"""Engine API with --tokenizer-worker-num > 1 (issue #15157). The parent Engine keeps generate() working through an in-process TokenizerWorker, and other processes join the same engine with Engine.attach_tokenizer_worker(). """ import multiprocessing as mp import os import unittest import sglang as sgl from sglang.test.ci.ci_register import register_cuda_ci from sglang.test.test_utils import DEFAULT_SMALL_MODEL_NAME_FOR_TEST, CustomTestCase register_cuda_ci(est_time=180, stage="base-b", runner_config="1-gpu-large") PROMPTS = ["The capital of France is", "1 + 1 ="] SAMPLING = {"temperature": 0, "max_new_tokens": 8} def _attached_generate(parent_pid, prompts, queue): engine = sgl.Engine.attach_tokenizer_worker(parent_pid) try: outs = engine.generate(prompts, SAMPLING) queue.put([(o["text"], o["meta_info"]["completion_tokens"]) for o in outs]) finally: engine.shutdown() class TestEngineMultiTokenizer(CustomTestCase): @classmethod def setUpClass(cls): cls.engine = sgl.Engine( model_path=DEFAULT_SMALL_MODEL_NAME_FOR_TEST, random_seed=42, tokenizer_worker_num=2, ) @classmethod def tearDownClass(cls): cls.engine.shutdown() def test_1_parent_generate_uses_worker(self): self.assertIsNotNone(self.engine.tokenizer_router) outs = self.engine.generate(PROMPTS, SAMPLING) self.assertEqual(len(outs), len(PROMPTS)) self.assertTrue(all(o["meta_info"]["completion_tokens"] > 0 for o in outs)) def test_2_attached_process_matches_parent(self): expected = [ (o["text"], o["meta_info"]["completion_tokens"]) for o in self.engine.generate(PROMPTS, SAMPLING) ] ctx = mp.get_context("spawn") queue = ctx.Queue() proc = ctx.Process( target=_attached_generate, args=(os.getpid(), PROMPTS, queue) ) proc.start() got = queue.get(timeout=600) proc.join(timeout=60) self.assertEqual(proc.exitcode, 0) self.assertEqual(got, expected) if __name__ == "__main__": unittest.main()