From cf346bb15de90c447ed498ad5e99aae407e04ce2 Mon Sep 17 00:00:00 2001
From: Kangyan-Zhou <zky314343421@gmail.com>
Date: Thu, 30 Apr 2026 16:45:10 -0700
Subject: [PATCH] [CI] Skip worker-dependent SMG e2e tests pending runner-image
 debug (#24166)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .github/workflows/pr-test-rust.yml            | 12 ++++++-
 .../python/tests/test_startup_sequence.py     |  7 ++++-
 sgl-model-gateway/e2e_test/fixtures/hooks.py  | 31 +++++++++++++++++++
 3 files changed, 48 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/pr-test-rust.yml b/.github/workflows/pr-test-rust.yml
index 8ead93319..feaa94e2d 100644
--- a/.github/workflows/pr-test-rust.yml
+++ b/.github/workflows/pr-test-rust.yml
@@ -5,11 +5,17 @@ on:
     branches: [ main ]
     paths:
       - "sgl-model-gateway/**"
+      - ".github/workflows/pr-test-rust.yml"
+      - "scripts/ci/cuda/ci_install_dependency.sh"
+      - "scripts/ci/cuda/ci_install_gateway_dependencies.sh"
   pull_request:
     branches: [ main ]
     types: [opened, synchronize, reopened, labeled]
     paths:
       - "sgl-model-gateway/**"
+      - ".github/workflows/pr-test-rust.yml"
+      - "scripts/ci/cuda/ci_install_dependency.sh"
+      - "scripts/ci/cuda/ci_install_gateway_dependencies.sh"
   workflow_dispatch:
 
 concurrency:
@@ -344,7 +350,11 @@ jobs:
   summarize-benchmarks:
     needs: gateway-e2e
     runs-on: ubuntu-latest
-    if: success()
+    # Disabled while e2e tests are skipped in
+    # sgl-model-gateway/e2e_test/fixtures/hooks.py — no benchmarks run, so the
+    # genai-bench-results-all-policies artifact is never produced and the
+    # download step would fail. Re-enable together with the skip removal.
+    if: false
 
     steps:
     - name: Checkout code
diff --git a/sgl-model-gateway/bindings/python/tests/test_startup_sequence.py b/sgl-model-gateway/bindings/python/tests/test_startup_sequence.py
index 6a40a67f4..8ae727d02 100644
--- a/sgl-model-gateway/bindings/python/tests/test_startup_sequence.py
+++ b/sgl-model-gateway/bindings/python/tests/test_startup_sequence.py
@@ -649,7 +649,10 @@ def _install_sglang_stubs(monkeypatch):
     entry_mod = types.ModuleType("sglang.srt.entrypoints")
     http_server_mod = types.ModuleType("sglang.srt.entrypoints.http_server")
     server_args_mod = types.ModuleType("sglang.srt.server_args")
+    # sglang.srt.utils was refactored from a module into a package; launch_server.py
+    # imports from sglang.srt.utils.network, so the stub must model the submodule.
     utils_mod = types.ModuleType("sglang.srt.utils")
+    network_mod = types.ModuleType("sglang.srt.utils.network")
 
     def launch_server(_args):
         return None
@@ -684,7 +687,8 @@ def _install_sglang_stubs(monkeypatch):
 
     http_server_mod.launch_server = launch_server
     server_args_mod.ServerArgs = ServerArgs
-    utils_mod.is_port_available = is_port_available
+    network_mod.is_port_available = is_port_available
+    utils_mod.network = network_mod
 
     # Also stub external deps imported at module top-level
     def _dummy_get(*_a, **_k):
@@ -706,6 +710,7 @@ def _install_sglang_stubs(monkeypatch):
     )
     monkeypatch.setitem(sys.modules, "sglang.srt.server_args", server_args_mod)
     monkeypatch.setitem(sys.modules, "sglang.srt.utils", utils_mod)
+    monkeypatch.setitem(sys.modules, "sglang.srt.utils.network", network_mod)
 
 
 def test_router_defaults_and_start(monkeypatch):
diff --git a/sgl-model-gateway/e2e_test/fixtures/hooks.py b/sgl-model-gateway/e2e_test/fixtures/hooks.py
index 2cbbb427a..ed8a5632b 100644
--- a/sgl-model-gateway/e2e_test/fixtures/hooks.py
+++ b/sgl-model-gateway/e2e_test/fixtures/hooks.py
@@ -233,6 +233,37 @@ def pytest_collection_modifyitems(
     else:
         logger.info("Scanned worker requirements: (none)")
 
+    # TEMPORARY: skip every test that would launch an sglang worker subprocess.
+    #
+    # Workers crash at import inside transformers.integrations.hub_kernels →
+    # kernels.deps with:
+    #   StrictDataclassFieldValidationError: Validation error for field
+    #   'import_name': TypeError: Unsupported type for field 'import_name': str | None
+    #
+    # The package combo (kernels==0.13.0 + huggingface_hub==1.12.2 +
+    # transformers==5.6.0) is NOT the bug — it imports cleanly in fresh
+    # python:3.10-slim and lmsysorg/sglang:dev containers. The crash is specific
+    # to the 4-gpu-a10 runner image (most likely a stale/partial huggingface_hub
+    # install where _BASIC_TYPE_VALIDATORS[types.UnionType] registration is
+    # missing). Remove this skip once the runner image is rebuilt.
+    #
+    # Filter on `model_pool` in fixturenames (covers setup_backend, model_client,
+    # model_base_url, backend_router — all transitively depend on model_pool) so
+    # tests without an explicit `@pytest.mark.e2e` marker are also skipped. Then
+    # clear scanned worker requirements so model_pool — if realized for any
+    # non-skipped fixture — starts empty and never spawns a worker.
+    skip_marker = pytest.mark.skip(
+        reason="worker-dependent tests disabled: SMG runner image crash on transformers.integrations.hub_kernels import"
+    )
+    for item in items:
+        if (
+            item.get_closest_marker("e2e") is not None
+            or "model_pool" in item.fixturenames
+        ):
+            item.add_marker(skip_marker)
+    _worker_counts.clear()
+    _first_seen_order.clear()
+
 
 # ---------------------------------------------------------------------------
 # Pool requirements