test: Add pre-download model test that run at beginning avoiding individual test timeout (#4989)

Signed-off-by: Jacky <18255193+kthui@users.noreply.github.com>

container/deps/requirements.test.txt

@@ -23,6 +23,7 @@ pytest-forked==1.6.0
 pytest-httpserver==1.1.3
 pytest-md-report==0.7.0
 pytest-mypy==1.0.1
+pytest-order==1.3.0
 pytest-timeout==2.4.0
 pytest-xdist==3.8.0
 # Triton client to Dynamo gRPC server

pyproject.toml

@@ -140,7 +140,6 @@ known_first_party = ["dynamo"]
 [tool.pytest.ini_options]
 minversion = "8.0"
 tmp_path_retention_policy = "failed"
-timeout_func_only = true
 
 # NOTE
 # We ignore model.py explicitly here to avoid mypy errors with duplicate modules

tests/test_predownload_models.py

+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""
+Test module to trigger model predownloading before any other tests run.
+
+This module uses pytest-order to ensure the predownload_models fixture
+is executed at the very beginning of the test session, downloading all
+required models before other tests attempt to use them.
+
+The tests use parametrization with pytest.param marks to match different
+CI marker expressions (e.g., "vllm and e2e and gpu_1").
+"""
+
+import pytest
+
+
+@pytest.mark.order("first")
+@pytest.mark.pre_merge
+@pytest.mark.parametrize(
+    "predownload_models_variant",
+    [
+        pytest.param(
+            "predownload_models_vllm_gpu1",
+            marks=[pytest.mark.vllm, pytest.mark.e2e, pytest.mark.gpu_1],
+        ),
+        pytest.param(
+            "predownload_models_sglang_gpu1",
+            marks=[pytest.mark.sglang, pytest.mark.e2e, pytest.mark.gpu_1],
+        ),
+        pytest.param(
+            "predownload_models_trtllm_gpu1",
+            marks=[pytest.mark.trtllm, pytest.mark.e2e, pytest.mark.gpu_1],
+        ),
+        pytest.param(
+            "predownload_models_vllm_gpu2",
+            marks=[pytest.mark.vllm, pytest.mark.e2e, pytest.mark.gpu_2],
+        ),
+        pytest.param(
+            "predownload_models_sglang_gpu2",
+            marks=[pytest.mark.sglang, pytest.mark.e2e, pytest.mark.gpu_2],
+        ),
+        pytest.param(
+            "predownload_models_trtllm_gpu2",
+            marks=[pytest.mark.trtllm, pytest.mark.e2e, pytest.mark.gpu_2],
+        ),
+    ],
+)
+def test_predownload_models(predownload_models_variant, predownload_models):
+    """Trigger the predownload_models fixture to download all required models.
+
+    This test runs before any other tests in the session due to pytest.mark.order("first").
+    By downloading models upfront, we ensure that model download time is not counted
+    against individual test timeouts. This prevents tests from failing due to slow
+    network downloads rather than actual test failures.
+
+    The actual download logic is handled by the predownload_models session-scoped fixture
+    defined in conftest.py.
+
+    The parametrization ensures this test is selected by any CI marker expression
+    that matches framework-specific E2E tests (e.g., "vllm and e2e and gpu_1").
+    """
+    # The fixture handles the download - this test just ensures it runs first
+    pass