[CI] Restore SMG e2e on 2-gpu-h100 / 4-gpu-h100 runners (#24222)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

sgl-model-gateway/e2e_test/pyproject.toml

@@ -9,9 +9,7 @@ dependencies = [
   "grpcio-health-checking",
   "httpx",
   "openai",
-  "py",  # Required for pytest-parallel with newer pytest versions
   "pytest",
-  "pytest-parallel",
   "pytest-rerunfailures",
 ]
 
@@ -32,13 +30,10 @@ addopts = "-v -s"
 # We configure logging manually in conftest.py
 log_cli = false
 
-# Parallel execution configuration:
-# Use --workers 1 --tests-per-worker N to run N tests concurrently as threads
-# within a single process. This enables true shared-worker parallelism where
-# the session-scoped model_pool fixture is shared across all threads.
-#
-# Example usage:
-#   pytest --workers 1 --tests-per-worker 4 e2e_test/router/
-#
-# The thread-safe ModelPool and GPUAllocator classes enable safe concurrent
-# access from multiple test threads.
+# Tests run serially under plain pytest. The pytest-parallel plugin
+# (last release 2019) was tried but its thread dispatch leaks fixture
+# references between tests, causing model_pool deadlocks; the parallel
+# speedup never materialized on a 2-GPU runner since the suite is
+# eviction-bound across 5 model:mode combos. ModelPool / GPUAllocator
+# remain thread-safe so re-introducing parallelism (xdist or otherwise)
+# stays a tractable option.