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

test/lm_eval_configs/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4.yaml

@@ -0,0 +1,13 @@
+model_name: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.847
+  - name: "exact_match,flexible-extract"
+    value: 0.444
+limit: 1319
+num_concurrent: 128
+num_fewshot: 5
+apply_chat_template: false
+fewshot_as_multiturn: true

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()

test/registered/unit/model_loader/test_modelopt_loader.py

@@ -15,6 +15,7 @@ from sglang.srt.configs.load_config import LoadConfig
 from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.layers.modelopt_utils import QUANT_CFG_CHOICES
 from sglang.srt.layers.quantization.modelopt_quant import (
+    ModelOptFp4Config,
     ModelOptMixedPrecisionConfig,
 )
 from sglang.srt.model_loader.loader import ModelOptModelLoader
@@ -646,7 +647,10 @@ class TestModelOptMixedPrecisionConfig(CustomTestCase):
         )
 
     def test_mixed_precision_uses_nvfp4_min_capability(self):
-        self.assertEqual(ModelOptMixedPrecisionConfig.get_min_capability(), 100)
+        self.assertEqual(
+            ModelOptMixedPrecisionConfig.get_min_capability(),
+            ModelOptFp4Config.get_min_capability(),
+        )
 
     def test_mixed_precision_quant_layer_resolution_after_mapping(self):
         quant_config = ModelOptMixedPrecisionConfig.from_config(