/winml/modelkit/commands/perf.py
+++ b/src/winml/modelkit/commands/perf.py
@@ -1051,8 +1051,16 @@ def _load_model(self) -> None:
             ),
         }
 
+        # VitisAI's compiler can hang during InferenceSession creation when
+        # native stderr points to a pipe, even when that pipe is drained
+        # concurrently. Keep its native handle untouched; other EPs retain the
+        # default warning filtering.
+        filter_native_warnings = (
+            self._ep_device.device.ep_name != "VitisAIExecutionProvider"
+        )
+
         if is_onnx:
-            with suppress_native_warnings(enabled=True):
+            with suppress_native_warnings(enabled=filter_native_warnings):
                 self._model = WinMLAutoModel.from_onnx(
                     onnx_path=model_path,
                     skip_build=self.config.skip_build,
@@ -1060,7 +1068,7 @@ def _load_model(self) -> None:
                     **common_kwargs,
                 )
         else:
-            with suppress_native_warnings(enabled=True):
+            with suppress_native_warnings(enabled=filter_native_warnings):
                 self._model = WinMLAutoModel.from_pretrained(
                     model_id,
                     **common_kwargs,
diff --git a/tests/unit/commands/test_perf_cli.py b/tests/unit/commands/test_perf_cli.py
index ace26947d..62eed01ef 100644
--- a/tests/unit/commands/test_perf_cli.py
+++ b/tests/unit/commands/test_perf_cli.py
@@ -512,6 +512,44 @@ def resolve_ep_device(self: PerfBenchmark) -> None:
         assert all(import_observed)
         assert load_observed == [True]
 
+    def test_load_model_does_not_redirect_native_stderr_for_vitisai(
+        self, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        """VitisAI session creation keeps the process native stderr handle."""
+        suppression_enabled: list[bool] = []
+
+        @contextmanager
+        def record_native_suppression(*_args: object, **kwargs: object):
+            suppression_enabled.append(bool(kwargs.get("enabled")))
+            yield
+
+        class FakeWinMLAutoModel:
+            @staticmethod
+            def from_pretrained(*args: object, **kwargs: object) -> MagicMock:
+                return MagicMock()
+
+        def resolve_ep_device(self: PerfBenchmark) -> None:
+            self._ep_device = MagicMock()
+            self._ep_device.device.ep_name = "VitisAIExecutionProvider"
+            self._resolved_device = "npu"
+            self._resolved_ep = "VitisAIExecutionProvider"
+
+        fake_models_pkg = ModuleType("winml.modelkit.models")
+        fake_models_pkg.WinMLAutoModel = FakeWinMLAutoModel
+        monkeypatch.setitem(sys.modules, "winml.modelkit.models", fake_models_pkg)
+        monkeypatch.setattr(
+            "winml.modelkit.commands.perf.suppress_native_warnings",
+            record_native_suppression,
+        )
+        monkeypatch.setattr(PerfBenchmark, "_resolve_device_ep", resolve_ep_device)
+
+        benchmark = PerfBenchmark(
+            BenchmarkConfig(model_id="microsoft/resnet-50", task="image-classification")
+        )
+        benchmark._load_model()
+
+        assert suppression_enabled == [True, False]
+
     def test_resolve_device_ep_filters_native_warnings_and_preserves_errors(
         self,
         monkeypatch: pytest.MonkeyPatch,