Simplify tests & Fix trtllm custom allreduce registration (#4252)

test/srt/test_bench_one_batch.py

@@ -3,6 +3,7 @@ import unittest
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_MOE_MODEL_NAME_FOR_TEST,
+    get_bool_env_var,
     is_in_ci,
     run_bench_one_batch,
     write_github_step_summary,
@@ -27,9 +28,13 @@ class TestBenchOneBatch(unittest.TestCase):
             DEFAULT_MOE_MODEL_NAME_FOR_TEST, ["--tp", "2", "--cuda-graph-max-bs", "2"]
         )
 
+        use_vllm_custom_allreduce = get_bool_env_var(
+            "USE_VLLM_CUSTOM_ALLREDUCE", default="true"
+        )
+
         if is_in_ci():
             write_github_step_summary(
-                f"### test_moe_tp2_bs1\n"
+                f"### test_moe_tp2_bs1 ({use_vllm_custom_allreduce=})\n"
                 f"output_throughput : {output_throughput:.2f} token/s\n"
             )
             self.assertGreater(output_throughput, 124)