feat(ci): add GB300 nightly benchmark test suites (#21487)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-28 21:54:03 -07:00
parent 166e9090ee
commit 9d64a82173
11 changed files with 874 additions and 4 deletions
@@ -0,0 +1,79 @@
+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
+
+register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300", nightly=True)
+
+MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
+
+COMMON_ARGS = [
+    "--trust-remote-code",
+    "--reasoning-parser=deepseek-v3",
+    "--tool-call-parser=deepseekv32",
+    "--mem-fraction-static=0.8",
+    "--enable-metrics",
+]
+
+MTP_ARGS = [
+    "--speculative-algorithm=EAGLE",
+    "--speculative-num-steps=3",
+    "--speculative-eagle-topk=1",
+    "--speculative-num-draft-tokens=4",
+]
+
+
+class TestDeepseekV32(unittest.TestCase):
+    """DeepSeek V3.2 on GB300 (4x B200 NVL4, tp=4)."""
+
+    def test_deepseek_v32(self):
+        variants = [
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS,
+                variant="TP4",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS
+                + [
+                    "--dp-size=4",
+                    "--ep-size=4",
+                    "--enable-dp-attention",
+                ],
+                variant="TP4+DP4+DPA",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS
+                + [
+                    "--dp-size=4",
+                    "--ep-size=4",
+                    "--enable-dp-attention",
+                ]
+                + MTP_ARGS,
+                variant="TP4+DP4+DPA+MTP",
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="DeepSeek-V3.2",
+            accuracy_params=AccuracyTestParams(
+                dataset="gsm8k", baseline_accuracy=0.935
+            ),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_gb300",
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,82 @@
+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
+
+register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300", nightly=True)
+
+MODEL_PATH = "nvidia/DeepSeek-V3.2-NVFP4"
+
+COMMON_ARGS = [
+    "--trust-remote-code",
+    "--reasoning-parser=deepseek-v3",
+    "--tool-call-parser=deepseekv32",
+    "--quantization=modelopt_fp4",
+    "--moe-runner-backend=flashinfer_trtllm",
+    "--kv-cache-dtype=bfloat16",
+    "--mem-fraction-static=0.8",
+    "--enable-metrics",
+]
+
+MTP_ARGS = [
+    "--speculative-algorithm=EAGLE",
+    "--speculative-num-steps=3",
+    "--speculative-eagle-topk=1",
+    "--speculative-num-draft-tokens=4",
+]
+
+
+class TestDeepseekV32Nvfp4(unittest.TestCase):
+    """DeepSeek V3.2 NVFP4 on GB300 (4x B200 NVL4, tp=4)."""
+
+    def test_deepseek_v32_nvfp4(self):
+        variants = [
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS,
+                variant="TP4",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS
+                + [
+                    "--dp-size=4",
+                    "--ep-size=4",
+                    "--enable-dp-attention",
+                ],
+                variant="TP4+DP4+DPA",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS
+                + [
+                    "--dp-size=4",
+                    "--ep-size=4",
+                    "--enable-dp-attention",
+                ]
+                + MTP_ARGS,
+                variant="TP4+DP4+DPA+MTP",
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="DeepSeek-V3.2-NVFP4",
+            accuracy_params=AccuracyTestParams(
+                dataset="gsm8k", baseline_accuracy=0.935
+            ),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_gb300",
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,68 @@
+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
+
+register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300", nightly=True)
+
+MODEL_PATH = "zai-org/GLM-5-FP8"
+
+COMMON_ARGS = [
+    "--trust-remote-code",
+    "--reasoning-parser=glm45",
+    "--tool-call-parser=glm47",
+    "--mem-fraction-static=0.9",
+    "--enable-metrics",
+]
+
+MTP_ARGS = [
+    "--speculative-algorithm=EAGLE",
+    "--speculative-num-steps=3",
+    "--speculative-eagle-topk=1",
+    "--speculative-num-draft-tokens=4",
+]
+
+
+class TestGlm5Fp8(unittest.TestCase):
+    """GLM-5 FP8 on GB300 (4x B200 NVL4, tp=4)."""
+
+    def test_glm5_fp8(self):
+        variants = [
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS,
+                variant="TP4",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS + ["--dp-size=4", "--enable-dp-attention"],
+                variant="TP4+DP4+DPA",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS
+                + ["--dp-size=4", "--enable-dp-attention"]
+                + MTP_ARGS,
+                variant="TP4+DP4+DPA+MTP",
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="GLM-5-FP8",
+            accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.92),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_gb300",
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,71 @@
+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
+
+register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300", nightly=True)
+
+MODEL_PATH = "nvidia/GLM-5-NVFP4"
+
+COMMON_ARGS = [
+    "--trust-remote-code",
+    "--reasoning-parser=glm45",
+    "--tool-call-parser=glm47",
+    "--quantization=modelopt_fp4",
+    "--moe-runner-backend=flashinfer_trtllm",
+    "--kv-cache-dtype=bfloat16",
+    "--mem-fraction-static=0.9",
+    "--enable-metrics",
+]
+
+MTP_ARGS = [
+    "--speculative-algorithm=EAGLE",
+    "--speculative-num-steps=3",
+    "--speculative-eagle-topk=1",
+    "--speculative-num-draft-tokens=4",
+]
+
+
+class TestGlm5Nvfp4(unittest.TestCase):
+    """GLM-5 NVFP4 on GB300 (4x B200 NVL4, tp=4)."""
+
+    def test_glm5_nvfp4(self):
+        variants = [
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS,
+                variant="TP4",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS + ["--dp-size=4", "--enable-dp-attention"],
+                variant="TP4+DP4+DPA",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS
+                + ["--dp-size=4", "--enable-dp-attention"]
+                + MTP_ARGS,
+                variant="TP4+DP4+DPA+MTP",
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="GLM-5-NVFP4",
+            accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.92),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_gb300",
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,58 @@
+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
+
+register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300", nightly=True)
+
+MODEL_PATH = "moonshotai/Kimi-K2.5"
+
+COMMON_ARGS = [
+    "--trust-remote-code",
+    "--reasoning-parser=kimi_k2",
+    "--tool-call-parser=kimi_k2",
+    "--mem-fraction-static=0.8",
+    "--enable-multimodal",
+    "--enable-metrics",
+]
+
+
+class TestKimiK25(unittest.TestCase):
+    """Kimi-K2.5 (native INT4) on GB300 (4x B200 NVL4, tp=4).
+
+    No EAGLE/MTP support for Kimi-K2.5 — only TP and TP+DP+DPA variants.
+    """
+
+    def test_kimi_k25(self):
+        variants = [
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS,
+                variant="TP4",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS + ["--dp-size=4", "--enable-dp-attention"],
+                variant="TP4+DP4+DPA",
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="Kimi-K2.5",
+            accuracy_params=AccuracyTestParams(
+                dataset="mmmu-pro", baseline_accuracy=0.69, repeat=1, max_tokens=32768
+            ),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_gb300",
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,61 @@
+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
+
+register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300", nightly=True)
+
+MODEL_PATH = "nvidia/Kimi-K2.5-NVFP4"
+
+COMMON_ARGS = [
+    "--trust-remote-code",
+    "--reasoning-parser=kimi_k2",
+    "--tool-call-parser=kimi_k2",
+    "--quantization=modelopt_fp4",
+    "--attention-backend=trtllm_mla",
+    "--moe-runner-backend=flashinfer_trtllm",
+    "--mem-fraction-static=0.8",
+    "--enable-multimodal",
+    "--enable-metrics",
+]
+
+
+class TestKimiK25Nvfp4(unittest.TestCase):
+    """Kimi-K2.5 NVFP4 on GB300 (4x B200 NVL4, tp=4).
+
+    No EAGLE/MTP support for Kimi-K2.5 — only TP and TP+DP+DPA variants.
+    """
+
+    def test_kimi_k25_nvfp4(self):
+        variants = [
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS,
+                variant="TP4",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS + ["--dp-size=4", "--enable-dp-attention"],
+                variant="TP4+DP4+DPA",
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="Kimi-K2.5-NVFP4",
+            accuracy_params=AccuracyTestParams(
+                dataset="mmmu-pro", baseline_accuracy=0.69, repeat=1, max_tokens=32768
+            ),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_gb300",
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,75 @@
+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
+
+register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300", nightly=True)
+
+MODEL_PATH = "Qwen/Qwen3.5-397B-A17B-FP8"
+
+COMMON_ARGS = [
+    "--trust-remote-code",
+    "--reasoning-parser=qwen3",
+    "--tool-call-parser=qwen3_coder",
+    "--enable-flashinfer-allreduce-fusion",
+    "--attention-backend=trtllm_mha",
+    "--mem-fraction-static=0.8",
+    "--enable-multimodal",
+    "--enable-metrics",
+]
+
+MTP_ARGS = [
+    "--speculative-algorithm=EAGLE",
+    "--speculative-num-steps=3",
+    "--speculative-eagle-topk=1",
+    "--speculative-num-draft-tokens=4",
+    "--mamba-scheduler-strategy=extra_buffer",
+    "--page-size=64",
+]
+
+
+class TestQwen35Fp8(unittest.TestCase):
+    """Qwen3.5-397B FP8 on GB300 (4x B200 NVL4, tp=4)."""
+
+    def test_qwen35_fp8(self):
+        variants = [
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS,
+                variant="TP4",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS + ["--dp-size=4", "--enable-dp-attention"],
+                variant="TP4+DP4+DPA",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS
+                + ["--dp-size=4", "--enable-dp-attention"]
+                + MTP_ARGS,
+                variant="TP4+DP4+DPA+MTP",
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="Qwen3.5-397B-FP8",
+            accuracy_params=AccuracyTestParams(
+                dataset="mmmu-pro", baseline_accuracy=0.78, repeat=1, max_tokens=32768
+            ),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_gb300",
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,79 @@
+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
+
+register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300", nightly=True)
+
+MODEL_PATH = "nvidia/Qwen3.5-397B-A17B-NVFP4"
+
+COMMON_ARGS = [
+    "--trust-remote-code",
+    "--reasoning-parser=qwen3",
+    "--tool-call-parser=qwen3_coder",
+    "--quantization=modelopt_fp4",
+    "--fp4-gemm-backend=flashinfer_cutlass",
+    "--moe-runner-backend=flashinfer_trtllm",
+    "--kv-cache-dtype=fp8_e4m3",
+    "--enable-flashinfer-allreduce-fusion",
+    "--attention-backend=trtllm_mha",
+    "--mem-fraction-static=0.8",
+    "--enable-multimodal",
+    "--enable-metrics",
+]
+
+MTP_ARGS = [
+    "--speculative-algorithm=EAGLE",
+    "--speculative-num-steps=3",
+    "--speculative-eagle-topk=1",
+    "--speculative-num-draft-tokens=4",
+    "--mamba-scheduler-strategy=extra_buffer",
+    "--page-size=64",
+]
+
+
+class TestQwen35Nvfp4(unittest.TestCase):
+    """Qwen3.5-397B NVFP4 on GB300 (4x B200 NVL4, tp=4)."""
+
+    def test_qwen35_nvfp4(self):
+        variants = [
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS,
+                variant="TP4",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS + ["--dp-size=4", "--enable-dp-attention"],
+                variant="TP4+DP4+DPA",
+            ),
+            ModelLaunchSettings(
+                MODEL_PATH,
+                tp_size=4,
+                extra_args=COMMON_ARGS
+                + ["--dp-size=4", "--enable-dp-attention"]
+                + MTP_ARGS,
+                variant="TP4+DP4+DPA+MTP",
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="Qwen3.5-397B-NVFP4",
+            accuracy_params=AccuracyTestParams(
+                dataset="mmmu-pro", baseline_accuracy=0.78, repeat=1, max_tokens=32768
+            ),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_gb300",
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -84,6 +84,8 @@ NIGHTLY_SUITES = {
        "nightly-eval-vlm-2-gpu",
        "nightly-perf-text-2-gpu",
        "nightly-perf-vlm-2-gpu",
+        # GB300 (4x B200 NVL4) nightly suite
+        "nightly-4-gpu-gb300",
    ],
    HWBackend.AMD: [
        "nightly-amd",