Support penalty in overlap mode; return logprob with chunked prefill; improve benchmark scripts (#3988)

Co-authored-by: SangBin Cho <rkooo567@gmail.com>
Co-authored-by: dhou-xai <dhou@x.ai>
Co-authored-by: Hanming Lu <hanming_lu@berkeley.edu>

test/srt/test_penalty.py

@@ -0,0 +1,91 @@
+import json
+import random
+import unittest
+from concurrent.futures import ThreadPoolExecutor
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.test_utils import (
+    DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
+
+
+class TestPenalty(unittest.TestCase):
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_SMALL_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,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def run_decode(self, sampling_params):
+        return_logprob = True
+        top_logprobs_num = 5
+        return_text = True
+        n = 1
+
+        response = requests.post(
+            self.base_url + "/generate",
+            json={
+                # prompt that is supposed to generate < 32 tokens
+                "text": "<|start_header_id|>user<|end_header_id|>\n\nWhat is the answer for 1 + 1 = ?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",
+                "sampling_params": {
+                    "max_new_tokens": 32,
+                    "n": n,
+                    **sampling_params,
+                },
+                "return_logprob": return_logprob,
+                "top_logprobs_num": top_logprobs_num,
+                "return_text_in_logprobs": return_text,
+                "logprob_start_len": 0,
+            },
+        )
+        self.assertEqual(response.status_code, 200)
+        print(json.dumps(response.json()))
+        print("=" * 100)
+
+    def test_default_values(self):
+        self.run_decode({})
+
+    def test_frequency_penalty(self):
+        self.run_decode({"frequency_penalty": 2})
+
+    def test_min_new_tokens(self):
+        self.run_decode({"min_new_tokens": 16})
+
+    def test_presence_penalty(self):
+        self.run_decode({"presence_penalty": 2})
+
+    def test_mixed(self):
+        args = [
+            {},
+            {},
+            {},
+            {"frequency_penalty": 2},
+            {"min_new_tokens": 16},
+            {"presence_penalty": 1},
+            {"frequency_penalty": 0.2},
+            {"min_new_tokens": 8},
+            {"presence_penalty": 0.4},
+            {"presence_penalty": 0.4, "frequency_penalty": 2},
+            {"min_new_tokens": 12, "frequency_penalty": 2},
+        ]
+        random.shuffle(args * 5)
+        with ThreadPoolExecutor(8) as executor:
+            list(executor.map(self.run_decode, args))
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=3)