Add multi-detokenizer support (#24944)

Co-authored-by: ybyang <10629930+whybeyoung@users.noreply.github.com> Co-authored-by: Shangming Cai <csmthu@gmail.com>
2026-05-15 17:26:26 -07:00
parent afc7c9f7f3
commit 5ba69f50fb
5 changed files with 308 additions and 32 deletions
@@ -0,0 +1,80 @@
+import unittest
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.kits.eval_accuracy_kit import MMLUMixin
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    auto_config_device,
+    get_benchmark_args,
+    is_in_amd_ci,
+    is_in_ci,
+    popen_launch_server,
+    run_benchmark,
+    write_github_step_summary,
+)
+
+register_cuda_ci(est_time=211, suite="stage-b-test-1-gpu-large")
+register_amd_ci(est_time=345, suite="stage-b-test-1-gpu-small-amd")
+
+
+class TestMultiDetokenizer(CustomTestCase, MMLUMixin):
+    mmlu_score_threshold = 0.65
+    mmlu_num_examples = 64
+    mmlu_num_threads = 32
+
+    @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=[
+                "--tokenizer-worker-num",
+                8,
+                "--detokenizer-worker-num",
+                4,
+                "--mem-fraction-static",
+                0.7,
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_multi_detokenizer_ttft(self):
+        args = get_benchmark_args(
+            base_url=self.base_url,
+            dataset_name="random",
+            dataset_path="",
+            tokenizer=None,
+            num_prompts=100,
+            random_input_len=4096,
+            random_output_len=2048,
+            sharegpt_context_len=None,
+            request_rate=1,
+            disable_stream=False,
+            disable_ignore_eos=False,
+            seed=0,
+            device=auto_config_device(),
+            lora_name=None,
+        )
+        res = run_benchmark(args)
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_multi_detokenizer_ttft\n"
+                f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
+            )
+            self.assertLess(res["median_e2e_latency_ms"], 11000)
+            self.assertLess(res["median_ttft_ms"], 130 if is_in_amd_ci() else 86)
+            self.assertLess(res["median_itl_ms"], 10)
+
+
+if __name__ == "__main__":
+    unittest.main()