[Deterministic] add deepseek v3 deterministic inference CI test (#12412)

test/srt/run_suite.py

@@ -75,6 +75,7 @@ suites = {
         TestFile("test_build_eagle_tree.py", 8),
         TestFile("test_chunked_prefill.py", 410),
         TestFile("test_create_kvindices.py", 2),
+        TestFile("test_deepseek_v3_deterministic.py", 240),
         TestFile("test_deterministic.py", 320),
         TestFile("test_eagle_infer_a.py", 370),
         TestFile("test_eagle_infer_b.py", 500),

test/srt/test_deepseek_v3_deterministic.py

@@ -0,0 +1,54 @@
+"""
+Usage:
+cd test/srt
+python3 -m unittest test_deepseek_v3_deterministic.TestFa3Deterministic
+"""
+
+import unittest
+
+from sglang.test.test_deterministic_utils import (
+    COMMON_SERVER_ARGS,
+    TestDeterministicBase,
+)
+
+DEEPSEEK_MODEL = "lmsys/sglang-ci-dsv3-test"
+
+
+class TestFa3Deterministic(TestDeterministicBase):
+    @classmethod
+    def get_model(cls):
+        return DEEPSEEK_MODEL
+
+    # Test with fa3 attention backend
+    @classmethod
+    def get_server_args(cls):
+        args = COMMON_SERVER_ARGS
+        args.extend(
+            [
+                "--attention-backend",
+                "fa3",
+            ]
+        )
+        return args
+
+
+class TestTritonDeterministic(TestDeterministicBase):
+    @classmethod
+    def get_model(cls):
+        return DEEPSEEK_MODEL
+
+    # Test with triton attention backend
+    @classmethod
+    def get_server_args(cls):
+        args = COMMON_SERVER_ARGS
+        args.extend(
+            [
+                "--attention-backend",
+                "triton",
+            ]
+        )
+        return args
+
+
+if __name__ == "__main__":
+    unittest.main()