Improve end-to-end throughput test and its coverage (#1039)

.github/workflows/e2e-test.yml

@@ -37,23 +37,16 @@ jobs:
 
     - name: Benchmark Serving Throughput
       run: |
-        python3 -m sglang.launch_server --model meta-llama/Meta-Llama-3.1-8B-Instruct --port 8413 --disable-radix-cache &
-        SERVER_PID=$!
-
-        echo "Waiting for server to start..."
-        for i in {1..120}; do
-          if curl -s http://127.0.0.1:8413/health; then
-            echo "Server is up!"
-            break
-          fi
-          if [ $i -eq 120 ]; then
-            echo "Server failed to start within 120 seconds"
-            exit 1
-          fi
-          sleep 1
-        done
-
-        cd $HOME && python3 -m sglang.bench_serving --backend sglang --port 8413 --dataset-name random --num-prompts 500 --random-input 4096 --random-output 2048
-
-        echo "Stopping server..."
-        kill -9 $SERVER_PID
+        cd test/srt
+        python3 -m unittest test_serving_throughput.TestServingThroughput.test_default
+
+    - name: Benchmark Serving Throughput (w/o RadixAttention)
+      run: |
+        cd test/srt
+        python3 -m unittest test_serving_throughput.TestServingThroughput.test_default_without_radix_cache
+
+    - name: Benchmark Serving Throughput (w/o FlashInfer)
+      run: |
+        cd test/srt
+        python3 -m unittest test_serving_throughput.TestServingThroughput.test_default_without_flashinfer
+

python/sglang/bench_serving.py

@@ -39,6 +39,8 @@ from transformers import (
 
 AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
 
+global args
+
 
 @dataclass
 class RequestFuncInput:
@@ -749,7 +751,11 @@ def check_chat_template(model_path):
         return False
 
 
-def fire(args: argparse.Namespace):
+def run_benchmark(args_: argparse.Namespace):
+    global args
+    args = args_
+
+    set_ulimit()
     random.seed(args.seed)
     np.random.seed(args.seed)
 
@@ -853,7 +859,7 @@ def fire(args: argparse.Namespace):
                 )
             )
     else:
-        asyncio.run(
+        return asyncio.run(
             benchmark(
                 backend=backend,
                 api_url=api_url,
@@ -962,11 +968,6 @@ if __name__ == "__main__":
         "Otherwise, we use Poisson process to synthesize the request arrival times. Default is 128.0.",
     )
     parser.add_argument("--seed", type=int, default=0, help="Default is 0.")
-    parser.add_argument(
-        "--disable-tqdm",
-        action="store_true",
-        help="Specify to disable tqdm progress bar.",
-    )
     parser.add_argument(
         "--multi",
         action="store_true",
@@ -979,6 +980,11 @@ if __name__ == "__main__":
         help="Range of request rates in the format start,stop,step. Default is 2,34,2. It also supports a list of request rates, requiring the parameters to not equal three.",
     )
     parser.add_argument("--output-file", type=str, help="Output JSONL file name.")
+    parser.add_argument(
+        "--disable-tqdm",
+        action="store_true",
+        help="Specify to disable tqdm progress bar.",
+    )
     parser.add_argument(
         "--disable-stream",
         action="store_true",
@@ -996,8 +1002,5 @@ if __name__ == "__main__":
         help="Append given JSON object to the request payload. You can use this to specify"
         "additional generate params like sampling params.",
     )
-
-    set_ulimit()
-
     args = parser.parse_args()
-    fire(args)
+    run_benchmark(args)

python/sglang/test/test_utils.py

@@ -21,6 +21,7 @@ from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
 from sglang.utils import get_exception_traceback
 
 DEFAULT_MODEL_NAME_FOR_TEST = "meta-llama/Meta-Llama-3.1-8B-Instruct"
+DEFAULT_URL_FOR_TEST = "http://127.0.0.1:8157"
 
 
 def call_generate_lightllm(prompt, temperature, max_tokens, stop=None, url=None):

test/srt/test_chunked_prefill.py

@@ -3,7 +3,11 @@ from types import SimpleNamespace
 
 from sglang.srt.utils import kill_child_process
 from sglang.test.run_eval import run_eval
-from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, popen_launch_server
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
 
 
 class TestAccuracy(unittest.TestCase):
@@ -11,7 +15,7 @@ class TestAccuracy(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
-        cls.base_url = "http://127.0.0.1:8157"
+        cls.base_url = DEFAULT_URL_FOR_TEST
         cls.process = popen_launch_server(
             cls.model,
             cls.base_url,

test/srt/test_embedding_openai_server.py

@@ -4,7 +4,7 @@ import openai
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
 from sglang.srt.utils import kill_child_process
-from sglang.test.test_utils import popen_launch_server
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, popen_launch_server
 
 
 class TestOpenAIServer(unittest.TestCase):
@@ -12,7 +12,7 @@ class TestOpenAIServer(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = "intfloat/e5-mistral-7b-instruct"
-        cls.base_url = "http://127.0.0.1:8157"
+        cls.base_url = DEFAULT_URL_FOR_TEST
         cls.api_key = "sk-123456"
         cls.process = popen_launch_server(
             cls.model, cls.base_url, timeout=300, api_key=cls.api_key

test/srt/test_eval_accuracy.py

@@ -3,7 +3,11 @@ from types import SimpleNamespace
 
 from sglang.srt.utils import kill_child_process
 from sglang.test.run_eval import run_eval
-from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, popen_launch_server
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
 
 
 class TestAccuracy(unittest.TestCase):
@@ -11,7 +15,7 @@ class TestAccuracy(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
-        cls.base_url = "http://127.0.0.1:8157"
+        cls.base_url = DEFAULT_URL_FOR_TEST
         cls.process = popen_launch_server(cls.model, cls.base_url, timeout=300)
 
     @classmethod

test/srt/test_large_max_new_tokens.py

@@ -8,7 +8,11 @@ import openai
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
 from sglang.srt.utils import kill_child_process
-from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, popen_launch_server
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
 
 
 class TestOpenAIServer(unittest.TestCase):
@@ -16,7 +20,7 @@ class TestOpenAIServer(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
-        cls.base_url = "http://127.0.0.1:8157"
+        cls.base_url = DEFAULT_URL_FOR_TEST
         cls.api_key = "sk-123456"
         cls.process = popen_launch_server(
             cls.model,

test/srt/test_openai_server.py

@@ -6,7 +6,11 @@ import openai
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
 from sglang.srt.utils import kill_child_process
-from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, popen_launch_server
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
 
 
 class TestOpenAIServer(unittest.TestCase):
@@ -14,7 +18,7 @@ class TestOpenAIServer(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
-        cls.base_url = "http://127.0.0.1:8157"
+        cls.base_url = DEFAULT_URL_FOR_TEST
         cls.api_key = "sk-123456"
         cls.process = popen_launch_server(
             cls.model, cls.base_url, timeout=300, api_key=cls.api_key

test/srt/test_serving_throughput.py

+import unittest
+from types import SimpleNamespace
+
+from sglang.bench_serving import run_benchmark
+from sglang.srt.utils import kill_child_process
+from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, popen_launch_server
+
+
+class TestServingThroughput(unittest.TestCase):
+
+    def run_test(self, disable_radix_cache, disable_flashinfer, chunked_prefill_size):
+        # Launch the server
+        other_args = []
+        if disable_radix_cache:
+            other_args.append("--disable-radix-cache")
+        if disable_flashinfer:
+            other_args.append("--disable-flashinfer")
+        other_args.extend(["--chunked-prefill-size", str(chunked_prefill_size)])
+
+        model = DEFAULT_MODEL_NAME_FOR_TEST
+        base_url = "http://127.0.0.1:9157"
+        process = popen_launch_server(
+            model, base_url, timeout=300, other_args=other_args
+        )
+
+        # Run benchmark
+        num_prompts = 400
+        args = SimpleNamespace(
+            backend="sglang",
+            base_url=base_url,
+            host=None,
+            port=None,
+            dataset_name="random",
+            dataset_path="",
+            model=None,
+            tokenizer=None,
+            num_prompts=num_prompts,
+            sharegpt_output_len=None,
+            random_input_len=4096,
+            random_output_len=2048,
+            random_range_ratio=0.0,
+            request_rate=float("inf"),
+            multi=None,
+            seed=0,
+            output_file=None,
+            disable_tqdm=False,
+            disable_stream=False,
+            disable_ignore_eos=False,
+            extra_request_body=None,
+        )
+
+        try:
+            res = run_benchmark(args)
+        finally:
+            kill_child_process(process.pid)
+
+        assert res["completed"] == num_prompts
+
+    def test_default(self):
+        self.run_test(
+            disable_radix_cache=False,
+            disable_flashinfer=False,
+            chunked_prefill_size=-1,
+        )
+
+    def test_default_without_radix_cache(self):
+        self.run_test(
+            disable_radix_cache=True,
+            disable_flashinfer=False,
+            chunked_prefill_size=-1,
+        )
+
+    def test_default_without_flashinfer(self):
+        self.run_test(
+            disable_radix_cache=False,
+            disable_flashinfer=True,
+            chunked_prefill_size=-1,
+        )
+
+    def test_all_cases(self):
+        for disable_radix_cache in [False, True]:
+            for disable_flashinfer in [False, True]:
+                for chunked_prefill_size in [-1, 2048]:
+                    self.run_test(
+                        disable_radix_cache=False,
+                        disable_flashinfer=False,
+                        chunked_prefill_size=-1,
+                    )
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/test_skip_tokenizer_init.py

@@ -4,7 +4,11 @@ import unittest
 import requests
 
 from sglang.srt.utils import kill_child_process
-from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, popen_launch_server
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
 
 
 class TestSkipTokenizerInit(unittest.TestCase):
@@ -12,7 +16,7 @@ class TestSkipTokenizerInit(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
-        cls.base_url = "http://127.0.0.1:8157"
+        cls.base_url = DEFAULT_URL_FOR_TEST
         cls.process = popen_launch_server(
             cls.model, cls.base_url, timeout=300, other_args=["--skip-tokenizer-init"]
         )

test/srt/test_srt_endpoint.py

@@ -4,7 +4,11 @@ import unittest
 import requests
 
 from sglang.srt.utils import kill_child_process
-from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, popen_launch_server
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
 
 
 class TestSRTEndpoint(unittest.TestCase):
@@ -12,7 +16,7 @@ class TestSRTEndpoint(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
-        cls.base_url = "http://127.0.0.1:8157"
+        cls.base_url = DEFAULT_URL_FOR_TEST
         cls.process = popen_launch_server(cls.model, cls.base_url, timeout=300)
 
     @classmethod

test/srt/test_throughput.py

+import json
+import unittest
+
+import requests
+
+from sglang.srt.utils import kill_child_process
+from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, popen_launch_server
+
+
+class TestSRTEndpoint(unittest.TestCase):
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST
+        cls.base_url = "http://127.0.0.1:8157"
+        cls.process = popen_launch_server(cls.model, cls.base_url, timeout=300)
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_child_process(cls.process.pid)
+
+    def run_decode(
+        self, return_logprob=False, top_logprobs_num=0, return_text=False, n=1
+    ):
+        response = requests.post(
+            self.base_url + "/generate",
+            json={
+                "text": "The capital of France is",
+                "sampling_params": {
+                    "temperature": 0 if n == 1 else 0.5,
+                    "max_new_tokens": 32,
+                    "n": n,
+                },
+                "stream": False,
+                "return_logprob": return_logprob,
+                "top_logprobs_num": top_logprobs_num,
+                "return_text_in_logprobs": return_text,
+                "logprob_start_len": 0,
+            },
+        )
+        print(json.dumps(response.json()))
+        print("=" * 100)
+
+    def test_simple_decode(self):
+        self.run_decode()
+
+    def test_parallel_sample(self):
+        self.run_decode(n=3)
+
+    def test_logprob(self):
+        for top_logprobs_num in [0, 3]:
+            for return_text in [True, False]:
+                self.run_decode(
+                    return_logprob=True,
+                    top_logprobs_num=top_logprobs_num,
+                    return_text=return_text,
+                )
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/test_torch_compile.py

@@ -3,7 +3,11 @@ from types import SimpleNamespace
 
 from sglang.srt.utils import kill_child_process
 from sglang.test.run_eval import run_eval
-from sglang.test.test_utils import DEFAULT_MODEL_NAME_FOR_TEST, popen_launch_server
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
 
 
 class TestAccuracy(unittest.TestCase):
@@ -11,7 +15,7 @@ class TestAccuracy(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
-        cls.base_url = "http://127.0.0.1:8157"
+        cls.base_url = DEFAULT_URL_FOR_TEST
         cls.process = popen_launch_server(
             cls.model, cls.base_url, timeout=300, other_args=["--enable-torch-compile"]
         )

test/srt/test_vision_openai_server.py

@@ -5,7 +5,7 @@ import openai
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
 from sglang.srt.utils import kill_child_process
-from sglang.test.test_utils import popen_launch_server
+from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, popen_launch_server
 
 
 class TestOpenAIVisionServer(unittest.TestCase):
@@ -13,7 +13,7 @@ class TestOpenAIVisionServer(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = "liuhaotian/llava-v1.6-vicuna-7b"
-        cls.base_url = "http://127.0.0.1:8157"
+        cls.base_url = DEFAULT_URL_FOR_TEST
         cls.api_key = "sk-123456"
         cls.process = popen_launch_server(
             cls.model,