[CI] Update nightly test models for H200/B200 (#22288)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

test/registered/8-gpu-models/test_deepseek_v31.py

@@ -1,14 +1,11 @@
 import unittest
 
 from sglang.test.accuracy_test_runner import AccuracyTestParams
-from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.performance_test_runner import PerformanceTestParams
 from sglang.test.run_combined_tests import run_combined_tests
 from sglang.test.test_utils import ModelLaunchSettings
 
-# Runs on both H200 and B200 via nightly-8-gpu-common suite
+# Manual-only: not registered in any CI suite
-register_cuda_ci(est_time=5400, suite="nightly-8-gpu-common", nightly=True)
-
 DEEPSEEK_V31_MODEL_PATH = "deepseek-ai/DeepSeek-V3.1"
 
 

test/registered/8-gpu-models/test_glm_46.py

@@ -1,52 +0,0 @@
-import unittest
-
-from sglang.test.accuracy_test_runner import AccuracyTestParams
-from sglang.test.ci.ci_register import register_cuda_ci
-from sglang.test.performance_test_runner import PerformanceTestParams
-from sglang.test.run_combined_tests import run_combined_tests
-from sglang.test.test_utils import ModelLaunchSettings
-
-# Runs on both H200 and B200 via nightly-8-gpu-common suite
-register_cuda_ci(est_time=1800, suite="nightly-8-gpu-common", nightly=True)
-
-GLM_4_6_MODEL_PATH = "zai-org/GLM-4.6"
-
-
-class TestGLM46(unittest.TestCase):
-    """Unified test class for GLM-4.6 performance and accuracy.
-
-    Single variant with simple TP=8 configuration.
-    GLM-4.6 is a 357B MoE model.
-    Runs BOTH:
-    - Performance test (using NightlyBenchmarkRunner)
-    - Accuracy test (using run_eval with mgsm_en)
-    """
-
-    def test_glm_46(self):
-        """Run performance and accuracy for GLM-4.6."""
-        base_args = [
-            "--tp=8",
-            "--trust-remote-code",
-        ]
-
-        variants = [
-            ModelLaunchSettings(
-                GLM_4_6_MODEL_PATH,
-                tp_size=8,
-                extra_args=base_args,
-                variant="TP8",
-            ),
-        ]
-
-        run_combined_tests(
-            models=variants,
-            test_name="GLM-4.6",
-            accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.80),
-            performance_params=PerformanceTestParams(
-                profile_dir="performance_profiles_glm_4_6",
-            ),
-        )
-
-
-if __name__ == "__main__":
-    unittest.main()

test/registered/8-gpu-models/test_glm_46_fp8.py

@@ -1,14 +1,11 @@
 import unittest
 
 from sglang.test.accuracy_test_runner import AccuracyTestParams
-from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.performance_test_runner import PerformanceTestParams
 from sglang.test.run_combined_tests import run_combined_tests
 from sglang.test.test_utils import ModelLaunchSettings
 
-# Runs on both H200 and B200 via nightly-8-gpu-common suite
+# Manual-only: not registered in any CI suite
-register_cuda_ci(est_time=1800, suite="nightly-8-gpu-common", nightly=True)
-
 GLM_4_6_FP8_MODEL_PATH = "zai-org/GLM-4.6-FP8"
 
 

test/registered/8-gpu-models/test_qwen35.py

@@ -9,7 +9,7 @@ from sglang.test.test_utils import ModelLaunchSettings
 # Runs on both H200 and B200 via nightly-8-gpu-common suite
 register_cuda_ci(est_time=1800, suite="nightly-8-gpu-common", nightly=True)
 
-QWEN35_MODEL_PATH = "Qwen/Qwen3.5-397B-A17B"
+QWEN35_MODEL_PATH = "Qwen/Qwen3.5-397B-A17B-FP8"
 
 
 class TestQwen35(unittest.TestCase):
@@ -30,6 +30,7 @@ class TestQwen35(unittest.TestCase):
             "--tool-call-parser=qwen3_coder",
             "--mem-fraction-static=0.8",
         ]
+        dp_args = ["--dp=8", "--enable-dp-attention"]
         mtp_args = [
             "--speculative-algorithm=EAGLE",
             "--speculative-num-steps=3",
@@ -48,8 +49,14 @@ class TestQwen35(unittest.TestCase):
             ModelLaunchSettings(
                 QWEN35_MODEL_PATH,
                 tp_size=8,
-                extra_args=base_args + mtp_args,
+                extra_args=base_args + dp_args,
-                variant="TP8+MTP",
+                variant="TP8+DP8",
+            ),
+            ModelLaunchSettings(
+                QWEN35_MODEL_PATH,
+                tp_size=8,
+                extra_args=base_args + dp_args + mtp_args,
+                variant="TP8+DP8+MTP",
                 env={"SGLANG_ENABLE_SPEC_V2": "1"},
             ),
         ]

test/registered/8-gpu-models/test_qwen3_235b.py

@@ -1,14 +1,11 @@
 import unittest
 
 from sglang.test.accuracy_test_runner import AccuracyTestParams
-from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.performance_test_runner import PerformanceTestParams
 from sglang.test.run_combined_tests import run_combined_tests
 from sglang.test.test_utils import ModelLaunchSettings, is_blackwell_system
 
-# Runs on both H200 and B200 via nightly-8-gpu-common suite
+# Manual-only: not registered in any CI suite
-register_cuda_ci(est_time=1800, suite="nightly-8-gpu-common", nightly=True)
-
 QWEN3_235B_FP8_MODEL_PATH = "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8"
 QWEN3_235B_EAGLE3_MODEL_PATH = (
     "lmsys/SGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-Meituan"