"""Archived test classes split out of test/registered/quant/test_nvfp4_gemm.py. Originally registered with `register_cuda_ci(...)`. Moved here as part of the per-commit pruning effort to keep the code reachable manually. Run with `python3 test/manual/quant/test_nvfp4_gemm_archived.py`. """ import unittest from types import SimpleNamespace from urllib.parse import urlparse from sglang.srt.utils import get_device_sm, kill_process_tree from sglang.test.run_eval import run_eval from sglang.test.test_utils import ( DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_URL_FOR_TEST, popen_launch_server, try_cached_model, ) MODEL_PATH = "nvidia/Llama-3.1-8B-Instruct-NVFP4" class FP4GemmBase: backend = None @classmethod def setUpClass(cls): if cls.backend is None: raise NotImplementedError("Subclass must set 'backend' attribute") cls.model = try_cached_model(MODEL_PATH) cls.base_url = DEFAULT_URL_FOR_TEST other_args = [ "--trust-remote-code", "--quantization", "modelopt_fp4", "--fp4-gemm-backend", cls.backend, ] cls.process = popen_launch_server( cls.model, cls.base_url, timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, other_args=other_args, ) @classmethod def tearDownClass(cls): kill_process_tree(cls.process.pid) def test_gsm8k(self): parsed_url = urlparse(self.base_url) args = SimpleNamespace( base_url=self.base_url, model=self.model, eval_name="gsm8k", api="completion", max_tokens=512, num_examples=1319, num_threads=200, ) metrics = run_eval(args) print(metrics) self.assertGreater(metrics["score"], 0.64) @unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher") class TestFP4GemmAuto(FP4GemmBase, unittest.TestCase): backend = "auto" if __name__ == "__main__": unittest.main()