From 15422ed3f748b2c26a9446db388979027a254199 Mon Sep 17 00:00:00 2001
From: Ryan Rock <ryan.rock@amd.com>
Date: Wed, 14 Jan 2026 22:01:42 -0600
Subject: [PATCH] [CI/Build][Hardware][AMD] Fix v1/shutdown (#31997)

Signed-off-by: Ryan Rock <ryan.rock@amd.com>
---
 tests/v1/shutdown/conftest.py           | 26 +++++++++++++++++++++++++
 tests/v1/shutdown/test_forward_error.py | 20 +++++++++++++++++--
 tests/v1/shutdown/test_startup_error.py | 21 +++++++++++++++++++-
 3 files changed, 64 insertions(+), 3 deletions(-)
 create mode 100644 tests/v1/shutdown/conftest.py

diff --git a/tests/v1/shutdown/conftest.py b/tests/v1/shutdown/conftest.py
new file mode 100644
index 000000000..b276d0fab
--- /dev/null
+++ b/tests/v1/shutdown/conftest.py
@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import os
+from collections.abc import Iterable
+from pathlib import Path
+
+import pytest
+
+from vllm.platforms import current_platform
+
+
+@pytest.fixture
+def rocm_sitecustomize_factory(monkeypatch, tmp_path: Path):
+    """Return a function that installs a given sitecustomize payload."""
+    if not current_platform.is_rocm():
+        return lambda _: None
+
+    def install(lines: Iterable[str]) -> None:
+        sc = tmp_path / "sitecustomize.py"
+        sc.write_text("\n".join(lines) + "\n")
+        monkeypatch.setenv(
+            "PYTHONPATH",
+            os.pathsep.join(filter(None, [str(tmp_path), os.getenv("PYTHONPATH")])),
+        )
+
+    return install
diff --git a/tests/v1/shutdown/test_forward_error.py b/tests/v1/shutdown/test_forward_error.py
index a751b2d91..4625bc174 100644
--- a/tests/v1/shutdown/test_forward_error.py
+++ b/tests/v1/shutdown/test_forward_error.py
@@ -3,6 +3,7 @@
 """Test that we handle an Error in model forward and shutdown."""
 
 import asyncio
+import inspect
 
 import pytest
 
@@ -38,11 +39,22 @@ def evil_forward(self, *args, **kwargs):
     return self.model(*args, **kwargs)
 
 
+@pytest.fixture
+def rocm_evil_forward(rocm_sitecustomize_factory):
+    lines = [
+        "from vllm.distributed import get_tensor_model_parallel_rank",
+        "from vllm.model_executor.models.llama import LlamaForCausalLM",
+        inspect.getsource(evil_forward),
+        f"LlamaForCausalLM.forward = {evil_forward.__name__}",
+    ]
+    rocm_sitecustomize_factory(lines)
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("tensor_parallel_size", [2, 1])
 @pytest.mark.parametrize("model", MODELS)
 async def test_async_llm_model_error(
-    monkeypatch, tensor_parallel_size: int, model: str
+    monkeypatch, rocm_evil_forward, tensor_parallel_size: int, model: str
 ) -> None:
     """Test that AsyncLLM propagates a forward pass error and frees memory.
 
@@ -104,7 +116,11 @@ async def test_async_llm_model_error(
 @pytest.mark.parametrize("tensor_parallel_size", [2, 1])
 @pytest.mark.parametrize("model", MODELS)
 def test_llm_model_error(
-    monkeypatch, tensor_parallel_size: int, enable_multiprocessing: bool, model: str
+    monkeypatch,
+    rocm_evil_forward,
+    tensor_parallel_size: int,
+    enable_multiprocessing: bool,
+    model: str,
 ) -> None:
     """Test that LLM propagates a forward pass error and frees memory.
     TODO(andy) - LLM without multiprocessing; LLM with multiprocessing
diff --git a/tests/v1/shutdown/test_startup_error.py b/tests/v1/shutdown/test_startup_error.py
index c1594cc2e..7925dc14b 100644
--- a/tests/v1/shutdown/test_startup_error.py
+++ b/tests/v1/shutdown/test_startup_error.py
@@ -2,6 +2,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Test that we handle a startup Error and shutdown."""
 
+import inspect
+
 import pytest
 
 from tests.utils import wait_for_gpu_memory_to_clear
@@ -28,12 +30,28 @@ def evil_method(self, *args, **kwargs):
     return self.model(*args, **kwargs, intermediate_tensors=None)
 
 
+@pytest.fixture
+def rocm_evil_method(rocm_sitecustomize_factory, request):
+    failing_method = request.getfixturevalue("failing_method")
+    lines = [
+        "from vllm.distributed import get_tensor_model_parallel_rank",
+        "from vllm.model_executor.models.llama import LlamaForCausalLM",
+        inspect.getsource(evil_method),
+        f"LlamaForCausalLM.{failing_method} = {evil_method.__name__}",
+    ]
+    rocm_sitecustomize_factory(lines)
+
+
 @pytest.mark.timeout(SHUTDOWN_TEST_TIMEOUT_SEC)
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("tensor_parallel_size", [2, 1])
 @pytest.mark.parametrize("failing_method", ["forward", "load_weights"])
 def test_async_llm_startup_error(
-    monkeypatch, model: str, tensor_parallel_size: int, failing_method: str
+    monkeypatch,
+    rocm_evil_method,
+    model: str,
+    tensor_parallel_size: int,
+    failing_method: str,
 ) -> None:
     """Test that AsyncLLM propagates an __init__ error & frees memory.
     Test profiling (forward()) and load weights failures.
@@ -67,6 +85,7 @@ def test_async_llm_startup_error(
 @pytest.mark.parametrize("failing_method", ["forward", "load_weights"])
 def test_llm_startup_error(
     monkeypatch,
+    rocm_evil_method,
     model: str,
     tensor_parallel_size: int,
     enable_multiprocessing: bool,