Fix unit tests for the frontend language part (#872)

.github/workflows/lint.yml

@@ -1,4 +1,4 @@
-name: lint
+name: Lint
 
 on: [push, pull_request]
 

python/sglang/lang/ir.py

@@ -99,7 +99,6 @@ class SglSamplingParams:
             "stop": self.stop or None,
             "temperature": self.temperature,
             "top_p": self.top_p,
-            "top_k": self.top_k,
             "frequency_penalty": self.frequency_penalty,
             "presence_penalty": self.presence_penalty,
         }

python/sglang/srt/server.py

@@ -479,6 +479,9 @@ class Runtime:
             parent.wait(timeout=5)
             self.pid = None
 
+    def cache_prefix(self, prefix: str):
+        self.endpoint.cache_prefix(prefix)
+
     def get_tokenizer(self):
         return get_tokenizer(
             self.server_args.tokenizer_path,

python/sglang/test/test_programs.py

@@ -113,15 +113,14 @@ def test_decode_json_regex():
             s += '  "population": ' + sgl.gen(regex=REGEX_INT + ",") + "\n"
             s += '  "area": ' + sgl.gen(regex=REGEX_INT + ",") + "\n"
             s += '  "latitude": ' + sgl.gen(regex=REGEX_FLOAT + ",") + "\n"
-            s += '  "country": ' + sgl.gen(regex=REGEX_STRING + ",") + "\n"
+            s += '  "country": ' + sgl.gen(regex=REGEX_STRING) + "\n"
-            s += '  "timezone": ' + sgl.gen(regex=REGEX_STRING) + "\n"
             s += "}"
 
-    ret = decode_json.run()
+    ret = decode_json.run(temperature=0.0)
     try:
         js_obj = json.loads(ret["json_output"])
     except json.decoder.JSONDecodeError:
-        print(ret["json_output"])
+        print("JSONDecodeError", ret["json_output"])
         raise
     assert isinstance(js_obj["name"], str)
     assert isinstance(js_obj["population"], int)
@@ -141,8 +140,12 @@ def test_decode_json():
             s += '  "timezone": ' + sgl.gen(dtype=str) + "\n"
             s += "}"
 
-    ret = decode_json.run()
+    ret = decode_json.run(max_new_tokens=64)
+    try:
         js_obj = json.loads(ret["json_output"])
+    except json.decoder.JSONDecodeError:
+        print("JSONDecodeError", ret["json_output"])
+        raise
     assert isinstance(js_obj["name"], str)
     assert isinstance(js_obj["population"], int)
 

test/lang/run_suite.py

@@ -7,6 +7,10 @@ import unittest
 
 from sglang.utils import run_with_timeout
 
+suites = {
+    "minimal": ["test_openai_backend.py", "test_srt_backend.py"],
+}
+
 
 def run_unittest_files(files, args):
     for filename in files:
@@ -45,9 +49,19 @@ if __name__ == "__main__":
         default=1000,
         help="The time limit for running one file in seconds.",
     )
+    arg_parser.add_argument(
+        "--suite",
+        type=str,
+        default=list(suites.keys())[0],
+        choices=list(suites.keys()) + ["all"],
+        help="The suite to run",
+    )
     args = arg_parser.parse_args()
 
+    if args.suite == "all":
         files = glob.glob("**/test_*.py", recursive=True)
+    else:
+        files = suites[args.suite]
 
     tic = time.time()
     success = run_unittest_files(files, args)

test/lang/test_anthropic_backend.py

@@ -7,12 +7,9 @@ from sglang.test.test_programs import test_mt_bench, test_stream
 
 class TestAnthropicBackend(unittest.TestCase):
     backend = None
-    chat_backend = None
 
-    def setUp(self):
+    @classmethod
-        cls = type(self)
+    def setUpClass(cls):
-
-        if cls.backend is None:
         cls.backend = Anthropic("claude-3-haiku-20240307")
         set_default_backend(cls.backend)
 
@@ -30,5 +27,5 @@ if __name__ == "__main__":
 
     # global_config.verbosity = 2
     # t = TestAnthropicBackend()
-    # t.setUp()
+    # t.setUpClass()
     # t.test_mt_bench()

test/lang/test_bind_cache.py

@@ -1,23 +1,19 @@
-"""
-Usage:
-python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000
-python3 test_bind_cache.py
-"""
-
 import unittest
 
 import sglang as sgl
-from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
 
 
 class TestBind(unittest.TestCase):
     backend = None
 
-    def setUp(self):
+    @classmethod
-        cls = type(self)
+    def setUpClass(cls):
+        cls.backend = sgl.Runtime(model_path="meta-llama/Meta-Llama-3-8B-Instruct")
+        sgl.set_default_backend(cls.backend)
 
-        if cls.backend is None:
+    @classmethod
-            cls.backend = RuntimeEndpoint(base_url="http://localhost:30000")
+    def tearDownClass(cls):
+        cls.backend.shutdown()
 
     def test_bind(self):
         @sgl.function
@@ -54,5 +50,5 @@ if __name__ == "__main__":
     unittest.main(warnings="ignore")
 
     # t = TestBind()
-    # t.setUp()
+    # t.setUpClass()
     # t.test_cache()

test/lang/test_litellm_backend.py

@@ -6,15 +6,12 @@ from sglang.test.test_programs import test_mt_bench, test_stream
 
 
 class TestAnthropicBackend(unittest.TestCase):
-    backend = None
     chat_backend = None
 
-    def setUp(self):
+    @classmethod
-        cls = type(self)
+    def setUpClass(cls):
-
+        cls.chat_backend = LiteLLM("gpt-3.5-turbo")
-        if cls.backend is None:
+        set_default_backend(cls.chat_backend)
-            cls.backend = LiteLLM("gpt-3.5-turbo")
-            set_default_backend(cls.backend)
 
     def test_mt_bench(self):
         test_mt_bench()

test/lang/test_openai_backend.py

@@ -20,20 +20,18 @@ from sglang.test.test_programs import (
 
 
 class TestOpenAIBackend(unittest.TestCase):
-    backend = None
+    instruct_backend = None
     chat_backend = None
     chat_vision_backend = None
 
-    def setUp(self):
+    @classmethod
-        cls = type(self)
+    def setUpClass(cls):
-
+        cls.instruct_backend = OpenAI("gpt-3.5-turbo-instruct")
-        if cls.backend is None:
-            cls.backend = OpenAI("gpt-3.5-turbo-instruct")
         cls.chat_backend = OpenAI("gpt-3.5-turbo")
         cls.chat_vision_backend = OpenAI("gpt-4-turbo")
 
     def test_few_shot_qa(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_few_shot_qa()
 
     def test_mt_bench(self):
@@ -41,35 +39,35 @@ class TestOpenAIBackend(unittest.TestCase):
         test_mt_bench()
 
     def test_select(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_select(check_answer=True)
 
     def test_decode_int(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_decode_int()
 
     def test_decode_json(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_decode_json()
 
     def test_expert_answer(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_expert_answer()
 
     def test_tool_use(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_tool_use()
 
     def test_react(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_react()
 
     def test_parallel_decoding(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_parallel_decoding()
 
     def test_parallel_encoding(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_parallel_encoding()
 
     def test_image_qa(self):
@@ -77,11 +75,11 @@ class TestOpenAIBackend(unittest.TestCase):
         test_image_qa()
 
     def test_stream(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_stream()
 
     def test_completion_speculative(self):
-        set_default_backend(self.backend)
+        set_default_backend(self.instruct_backend)
         test_completion_speculative()
 
     def test_chat_completion_speculative(self):
@@ -96,5 +94,5 @@ if __name__ == "__main__":
 
     # global_config.verbosity = 2
     # t = TestOpenAIBackend()
-    # t.setUp()
+    # t.setUpClass()
-    # t.test_chat_completion_speculative()
+    # t.test_stream()

test/lang/test_srt_backend.py

@@ -1,9 +1,3 @@
-"""
-Usage:
-python3 -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct --port 30000
-python3 test_srt_backend.py
-"""
-
 import json
 import unittest
 
@@ -15,8 +9,6 @@ from sglang.test.test_programs import (
     test_few_shot_qa,
     test_mt_bench,
     test_parallel_decoding,
-    test_parallel_encoding,
-    test_react,
     test_regex,
     test_select,
     test_stream,
@@ -27,13 +19,15 @@ from sglang.test.test_programs import (
 class TestSRTBackend(unittest.TestCase):
     backend = None
 
-    def setUp(self):
+    @classmethod
-        cls = type(self)
+    def setUpClass(cls):
-
+        cls.backend = sgl.Runtime(model_path="meta-llama/Meta-Llama-3-8B-Instruct")
-        if cls.backend is None:
-            cls.backend = sgl.RuntimeEndpoint(base_url="http://localhost:30000")
         sgl.set_default_backend(cls.backend)
 
+    @classmethod
+    def tearDownClass(cls):
+        cls.backend.shutdown()
+
     def test_few_shot_qa(self):
         test_few_shot_qa()
 
@@ -64,9 +58,6 @@ class TestSRTBackend(unittest.TestCase):
     def test_regex(self):
         test_regex()
 
-    # def test_parallel_encoding(self):
-    #     test_parallel_encoding(check_answer=False)
-
 
 if __name__ == "__main__":
     unittest.main(warnings="ignore")
@@ -75,5 +66,6 @@ if __name__ == "__main__":
 
     # global_config.verbosity = 2
     # t = TestSRTBackend()
-    # t.setUp()
+    # t.setUpClass()
-    # t.test_regex()
+    # t.test_few_shot_qa()
+    # t.tearDownClass()

test/lang/test_vertexai_backend.py

@@ -17,10 +17,8 @@ class TestVertexAIBackend(unittest.TestCase):
     chat_backend = None
     chat_vision_backend = None
 
-    def setUp(self):
+    @classmethod
-        cls = type(self)
+    def setUpClass(cls):
-
-        if cls.backend is None:
         cls.backend = VertexAI("gemini-pro")
         cls.chat_backend = VertexAI("gemini-pro")
         cls.chat_vision_backend = VertexAI("gemini-pro-vision")
@@ -61,5 +59,5 @@ if __name__ == "__main__":
 
     # global_config.verbosity = 2
     # t = TestVertexAIBackend()
-    # t.setUp()
+    # t.setUpClass()
     # t.test_stream()