Migrate 4-GPU/8-GPU workflow jobs to stage-c and add CI registry decorators (#17299)

test/registered/models/test_dummy_grok_models.py

@@ -0,0 +1,41 @@
+import unittest
+
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.test_utils import CustomTestCase, is_in_ci, run_bench_one_batch
+
+register_cuda_ci(
+    est_time=120,
+    suite="stage-b-test-large-2-gpu",
+    disabled="Temporarily disabled",
+)
+
+
+class TestDummyGrok1(CustomTestCase):
+
+    def test_dummy_grok_1(self):
+        _, output_throughput, _ = run_bench_one_batch(
+            None,
+            [
+                "--model",
+                "/dummy-grok",
+                "--tokenizer-path",
+                "Xenova/grok-1-tokenizer",
+                "--batch-size",
+                "2",
+                "--tp",
+                "2",
+                "--quantization",
+                "fp8",
+                "--load-format",
+                "dummy",
+                "--json-model-override-args",
+                '{"num_hidden_layers": 2}',
+            ],
+        )
+
+        if is_in_ci():
+            self.assertGreater(output_throughput, 0)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/models/test_ministral3_models.py

@@ -0,0 +1,33 @@
+import unittest
+
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
+from sglang.test.kits.mmmu_vlm_kit import MMMUMixin
+from sglang.test.server_fixtures.default_fixture import DefaultServerBase
+from sglang.test.server_fixtures.mmmu_fixture import MMMUServerBase
+
+register_cuda_ci(
+    est_time=200,
+    suite="stage-b-test-small-1-gpu",
+    disabled="Temporarily disabled",
+)
+
+MODEL = "mistralai/Ministral-3-3B-Instruct-2512"
+
+
+class TestMinistral3TextOnly(GSM8KMixin, DefaultServerBase):
+    gsm8k_accuracy_thres = 0.6
+    model = MODEL
+    other_args = ["--trust-remote-code"]
+
+
+class TestMinistral3MMMU(MMMUMixin, MMMUServerBase):
+    accuracy = 0.3
+    model = MODEL
+    other_args = ["--trust-remote-code"]
+    mmmu_args = ["--limit=0.1"]
+    """`--limit=0.1`: 10 percent of each task - this is fine for testing since the nominal result isn't interesting - this run is just to prevent relative regressions."""
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/models/test_qwen3_next_models_pcg.py

@@ -0,0 +1,77 @@
+"""
+Qwen3 Next piecewise CUDA graph tests.
+
+DISABLED: See https://github.com/sgl-project/sglang/issues/17039
+PCG tests for Qwen3 Next have intermittent failures (5-10% probability).
+Investigation ongoing by @YuweiAn.
+"""
+
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.few_shot_gsm8k import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_cuda_ci(
+    est_time=400,
+    suite="stage-c-test-4-gpu-h100",
+    disabled="Intermittent failures, see #17039",
+)
+
+QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
+
+ACC_THRESHOLDS = {
+    QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
+}
+
+
+@unittest.skip("Disabled: intermittent failures, see #17039")
+class TestQwen3NextPiecewiseCudaGraph(CustomTestCase):
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = QWEN3_NEXT_MODEL
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--tp",
+                "4",
+                "--enable-piecewise-cuda-graph",
+                "--piecewise-cuda-graph-compiler",
+                "eager",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval(args)
+        print(f"{metrics=}")
+        self.assertGreaterEqual(
+            metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()