Feat/add w4a16 moe support to nemotron (#25655)

test/manual/models/test_nvidia_nemotron_3_nano_archived.py

@@ -7,6 +7,7 @@ Run with `python3 test/manual/models/test_nvidia_nemotron_3_nano_archived.py`.
 
 import unittest
 
+from sglang.srt.utils import is_sm80_supported, is_sm90_supported
 from sglang.test.kits.lm_eval_kit import LMEvalMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
@@ -43,5 +44,27 @@ class TestNvidiaNemotron3Nano30BBF16FlashInfer(LMEvalMixin, DefaultServerBase):
     ] + NEMOTRON_3_NANO_THINKING_ARGS
 
 
+@unittest.skip("Skip, test pass locally but compiling takes too long in CI")
+@unittest.skipIf(
+    not (is_sm80_supported() or is_sm90_supported()),
+    "NVFP4 Marlin fallback test requires CUDA SM8X/SM9X",
+)
+class TestNvidiaNemotron3Nano30BNVFP4Marlin(LMEvalMixin, DefaultServerBase):
+    """Test Nemotron-3-Nano-30B NVFP4 model with the Marlin path."""
+
+    model = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4"
+    model_config_name = "lm_eval_configs/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4.yaml"
+    other_args = [
+        "--tp-size",
+        "1",
+        "--quantization",
+        "modelopt_fp4",
+        "--fp4-gemm-backend",
+        "marlin",
+        "--moe-runner-backend",
+        "marlin",
+    ] + NEMOTRON_3_NANO_THINKING_ARGS
+
+
 if __name__ == "__main__":
     unittest.main()