[Refactor] Add -fp4-gemm-backend to replace SGLANG_FLASHINFER_FP4_GEMM_BACKEND (#16534)

Co-authored-by: Vincent Zhong <207368749+vincentzed@users.noreply.github.com>

test/srt/test_nvfp4_gemm.py

@@ -0,0 +1,82 @@
+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.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+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(
+            num_shots=5,
+            data_path=None,
+            num_questions=1319,
+            max_new_tokens=512,
+            parallel=200,
+            host=f"{parsed_url.scheme}://{parsed_url.hostname}",
+            port=parsed_url.port,
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(metrics)
+
+        self.assertGreater(metrics["accuracy"], 0.64)
+
+
+@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
+class TestFP4GemmAuto(FP4GemmBase, unittest.TestCase):
+    backend = "auto"
+
+
+@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
+class TestFP4GemmCutlass(FP4GemmBase, unittest.TestCase):
+    backend = "cutlass"
+
+
+@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
+class TestFP4GemmCudnn(FP4GemmBase, unittest.TestCase):
+    backend = "cudnn"
+
+
+@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
+class TestFP4GemmTrtllm(FP4GemmBase, unittest.TestCase):
+    backend = "trtllm"
+
+
+if __name__ == "__main__":
+    unittest.main()