Add test for Radix cache variants

test/srt/run_suite.py

@@ -63,6 +63,7 @@
         "test_block_int8.py",
         "test_int8_kernel.py",
         "test_reasoning_content.py",
+        "test_cache.py",
     ],
     "nightly": [
         "test_nightly_gsm8k_eval.py",

test/srt/test_cache.py

@@ -0,0 +1,221 @@
+"""
+python3 -m unittest test_cache
+"""
+
+import time
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+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,
+)
+
+
+# No need to test ChunkCache, as it's tested in test_chunked_prefill.py
+class TestDisableRadixCache(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        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",
+                "--enable-cache-report",
+                "--enable-metrics",
+            ],
+        )
+
+    @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)
+
+    def run_decode(self, max_new_tokens):
+        response = requests.post(
+            self.base_url + "/generate",
+            json={
+                "text": "The capital of France is",
+                "sampling_params": {
+                    "temperature": 0,
+                    "max_new_tokens": max_new_tokens,
+                    "ignore_eos": True,
+                },
+            },
+        )
+        return response.json()
+
+    def test_throughput(self):
+        # Warmup
+        res = self.run_decode(16)
+
+        max_tokens = 256
+        tic = time.time()
+        res = self.run_decode(max_tokens)
+        tok = time.time()
+        print(f"{res=}")
+        throughput = max_tokens / (tok - tic)
+        print(f"Throughput: {throughput} tokens/s")
+        self.assertGreaterEqual(throughput, 80)
+
+
+class TestRadixCache(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        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=[
+                "--enable-cache-report",
+                "--enable-metrics",
+            ],
+        )
+
+    @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)
+
+    def run_decode(self, max_new_tokens):
+        response = requests.post(
+            self.base_url + "/generate",
+            json={
+                "text": "The capital of France is",
+                "sampling_params": {
+                    "temperature": 0,
+                    "max_new_tokens": max_new_tokens,
+                    "ignore_eos": True,
+                },
+            },
+        )
+        return response.json()
+
+    def test_evict(self, num_iters=2):
+        # Run a few times to ensure eviction happens
+        print(f"Running test_mmlu {num_iters} times...")
+        start_time = time.time()
+        for i in range(num_iters):
+            print(f"Running iteration {i+1}/{num_iters}")
+            self.test_mmlu()
+        end_time = time.time()
+        print(f"Total time: {end_time - start_time:.2f} seconds")
+
+    def test_throughput(self):
+        # Warmup
+        res = self.run_decode(16)
+
+        max_tokens = 256
+        tic = time.time()
+        res = self.run_decode(max_tokens)
+        tok = time.time()
+        print(f"{res=}")
+        throughput = max_tokens / (tok - tic)
+        print(f"Throughput: {throughput} tokens/s")
+        self.assertGreaterEqual(throughput, 80)
+
+
+class TestHiCache(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        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=[
+                "--enable-hierarchical-cache",
+                "--enable-cache-report",
+                "--enable-metrics",
+            ],
+        )
+
+    @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)
+
+    def run_decode(self, max_new_tokens):
+        response = requests.post(
+            self.base_url + "/generate",
+            json={
+                "text": "The capital of France is",
+                "sampling_params": {
+                    "temperature": 0,
+                    "max_new_tokens": max_new_tokens,
+                    "ignore_eos": True,
+                },
+            },
+        )
+        return response.json()
+
+    def test_evict(self, num_iters=2):
+        print(f"Running test_mmlu {num_iters} times...")
+        start_time = time.time()
+        for i in range(num_iters):
+            print(f"Running iteration {i+1}/{num_iters}")
+            self.test_mmlu()
+        end_time = time.time()
+        print(f"Total time: {end_time - start_time:.2f} seconds")
+
+    def test_throughput(self):
+        # Warmup
+        res = self.run_decode(16)
+
+        max_tokens = 256
+        tic = time.time()
+        res = self.run_decode(max_tokens)
+        tok = time.time()
+        print(f"{res=}")
+        throughput = max_tokens / (tok - tic)
+        print(f"Decode Throughput: {throughput} tokens/s")
+        self.assertGreaterEqual(throughput, 80)
+
+
+if __name__ == "__main__":
+    unittest.main()