[CI] Pin NCCL ports for GB300 PR tests (#33044)

test/registered/4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py

@@ -7,6 +7,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_DEEPSEEK_NVFP4_MODEL_FOR_TEST,
+    DEFAULT_PORT_FOR_SRT_TEST_RUNNER,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
@@ -16,6 +17,9 @@ from sglang.test.test_utils import (
 
 register_cuda_ci(est_time=1800, stage="base-c", runner_config="4-gpu-gb300")
 
+# Keep rendezvous ports below the ephemeral range on the 4-GPU GB300 runner.
+NCCL_PORT_BASE = DEFAULT_PORT_FOR_SRT_TEST_RUNNER + 100
+
 
 class TestDeepseekR1Nvfp4CuteDSLDeepEP(CustomTestCase):
     @classmethod
@@ -42,6 +46,8 @@ class TestDeepseekR1Nvfp4CuteDSLDeepEP(CustomTestCase):
             "--moe-dense-tp-size",
             "1",
             "--enable-dp-attention",
+            "--nccl-port",
+            str(NCCL_PORT_BASE),
             "--quantization",
             "modelopt_fp4",
             "--attention-backend",
@@ -114,6 +120,8 @@ class TestDummyWithSBO(CustomTestCase):
             "--moe-dense-tp-size",
             "1",
             "--enable-dp-attention",
+            "--nccl-port",
+            str(NCCL_PORT_BASE + 1),
             "--quantization",
             "modelopt_fp4",
             "--attention-backend",

test/registered/disaggregation/test_disaggregation_aarch64.py

@@ -8,6 +8,7 @@ from sglang.test.server_fixtures.disaggregation_fixture import (
     PDDisaggregationServerBase,
 )
 from sglang.test.test_utils import (
+    DEFAULT_PORT_FOR_SRT_TEST_RUNNER,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     popen_launch_pd_server,
 )
@@ -16,6 +17,9 @@ QWEN3_8B_MODEL_PATH = "Qwen/Qwen3-8B"
 
 register_cuda_ci(est_time=300, stage="base-c", runner_config="4-gpu-gb300")
 
+# Keep rendezvous ports below the ephemeral range on the 4-GPU GB300 runner.
+NCCL_PORT_BASE = DEFAULT_PORT_FOR_SRT_TEST_RUNNER + 120
+
 
 class TestDisaggregationMooncakeAARCH64Accuracy(PDDisaggregationServerBase):
     @classmethod
@@ -51,6 +55,8 @@ class TestDisaggregationMooncakeAARCH64Accuracy(PDDisaggregationServerBase):
             cls.bootstrap_port,
             "--tp",
             "2",
+            "--nccl-port",
+            str(NCCL_PORT_BASE),
         ]
         prefill_args += cls.transfer_backend + cls.rdma_devices
         cls.process_prefill = popen_launch_pd_server(
@@ -72,6 +78,8 @@ class TestDisaggregationMooncakeAARCH64Accuracy(PDDisaggregationServerBase):
             "2",
             "--base-gpu-id",
             "2",
+            "--nccl-port",
+            str(NCCL_PORT_BASE + 1),
         ]
         decode_args += cls.transfer_backend + cls.rdma_devices
         cls.process_decode = popen_launch_pd_server(

test/registered/ep/test_flashinfer_a2a.py

@@ -7,6 +7,7 @@ from sglang.srt.utils import kill_process_tree
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
+    DEFAULT_PORT_FOR_SRT_TEST_RUNNER,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
     popen_launch_server,
@@ -14,6 +15,9 @@ from sglang.test.test_utils import (
 
 register_cuda_ci(est_time=500, stage="base-c", runner_config="4-gpu-gb300")
 
+# Keep rendezvous ports below the ephemeral range on the 4-GPU GB300 runner.
+NCCL_PORT_BASE = DEFAULT_PORT_FOR_SRT_TEST_RUNNER + 110
+
 DEEPSEEK_V3_FP4_MODEL = "nvidia/DeepSeek-V3-0324-FP4"
 GLM52_NVFP4_MODEL = "nvidia/GLM-5.2-NVFP4"
 QWEN3_FP8_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct-FP8"
@@ -42,6 +46,8 @@ class TestFlashinferA2ATrtllmRoutedFP4(CustomTestCase):
                 "--dp",
                 "4",
                 "--enable-dp-attention",
+                "--nccl-port",
+                str(NCCL_PORT_BASE),
                 "--moe-a2a-backend",
                 "flashinfer",
                 "--moe-runner-backend",
@@ -93,6 +99,8 @@ class TestFlashinferA2ACutedslStaticFP4(CustomTestCase):
                 "--dp",
                 "4",
                 "--enable-dp-attention",
+                "--nccl-port",
+                str(NCCL_PORT_BASE + 1),
                 "--moe-a2a-backend",
                 "flashinfer",
                 "--moe-runner-backend",
@@ -152,6 +160,8 @@ class TestFlashinferA2ATrtllmRoutedFP8(CustomTestCase):
                 "--dp",
                 "4",
                 "--enable-dp-attention",
+                "--nccl-port",
+                str(NCCL_PORT_BASE + 2),
                 "--moe-a2a-backend",
                 "flashinfer",
                 "--moe-runner-backend",