test: remove pytest_runtestloop (#4886)

Signed-off-by: PeaBrane <yanrpei@gmail.com>

pyproject.toml

@@ -140,6 +140,7 @@ 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/conftest.py

@@ -5,7 +5,6 @@ import logging
 import os
 import shutil
 import tempfile
-import time
 from pathlib import Path
 from typing import Optional
 
@@ -223,26 +222,6 @@ def pytest_collection_modifyitems(config, items):
         config.models_to_download = models_to_download
 
 
-def pytest_runtestloop(session):
-    """Download models after collection but before any tests run.
-
-    This hook runs after pytest_collection_modifyitems (so models are collected)
-    but before any test execution, ensuring model downloads don't count against test timeouts.
-    """
-    models = getattr(session.config, "models_to_download", None)
-
-    if models:
-        logging.info(
-            f"Downloading {len(models)} models before test execution\nModels: {models}"
-        )
-        start_time = time.time()
-
-        download_models(model_list=list(models))
-
-        download_duration = time.time() - start_time
-        logging.info(f"Model download completed in {download_duration:.1f}s")
-
-
 class EtcdServer(ManagedProcess):
     def __init__(self, request, port=2379, timeout=300):
         # Allocate free ports if port is 0

tests/fault_tolerance/cancellation/test_sglang.py

@@ -187,7 +187,7 @@ class DynamoWorkerProcess(ManagedProcess):
 @pytest.mark.xfail(strict=False)
 @pytest.mark.parametrize("request_plane", ["nats", "tcp"], indirect=True)
 def test_request_cancellation_sglang_aggregated(
-    request, runtime_services_dynamic_ports
+    request, runtime_services_dynamic_ports, predownload_models
 ):
     """
     End-to-end test for request cancellation functionality in aggregated mode.
@@ -305,7 +305,7 @@ def test_request_cancellation_sglang_aggregated(
     indirect=True,
 )
 def test_request_cancellation_sglang_decode_cancel(
-    request, runtime_services_dynamic_ports
+    request, runtime_services_dynamic_ports, predownload_models
 ):
     """
     End-to-end test for request cancellation during decode phase.

tests/fault_tolerance/cancellation/test_trtllm.py

@@ -166,7 +166,7 @@ class DynamoWorkerProcess(ManagedProcess):
 @pytest.mark.timeout(140)  # 3x average
 @pytest.mark.parametrize("request_plane", ["nats", "tcp"], indirect=True)
 def test_request_cancellation_trtllm_aggregated(
-    request, runtime_services_dynamic_ports
+    request, runtime_services_dynamic_ports, predownload_models
 ):
     """
     End-to-end test for request cancellation functionality in aggregated mode.
@@ -264,7 +264,7 @@ def test_request_cancellation_trtllm_aggregated(
     indirect=True,
 )
 def test_request_cancellation_trtllm_decode_cancel(
-    request, runtime_services_dynamic_ports
+    request, runtime_services_dynamic_ports, predownload_models
 ):
     """
     End-to-end test for request cancellation during decode phase with unified frontend.
@@ -359,7 +359,7 @@ def test_request_cancellation_trtllm_decode_cancel(
     indirect=True,
 )
 def test_request_cancellation_trtllm_prefill_cancel(
-    request, runtime_services_dynamic_ports
+    request, runtime_services_dynamic_ports, predownload_models
 ):
     """
     End-to-end test for request cancellation during prefill phase with unified frontend.
@@ -456,7 +456,7 @@ def test_request_cancellation_trtllm_prefill_cancel(
     strict=False,
 )
 def test_request_cancellation_trtllm_kv_transfer_cancel(
-    request, runtime_services_dynamic_ports
+    request, runtime_services_dynamic_ports, predownload_models
 ):
     """
     End-to-end test for request cancellation during prefill to decode KV transfer phase.

tests/fault_tolerance/cancellation/test_vllm.py

@@ -167,7 +167,9 @@ class DynamoWorkerProcess(ManagedProcess):
 
 @pytest.mark.timeout(110)  # 3x average
 @pytest.mark.parametrize("request_plane", ["nats", "tcp"], indirect=True)
-def test_request_cancellation_vllm_aggregated(request, runtime_services_dynamic_ports):
+def test_request_cancellation_vllm_aggregated(
+    request, runtime_services_dynamic_ports, predownload_models
+):
     """
     End-to-end test for request cancellation functionality in aggregated mode.
 
@@ -258,7 +260,7 @@ def test_request_cancellation_vllm_aggregated(request, runtime_services_dynamic_
     indirect=True,
 )
 def test_request_cancellation_vllm_decode_cancel(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for request cancellation during decode phase.
@@ -350,7 +352,7 @@ def test_request_cancellation_vllm_decode_cancel(
     indirect=True,
 )
 def test_request_cancellation_vllm_prefill_cancel(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for request cancellation during prefill phase.

tests/fault_tolerance/migration/test_sglang.py

@@ -151,7 +151,7 @@ class DynamoWorkerProcess(ManagedProcess):
     indirect=True,
 )
 def test_request_migration_sglang_worker_failure(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with migration support using SGLang.
@@ -228,7 +228,7 @@ def test_request_migration_sglang_worker_failure(
     indirect=True,
 )
 def test_request_migration_sglang_graceful_shutdown(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with graceful shutdown and migration support using SGLang.
@@ -308,7 +308,7 @@ def test_request_migration_sglang_graceful_shutdown(
     indirect=True,
 )
 def test_no_request_migration_sglang_worker_failure(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with migration disabled using SGLang.
@@ -403,7 +403,7 @@ def test_no_request_migration_sglang_worker_failure(
     indirect=True,
 )
 def test_no_request_migration_sglang_graceful_shutdown(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with graceful shutdown and migration disabled using SGLang.

tests/fault_tolerance/migration/test_trtllm.py

@@ -149,7 +149,7 @@ class DynamoWorkerProcess(ManagedProcess):
     indirect=True,
 )
 def test_request_migration_trtllm_worker_failure(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with migration support using TRT-LLM.
@@ -213,7 +213,7 @@ def test_request_migration_trtllm_worker_failure(
     indirect=True,
 )
 def test_request_migration_trtllm_graceful_shutdown(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with graceful shutdown and migration support using TRT-LLM.
@@ -281,7 +281,7 @@ def test_request_migration_trtllm_graceful_shutdown(
     indirect=True,
 )
 def test_no_request_migration_trtllm_worker_failure(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with migration disabled using TRT-LLM.
@@ -369,7 +369,7 @@ def test_no_request_migration_trtllm_worker_failure(
     indirect=True,
 )
 def test_no_request_migration_trtllm_graceful_shutdown(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with graceful shutdown and migration disabled using TRT-LLM.

tests/fault_tolerance/migration/test_vllm.py

@@ -159,7 +159,7 @@ class DynamoWorkerProcess(ManagedProcess):
     indirect=True,
 )
 def test_request_migration_vllm_worker_failure(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with migration support.
@@ -223,7 +223,7 @@ def test_request_migration_vllm_worker_failure(
     indirect=True,
 )
 def test_request_migration_vllm_graceful_shutdown(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with graceful shutdown and migration support.
@@ -291,7 +291,7 @@ def test_request_migration_vllm_graceful_shutdown(
     indirect=True,
 )
 def test_no_request_migration_vllm_worker_failure(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with migration disabled.
@@ -373,7 +373,7 @@ def test_no_request_migration_vllm_worker_failure(
     indirect=True,
 )
 def test_no_request_migration_vllm_graceful_shutdown(
-    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm
+    request, runtime_services_dynamic_ports, set_ucx_tls_no_mm, predownload_models
 ):
     """
     End-to-end test for worker fault tolerance with graceful shutdown and migration disabled.

tests/frontend/test_completion_mocker_engine.py

@@ -142,7 +142,7 @@ def runtime_services(request):
 
 
 @pytest.fixture(scope="module")
-def start_services(request, runtime_services):
+def start_services(request, runtime_services, predownload_tokenizers):
     """Start frontend and worker processes once for this module's tests."""
     with DynamoFrontendProcess(request):
         logger.info("Frontend started for tests")