import unittest import requests 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=220, stage="base-b", runner_config="1-gpu-large") register_amd_ci(est_time=355, suite="stage-b-test-1-gpu-small-amd") class TestMultiTokenizer(CustomTestCase, MMLUMixin): """One server covering both worker pools: multi-tokenizer and multi-detokenizer (the flags are orthogonal).""" mmlu_score_threshold = 0.64 mmlu_num_examples = 256 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_tokenizer_ttft(self): # from test_bench_serving.py run_bench_serving 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_tokenizer_ttft\n" f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n" ) self.assertLess(res["median_e2e_latency_ms"], 11000) # relax for mi300x self.assertLess(res["median_ttft_ms"], 130 if is_in_amd_ci() else 86) self.assertLess(res["median_itl_ms"], 10) def test_batch_input_ids_routing(self): # Regression guard for sgl-project/sglang#29878 (introduced by #29214). # # A batch of pre-tokenized `input_ids` (no text / multimodal) is the one # case that takes the batch-tokenization path (_send_batch_request -> # BatchTokenizedGenerateReqInput). In multi-tokenizer mode this batch # must stamp each sub-request's `http_worker_ipc` so the scheduler can # route every reply back to its owning tokenizer worker. If it is missing, # the requests hang forever. # # The existing ttft test only sends *text*, so it never exercises this # path — this case does, and uses a short timeout so a routing hang # fails fast instead of stalling until the server launch timeout. batch_input_ids = [ [1, 2, 3, 4, 5], [10, 11, 12, 13, 14], [20, 21, 22, 23, 24], [30, 31, 32, 33, 34], ] response = requests.post( self.base_url + "/generate", json={ "input_ids": batch_input_ids, "sampling_params": {"max_new_tokens": 8, "temperature": 0}, }, timeout=60, ) self.assertEqual(response.status_code, 200, response.text) results = response.json() # Every batched request must get its reply routed back — not hang. self.assertEqual(len(results), len(batch_input_ids)) for result in results: self.assertIn("text", result) if __name__ == "__main__": unittest.main()