[CI] Move manual-only nightly tests out of test/registered/ (#22298)

Co-authored-by: Alison Shao <alison.shao@Mac.attlocal.net>
2026-04-07 21:03:52 -07:00
parent 4e4b4ac153
commit 36f05810c9
3 changed files with 0 additions and 12 deletions
@@ -0,0 +1,108 @@
+import unittest
+
+from sglang.test.accuracy_test_runner import AccuracyTestParams
+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
+
+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"
+)
+
+
+class TestQwen3235BFP8(unittest.TestCase):
+    """Test class for Qwen3-235B-FP8 performance and accuracy.
+
+    Three variants:
+    - basic: TP=8
+    - eagle3: TP=8 + EP=2 + EAGLE3 speculative decoding
+    - TP8+CP2+EP2: TP=8 + CP=2 + EP=2 context parallel
+
+    Each variant runs BOTH:
+    - Performance test (using NightlyBenchmarkRunner)
+    - Accuracy test (using run_eval with gsm8k)
+    """
+
+    def test_qwen3_235b_fp8_all_variants(self):
+        """Run performance and accuracy for Qwen3-235B-FP8."""
+        base_args = [
+            "--tp=8",
+            "--ep=2",
+            "--trust-remote-code",
+        ]
+        eagle3_args = [
+            "--speculative-algorithm=EAGLE3",
+            f"--speculative-draft-model-path={QWEN3_235B_EAGLE3_MODEL_PATH}",
+            "--speculative-num-steps=3",
+            "--speculative-eagle-topk=1",
+            "--speculative-num-draft-tokens=4",
+        ]
+
+        variants = [
+            # Variant: "basic" - TP=8
+            ModelLaunchSettings(
+                QWEN3_235B_FP8_MODEL_PATH,
+                tp_size=8,
+                extra_args=base_args,
+                variant="TP8",
+            ),
+            # Variant: "eagle3" - TP=8 + EP=2 + EAGLE3 speculative decoding
+            ModelLaunchSettings(
+                QWEN3_235B_FP8_MODEL_PATH,
+                tp_size=8,
+                extra_args=base_args + eagle3_args,
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
+                variant="TP8+EP2+EAGLE3",
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="Qwen3-235B-FP8",
+            accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.88),
+            performance_params=PerformanceTestParams(
+                profile_dir="performance_profiles_qwen3_235b_fp8",
+            ),
+        )
+
+    @unittest.skipIf(is_blackwell_system(), "Requires H200 system")
+    def test_qwen3_235b_fp8_cp(self):
+        """Run performance and accuracy for Qwen3-235B-FP8 with context parallelism."""
+
+        BASE_ARGS = [
+            "--trust-remote-code",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true, "num_threads": 64}',
+        ]
+
+        DP_ARGS = [
+            "--tp=8",
+            "--moe-dp-size=2",
+            "--attn-cp-size=2",
+            "--ep-size=4",
+            "--enable-prefill-context-parallel",
+        ]
+
+        MTP_ARGS = [
+            "--cuda-graph-max-bs=32",
+            "--max-running-requests=32",
+        ]
+        variants = [
+            ModelLaunchSettings(
+                QWEN3_235B_FP8_MODEL_PATH,
+                tp_size=8,
+                extra_args=BASE_ARGS + DP_ARGS + MTP_ARGS,
+                variant="TP8+CP2+EP2",
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="Qwen3-235B-FP8 Context Parallel",
+            accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.88),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()