[Test] Add fwd_occupancy sanity kit (#25886)

test/registered/core/test_basic_sanity.py

@@ -1,6 +1,7 @@
-"""Basic sanity: small-but-broad server smoke that downstream stages
-depend on. Three sanity kits, one shared server, covering protocol
-contract, decode correctness, and scheduler stress paths."""
+"""Stage-a basic sanity: small-but-broad server smoke that downstream
+stages depend on. Multiple sanity-kit mixins driving one shared server,
+covering protocol, decode correctness, scheduler stress, occupancy, and
+hellaswag accuracy."""
 
 import unittest
 
@@ -9,6 +10,8 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.kits.basic_api_contract_kit import BasicAPIContractMixin
 from sglang.test.kits.basic_decode_correctness_kit import BasicDecodeCorrectnessMixin
 from sglang.test.kits.basic_scheduler_stress_kit import BasicSchedulerStressMixin
+from sglang.test.kits.fwd_occupancy_kit import FwdOccupancyMixin
+from sglang.test.kits.hellaswag_kit import HellaswagMixin
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -17,17 +20,22 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=120, stage="base-a", runner_config="1-gpu-small")
-register_amd_ci(est_time=120, suite="stage-a-test-1-gpu-small-amd")
+register_cuda_ci(est_time=160, stage="base-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=160, suite="stage-a-test-1-gpu-small-amd")
 
 
 class TestBasicSanity(
     BasicAPIContractMixin,
     BasicDecodeCorrectnessMixin,
     BasicSchedulerStressMixin,
+    FwdOccupancyMixin,
+    HellaswagMixin,
     CustomTestCase,
 ):
     served_model_name = DEFAULT_MODEL_NAME_FOR_TEST
+    # 5090 + Llama-3.1-8B single-batch decode with overlap scheduler +
+    # cuda graph measured ~99 median in CI; keep ~2pp headroom.
+    fwd_occupancy_threshold = 97.0
 
     @classmethod
     def setUpClass(cls):
@@ -43,32 +51,13 @@ class TestBasicSanity(
                 "0.7",
                 "--enable-metrics",
             ],
+            env={"SGLANG_ENABLE_METRICS_DEVICE_TIMER": "1"},
         )
 
     @classmethod
     def tearDownClass(cls):
         kill_process_tree(cls.process.pid)
 
-    def test_accuracy_floor(self):
-        # Stage-a-private accuracy guard: hellaswag via the frontend DSL
-        # bound to this server. Catches systematic regressions that pass
-        # every cheap probe in the mixed-in kits but tank multi-choice
-        # reasoning. Not part of any reusable mixin -- accuracy gating
-        # is the gate test's own responsibility.
-        import sglang as sgl
-        from sglang.test.test_programs import test_hellaswag_select
-
-        sgl.set_default_backend(sgl.RuntimeEndpoint(self.base_url))
-        try:
-            accuracy, _ = test_hellaswag_select()
-        finally:
-            sgl.set_default_backend(None)
-        self.assertGreater(
-            accuracy,
-            0.60,
-            f"hellaswag accuracy floor breached: {accuracy:.3f}",
-        )
-
 
 if __name__ == "__main__":
     unittest.main()

test/registered/core/test_basic_sanity_eagle3.py

@@ -0,0 +1,79 @@
+"""Stage-a basic sanity with EAGLE3 spec decoding enabled. Mirrors
+test_basic_sanity.py with the spec-decoding path active."""
+
+import unittest
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.kits.basic_api_contract_kit import BasicAPIContractMixin
+from sglang.test.kits.basic_decode_correctness_kit import BasicDecodeCorrectnessMixin
+from sglang.test.kits.basic_scheduler_stress_kit import BasicSchedulerStressMixin
+from sglang.test.kits.fwd_occupancy_kit import FwdOccupancyMixin
+from sglang.test.kits.hellaswag_kit import HellaswagMixin
+from sglang.test.test_utils import (
+    DEFAULT_DRAFT_MODEL_EAGLE3,
+    DEFAULT_TARGET_MODEL_EAGLE3,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_cuda_ci(est_time=200, stage="base-a", runner_config="1-gpu-small")
+register_amd_ci(est_time=200, suite="stage-a-test-1-gpu-small-amd")
+
+
+class TestBasicSanityEagle3(
+    BasicAPIContractMixin,
+    BasicDecodeCorrectnessMixin,
+    BasicSchedulerStressMixin,
+    FwdOccupancyMixin,
+    HellaswagMixin,
+    CustomTestCase,
+):
+    served_model_name = DEFAULT_TARGET_MODEL_EAGLE3
+    # Match vanilla gate at 97; EAGLE3 spec should sustain similar
+    # single-batch occupancy. Adjust per CI calibration.
+    fwd_occupancy_threshold = 97.0
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            DEFAULT_TARGET_MODEL_EAGLE3,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                # Canonical EAGLE3 sglang config: fp16 + triton attention.
+                # bf16 + flashinfer cutlass RMSNorm hits a SM120 dtype
+                # mismatch on the draft model's input_layernorm.
+                "--dtype",
+                "float16",
+                "--attention-backend",
+                "triton",
+                "--speculative-algorithm",
+                "EAGLE3",
+                "--speculative-draft-model-path",
+                DEFAULT_DRAFT_MODEL_EAGLE3,
+                "--speculative-num-steps",
+                "3",
+                "--speculative-eagle-topk",
+                "1",
+                "--speculative-num-draft-tokens",
+                "4",
+                "--cuda-graph-max-bs",
+                "4",
+                "--mem-fraction-static",
+                "0.7",
+                "--enable-metrics",
+            ],
+            env={"SGLANG_ENABLE_METRICS_DEVICE_TIMER": "1"},
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()