Support page size > 1 (#4356)

test/srt/run_suite.py

@@ -45,6 +45,7 @@ suites = {
         TestFile("test_no_overlap_scheduler.py", 262),
         TestFile("test_openai_server.py", 124),
         TestFile("test_penalty.py", 41),
+        TestFile("test_page_size.py", 60),
         TestFile("test_pytorch_sampling_backend.py", 66),
         TestFile("test_radix_attention.py", 167),
         TestFile("test_reasoning_content.py", 89),

test/srt/test_dp_attention.py

@@ -42,7 +42,8 @@ class TestDPAttention(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.5
+        print(f"{metrics=}")
+        self.assertGreater(metrics["score"], 0.5)
 
     def test_mgsm_en(self):
         args = SimpleNamespace(
@@ -54,7 +55,8 @@ class TestDPAttention(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.8
+        print(f"{metrics=}")
+        self.assertGreater(metrics["score"], 0.8)
 
 
 if __name__ == "__main__":

test/srt/test_gptqmodel_dynamic.py

@@ -184,6 +184,7 @@ class TestGPTQModelDynamicWithMarlin(unittest.TestCase):
                 "text": "The capital of France is",
                 "sampling_params": {
                     "max_new_tokens": max_new_tokens,
+                    "temperature": 0.001,
                 },
             },
         )

test/srt/test_mla_deepseek_v3.py

@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
 )
 
 
-class TestDeepseekV3(unittest.TestCase):
+class TestMLADeepseekV3(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = "lmsys/sglang-ci-dsv3-test"

test/srt/test_page_size.py

@@ -0,0 +1,46 @@
+import os
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
+
+
+class TestPageSize(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        os.environ["SGLANG_DEBUG_MEMORY_POOL"] = "1"
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST
+        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=["--page-size", 4, "--chunked-prefill-size", 128],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_mmlu(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mmlu",
+            num_examples=64,
+            num_threads=32,
+        )
+
+        metrics = run_eval(args)
+        self.assertGreaterEqual(metrics["score"], 0.65)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/test_retract_decode.py

@@ -1,3 +1,4 @@
+import os
 import unittest
 from types import SimpleNamespace
 
@@ -14,6 +15,8 @@ from sglang.test.test_utils import (
 class TestRetractDecode(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
+        os.environ["SGLANG_TEST_RETRACT"] = "1"
+
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
         cls.base_url = DEFAULT_URL_FOR_TEST
         cls.process = popen_launch_server(
@@ -37,5 +40,20 @@ class TestRetractDecode(unittest.TestCase):
         self.assertGreaterEqual(metrics["score"], 0.65)
 
 
+class TestRetractDecodeChunkCache(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        os.environ["SGLANG_TEST_RETRACT"] = "1"
+
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST
+        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=["--disable-radix-cache", "--chunked-prefill-size", 128],
+        )
+
+
 if __name__ == "__main__":
     unittest.main()