ci: run GB300 nightly suite in the standard Nvidia nightly workflow (#28536)

Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
Co-authored-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>

.github/workflows/nightly-test-nvidia.yml

@@ -24,6 +24,7 @@ on:
           - 'nightly-test-perf-4-gpu-b200'
           - 'nightly-test-perf-8-gpu-b200'
           - 'nightly-test-specialized-8-gpu-b200'
+          - 'nightly-test-perf-4-gpu-gb300'
           - 'nightly-test-kernel-1-gpu-h100'
           - 'nightly-test-diffusion-comparison'
           - 'nightly-test-kernel-8-gpu-h200'
@@ -512,6 +513,31 @@ jobs:
       - uses: ./.github/actions/upload-cuda-coredumps
         if: failure()
 
+  # GB300 (Grace-Blackwell NVL4) performance tests - 4 GPU (ARM64)
+  nightly-test-perf-4-gpu-gb300:
+    if: github.repository == 'sgl-project/sglang' && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-test-perf-4-gpu-gb300')
+    runs-on: 4-gpu-gb300
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.ref }}
+
+      - uses: ./.github/actions/check-maintenance
+
+      - name: Install dependencies
+        run: |
+          bash scripts/ci/cuda/ci_install_dependency.sh
+
+      - name: Run test
+        timeout-minutes: 600
+        run: |
+          cd test
+          python3 run_suite.py --hw cuda --suite nightly-4-gpu-gb300 --nightly --continue-on-error --timeout-per-file 7200
+
+      - uses: ./.github/actions/upload-cuda-coredumps
+        if: failure()
+
   # Specialized B200 tests - 8 GPU, for specific backends and configs
   nightly-test-specialized-8-gpu-b200:
     if: github.repository == 'sgl-project/sglang' && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-test-perf-8-gpu-b200' || inputs.job_filter == 'nightly-test-specialized-8-gpu-b200')
@@ -695,6 +721,7 @@ jobs:
       - nightly-test-vlm-perf-2-gpu-h100
       - nightly-test-perf-4-gpu-b200
       - nightly-test-specialized-8-gpu-b200
+      - nightly-test-perf-4-gpu-gb300
       - nightly-test-diffusion-comparison
       - nightly-test-precision-8-gpu-h200
       - consolidate-metrics

python/sglang/test/performance_test_runner.py

@@ -10,8 +10,8 @@ from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, ModelLaunchSettings
 class PerformanceTestParams:
     """Parameters for performance testing."""
 
-    batch_sizes: List[int] = field(default_factory=lambda: [1, 8, 16, 64])
+    batch_sizes: List[int] = field(default_factory=lambda: [1, 8, 16])
-    input_lens: Tuple[int, ...] = (4096,)
+    input_lens: Tuple[int, ...] = (8192,)
     output_lens: Tuple[int, ...] = (512,)
     profile_dir: Optional[str] = None  # None = auto-generate based on is_vlm
     dataset_name: str = "mmmu"  # For VLM perf test

test/registered/gb300/test_deepseek_v32.py

@@ -1,78 +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
-
-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",
-            ),
-        ]
-
-        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()

test/registered/gb300/test_deepseek_v32_nvfp4.py

@@ -1,81 +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
-
-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",
-            ),
-        ]
-
-        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()

test/registered/gb300/test_glm5_fp8.py

@@ -6,7 +6,9 @@ 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)
+register_cuda_ci(
+    est_time=7200, suite="nightly-4-gpu-gb300", nightly=True, disabled="not needed"
+)
 
 MODEL_PATH = "zai-org/GLM-5.1-FP8"
 
@@ -27,7 +29,7 @@ MTP_ARGS = [
 
 
 class TestGlm5Fp8(unittest.TestCase):
-    """GLM-5.1 FP8 on GB300 (4x B200 NVL4, tp=4)."""
+    """GLM-5.1 FP8 on GB300 (4x GB300 NVL4, tp=4)."""
 
     def test_glm5_fp8(self):
         variants = [

test/registered/gb300/test_glm5_nvfp4.py

@@ -16,42 +16,42 @@ COMMON_ARGS = [
     "--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 = [
+TP_MTP_ARGS = [
     "--speculative-algorithm=EAGLE",
     "--speculative-num-steps=3",
     "--speculative-eagle-topk=1",
     "--speculative-num-draft-tokens=4",
 ]
 
+DP_MTP_ARGS = [
+    "--speculative-algorithm=EAGLE",
+    "--speculative-num-steps=1",
+    "--speculative-eagle-topk=1",
+    "--speculative-num-draft-tokens=2",
+]
+
 
 class TestGlm5Nvfp4(unittest.TestCase):
-    """GLM-5 NVFP4 on GB300 (4x B200 NVL4, tp=4)."""
+    """GLM-5 NVFP4 on GB300 (4x GB300 NVL4, tp=4)."""
 
     def test_glm5_nvfp4(self):
         variants = [
             ModelLaunchSettings(
                 MODEL_PATH,
                 tp_size=4,
-                extra_args=COMMON_ARGS,
+                extra_args=COMMON_ARGS + TP_MTP_ARGS,
-                variant="TP4",
+                variant="TP4+MTP",
-            ),
-            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,
+                + DP_MTP_ARGS,
                 variant="TP4+DP4+DPA+MTP",
             ),
         ]

test/registered/gb300/test_kimi_k25.py

@@ -6,7 +6,9 @@ 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)
+register_cuda_ci(
+    est_time=7200, suite="nightly-4-gpu-gb300", nightly=True, disabled="not needed"
+)
 
 MODEL_PATH = "moonshotai/Kimi-K2.5"
 
@@ -21,7 +23,7 @@ COMMON_ARGS = [
 
 
 class TestKimiK25(unittest.TestCase):
-    """Kimi-K2.5 (native INT4) on GB300 (4x B200 NVL4, tp=4).
+    """Kimi-K2.5 (native INT4) on GB300 (4x GB300 NVL4, tp=4).
 
     No EAGLE/MTP support for Kimi-K2.5 — only TP and TP+DP+DPA variants.
     """

test/registered/gb300/test_kimi_k25_nvfp4.py

@@ -15,7 +15,8 @@ COMMON_ARGS = [
     "--reasoning-parser=kimi_k2",
     "--tool-call-parser=kimi_k2",
     "--quantization=modelopt_fp4",
-    "--attention-backend=trtllm_mla",
+    "--attention-backend=tokenspeed_mla",
+    "--kv-cache-dtype=fp8_e4m3",
     "--moe-runner-backend=flashinfer_trtllm",
     "--mem-fraction-static=0.8",
     "--enable-multimodal",
@@ -24,7 +25,7 @@ COMMON_ARGS = [
 
 
 class TestKimiK25Nvfp4(unittest.TestCase):
-    """Kimi-K2.5 NVFP4 on GB300 (4x B200 NVL4, tp=4).
+    """Kimi-K2.5 NVFP4 on GB300 (4x GB300 NVL4, tp=4).
 
     No EAGLE/MTP support for Kimi-K2.5 — only TP and TP+DP+DPA variants.
     """

test/registered/gb300/test_qwen35_fp8.py

@@ -17,43 +17,43 @@ COMMON_ARGS = [
     "--enable-flashinfer-allreduce-fusion",
     "--attention-backend=trtllm_mha",
     "--mem-fraction-static=0.8",
+    "--mamba-scheduler-strategy=extra_buffer",
     "--enable-multimodal",
     "--enable-metrics",
 ]
 
-MTP_ARGS = [
+TP_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",
+
+DP_MTP_ARGS = [
+    "--speculative-algorithm=EAGLE",
+    "--speculative-num-steps=1",
+    "--speculative-eagle-topk=1",
+    "--speculative-num-draft-tokens=2",
 ]
 
 
 class TestQwen35Fp8(unittest.TestCase):
-    """Qwen3.5-397B FP8 on GB300 (4x B200 NVL4, tp=4)."""
+    """Qwen3.5-397B FP8 on GB300 (4x GB300 NVL4, tp=4)."""
 
     def test_qwen35_fp8(self):
         variants = [
             ModelLaunchSettings(
                 MODEL_PATH,
                 tp_size=4,
-                extra_args=COMMON_ARGS,
+                extra_args=COMMON_ARGS + TP_MTP_ARGS,
-                variant="TP4",
+                variant="TP4+MTP",
-            ),
-            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,
+                + DP_MTP_ARGS,
                 variant="TP4+DP4+DPA+MTP",
             ),
         ]
@@ -62,7 +62,7 @@ class TestQwen35Fp8(unittest.TestCase):
             models=variants,
             test_name="Qwen3.5-397B-FP8",
             accuracy_params=AccuracyTestParams(
-                dataset="mmmu-pro", baseline_accuracy=0.78, repeat=1, max_tokens=32768
+                dataset="mmmu-pro", baseline_accuracy=0.76, repeat=1, max_tokens=32768
             ),
             performance_params=PerformanceTestParams(
                 profile_dir="performance_profiles_gb300",

test/registered/gb300/test_qwen35_nvfp4.py

@@ -6,7 +6,9 @@ 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)
+register_cuda_ci(
+    est_time=7200, suite="nightly-4-gpu-gb300", nightly=True, disabled="not needed"
+)
 
 MODEL_PATH = "nvidia/Qwen3.5-397B-A17B-NVFP4"
 
@@ -36,7 +38,7 @@ MTP_ARGS = [
 
 
 class TestQwen35Nvfp4(unittest.TestCase):
-    """Qwen3.5-397B NVFP4 on GB300 (4x B200 NVL4, tp=4)."""
+    """Qwen3.5-397B NVFP4 on GB300 (4x GB300 NVL4, tp=4)."""
 
     def test_qwen35_nvfp4(self):
         variants = [
@@ -66,7 +68,7 @@ class TestQwen35Nvfp4(unittest.TestCase):
             models=variants,
             test_name="Qwen3.5-397B-NVFP4",
             accuracy_params=AccuracyTestParams(
-                dataset="mmmu-pro", baseline_accuracy=0.78, repeat=1, max_tokens=32768
+                dataset="mmmu-pro", baseline_accuracy=0.76, repeat=1, max_tokens=32768
             ),
             performance_params=PerformanceTestParams(
                 profile_dir="performance_profiles_gb300",

test/run_suite.py

@@ -121,7 +121,7 @@ NIGHTLY_SUITES = {
         "nightly-eval-vlm-2-gpu",
         "nightly-perf-text-2-gpu",
         "nightly-perf-vlm-2-gpu",
-        # GB300 (4x B200 NVL4) nightly suite
+        # GB300 (4x GB300 NVL4) nightly suite
         "nightly-4-gpu-gb300",
         # Nightly precision regression (per-layer hidden state comparison)
         "nightly-precision-8-gpu-h200",