chore(guardrails): speed up IGW middleware integration test fixtures

plugins/example-plugin/tests/integration/conftest.py

@@ -1,15 +1,22 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Pytest fixture re-export for the example plugin integration tests.
+"""Pytest fixture re-exports for the example plugin integration tests.
 
-Re-exporting :func:`igw_plugin_harness` from a project-level ``conftest.py``
-is the standard pytest pattern for sharing a fixture across a test package
-without importing it at the top of every test module. Listing it in
-``__all__`` makes the re-export explicit so it isn't flagged as an unused
-import.
+The module-scope helpers (``_igw_app_context``, ``_igw_extra_services``)
+are re-imported so pytest can resolve :func:`igw_plugin_harness`'s
+dependency chain. The default empty ``_igw_extra_services`` tuple
+applies — no services beyond IGW + Models are mounted.
 """
 
-from nmp.core.inference_gateway.testing.fixtures import igw_plugin_harness
+from nmp.core.inference_gateway.testing.fixtures import (
+    _igw_app_context,
+    _igw_extra_services,
+    igw_plugin_harness,
+)
 
-__all__ = ["igw_plugin_harness"]
+__all__ = [
+    "_igw_app_context",
+    "_igw_extra_services",
+    "igw_plugin_harness",
+]

plugins/example-plugin/tests/integration/test_inference_middleware.py

@@ -34,7 +34,6 @@
 
 pytestmark = [pytest.mark.integration]
 
-DEFAULT_WORKSPACE = "default"
 EXAMPLE_PLUGIN_NAME = "nemo-example-middleware"
 EXAMPLE_PLUGIN_CONFIG_TYPE = ExampleMiddlewareConfig.__entity_type__
 
@@ -77,21 +76,21 @@ def test_safe_input_is_proxied_to_backend(self, igw_plugin_harness: IGWPluginHar
             ],
         )
         h.add_provider(
-            workspace=DEFAULT_WORKSPACE,
+            workspace=h.workspace,
             name=f"example-provider-{test_id}",
             served_models={model_name: model_name},
         )
 
         with h.use_plugin(EXAMPLE_PLUGIN_NAME, ExampleInferenceMiddleware()):
             h.add_virtual_model(
-                workspace=DEFAULT_WORKSPACE,
+                workspace=h.workspace,
                 name=virtual_model_name,
-                default_model_entity=f"{DEFAULT_WORKSPACE}/{model_name}",
+                default_model_entity=f"{h.workspace}/{model_name}",
                 request_middleware=[_build_middleware_call(blocked_keywords=["violence"])],
             )
 
             response = h.chat_completions(
-                workspace=DEFAULT_WORKSPACE,
+                workspace=h.workspace,
                 body={
                     "model": virtual_model_name,
                     "messages": [{"role": "user", "content": "Tell me about flowers."}],
@@ -109,16 +108,16 @@ def test_blocked_input_short_circuits_proxy(self, igw_plugin_harness: IGWPluginH
         block_message = "That topic is off-limits."
 
         h.add_provider(
-            workspace=DEFAULT_WORKSPACE,
+            workspace=h.workspace,
             name=f"example-provider-{test_id}",
             served_models={model_name: model_name},
         )
 
         with h.use_plugin(EXAMPLE_PLUGIN_NAME, ExampleInferenceMiddleware()):
             h.add_virtual_model(
-                workspace=DEFAULT_WORKSPACE,
+                workspace=h.workspace,
                 name=virtual_model_name,
-                default_model_entity=f"{DEFAULT_WORKSPACE}/{model_name}",
+                default_model_entity=f"{h.workspace}/{model_name}",
                 request_middleware=[
                     _build_middleware_call(
                         blocked_keywords=["violence"],
@@ -128,7 +127,7 @@ def test_blocked_input_short_circuits_proxy(self, igw_plugin_harness: IGWPluginH
             )
 
             response = h.chat_completions(
-                workspace=DEFAULT_WORKSPACE,
+                workspace=h.workspace,
                 body={
                     "model": virtual_model_name,
                     "messages": [{"role": "user", "content": "Tell me about violence."}],
@@ -171,21 +170,21 @@ def test_backend_response_is_redacted(self, igw_plugin_harness: IGWPluginHarness
             ],
         )
         h.add_provider(
-            workspace=DEFAULT_WORKSPACE,
+            workspace=h.workspace,
             name=f"example-provider-{test_id}",
             served_models={model_name: model_name},
         )
 
         with h.use_plugin(EXAMPLE_PLUGIN_NAME, ExampleInferenceMiddleware()):
             h.add_virtual_model(
-                workspace=DEFAULT_WORKSPACE,
+                workspace=h.workspace,
                 name=virtual_model_name,
-                default_model_entity=f"{DEFAULT_WORKSPACE}/{model_name}",
+                default_model_entity=f"{h.workspace}/{model_name}",
                 response_middleware=[_build_middleware_call(blocked_keywords=["secret"])],
             )
 
             response = h.chat_completions(
-                workspace=DEFAULT_WORKSPACE,
+                workspace=h.workspace,
                 body={
                     "model": virtual_model_name,
                     "messages": [{"role": "user", "content": "Share the answer."}],
@@ -233,21 +232,21 @@ def test_streaming_backend_response_is_redacted(self, igw_plugin_harness: IGWPlu
             ],
         )
         h.add_provider(
-            workspace=DEFAULT_WORKSPACE,
+            workspace=h.workspace,
             name=f"example-stream-provider-{test_id}",
             served_models={model_name: model_name},
         )
 
         with h.use_plugin(EXAMPLE_PLUGIN_NAME, ExampleInferenceMiddleware()):
             h.add_virtual_model(
-                workspace=DEFAULT_WORKSPACE,
+                workspace=h.workspace,
                 name=virtual_model_name,
-                default_model_entity=f"{DEFAULT_WORKSPACE}/{model_name}",
+                default_model_entity=f"{h.workspace}/{model_name}",
                 response_middleware=[_build_middleware_call(blocked_keywords=["secret"])],
             )
 
             chunks = h.stream_chat_completions(
-                workspace=DEFAULT_WORKSPACE,
+                workspace=h.workspace,
                 body={
                     "model": virtual_model_name,
                     "messages": [{"role": "user", "content": "Share the answer."}],

plugins/nemo-guardrails/tests/integration/conftest.py

@@ -3,23 +3,51 @@
 
 """Pytest fixtures for the guardrails plugin integration tests.
 
-Re-exports the IGW harness fixtures so test modules don't have to import
-them at the top of every file, and provides an autouse fixture that
-keeps ``nemoguardrails`` from reaching out to HuggingFace at startup.
-
-- :func:`igw_plugin_harness` — default; no real port for IGW.
-- :func:`igw_loopback_harness` — opt-in; IGW additionally bound on a real
-  ``127.0.0.1:<port>`` for tests that need IGW's loopback URL. Call it with
-  extra services to mount additional routes.
+Re-exports the IGW harness fixtures so test modules don't have to
+import them. ``_igw_extra_services`` is overridden below to mount
+:class:`GuardrailsService` on the module-scoped app — entity-backed
+guardrail-config tests need its CRUD routes. The module-scope helpers
+(``_igw_app_context``, ``_igw_loopback_context``) are re-imported so
+pytest can resolve the dependency chain from this conftest's scope.
+
+``HF_HUB_OFFLINE`` is set at conftest import time, **before** the
+``GuardrailsService`` import below — importing ``GuardrailsService``
+transitively imports ``nemoguardrails``, which reaches HuggingFace at
+import time if not told to stay offline. A function-scoped autouse
+``monkeypatch`` fixture would be too late: it doesn't run until after
+the module-scoped fixture setup that triggers these imports.
 """
 
+import os
+
 import pytest
-from nmp.core.inference_gateway.testing.fixtures import igw_loopback_harness, igw_plugin_harness
+from nmp.core.inference_gateway.testing.fixtures import (
+    _igw_app_context,
+    _igw_loopback_context,
+    igw_loopback_harness,
+    igw_plugin_harness,
+)
+from nmp.testing.client import ServiceFactory
+
+# Must precede the ``nemoguardrails``-pulling import below.
+os.environ.setdefault("HF_HUB_OFFLINE", "1")
+
+from nmp.guardrails.service import GuardrailsService  # noqa: E402
+
+__all__ = [
+    "_igw_app_context",
+    "_igw_loopback_context",
+    "igw_loopback_harness",
+    "igw_plugin_harness",
+]
 
-__all__ = ["igw_loopback_harness", "igw_plugin_harness"]
 
+@pytest.fixture(scope="module")
+def _igw_extra_services() -> tuple[ServiceFactory, ...]:
+    """Mount :class:`GuardrailsService` on the module-scoped IGW + Models app.
 
-@pytest.fixture(autouse=True)
-def offline_huggingface(monkeypatch: pytest.MonkeyPatch) -> None:
-    """Skip ``nemoguardrails`` HuggingFace tokenizer downloads — they time out offline."""
-    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    Every integration test here gets Guardrails CRUD routes whether it
+    uses them or not — the startup cost amortises across the module,
+    so files that only touch inline configs pay almost nothing extra.
+    """
+    return (GuardrailsService,)