From a2ea5941d562faa5a9a20083ffc81dd0d9dd2d6a Mon Sep 17 00:00:00 2001
From: Zheng Duan <200704041+zhengd-nv@users.noreply.github.com>
Date: Sat, 28 Feb 2026 10:28:47 +0800
Subject: [PATCH] [feat] Support nvfp4 quantized model of Qwen3-Next (#17627)

---
 python/sglang/srt/models/qwen3_next.py        | 13 +++-
 .../models/test_qwen3_next_models_fp4.py      | 71 +++++++++++++++++++
 2 files changed, 83 insertions(+), 1 deletion(-)
 create mode 100644 test/registered/models/test_qwen3_next_models_fp4.py

diff --git a/python/sglang/srt/models/qwen3_next.py b/python/sglang/srt/models/qwen3_next.py
index 6000e92db..e742dd543 100644
--- a/python/sglang/srt/models/qwen3_next.py
+++ b/python/sglang/srt/models/qwen3_next.py
@@ -625,13 +625,19 @@ class Qwen3HybridAttentionDecoderLayer(nn.Module):
             dtype=torch.get_default_dtype(),  # see impl of get_rope
         )
 
+        # qkv_proj is not quantized for fp4
         self.qkv_proj = QKVParallelLinear(
             config.hidden_size,
             self.head_dim,
             self.total_num_heads * (1 + self.attn_output_gate),
             self.total_num_kv_heads,
             bias=False,
-            quant_config=quant_config,
+            quant_config=(
+                quant_config
+                if quant_config is not None
+                and quant_config.get_name() != "modelopt_fp4"
+                else None
+            ),
             tp_rank=self.attn_tp_rank,
             tp_size=self.attn_tp_size,
             prefix=add_prefix("qkv_proj", prefix),
@@ -1123,6 +1129,11 @@ class Qwen3NextForCausalLM(nn.Module):
                     # if is_pp_missing_parameter(name, self):
                     #     continue
 
+                    if name.endswith("_scale") and name not in params_dict:
+                        assert (
+                            abs(loaded_weight.item() - 1.0) < 1e-6
+                        ), f"Expected 1.0, got {loaded_weight.item()} in skipped {name}"
+                        continue
                     param = params_dict[name]
                     weight_loader = getattr(
                         param, "weight_loader", default_weight_loader
diff --git a/test/registered/models/test_qwen3_next_models_fp4.py b/test/registered/models/test_qwen3_next_models_fp4.py
new file mode 100644
index 000000000..14be5ffa0
--- /dev/null
+++ b/test/registered/models/test_qwen3_next_models_fp4.py
@@ -0,0 +1,71 @@
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import get_device_sm, kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.few_shot_gsm8k import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_cuda_ci(est_time=500, suite="nightly-4-gpu-b200", nightly=True)
+
+QWEN3_NEXT_MODEL_FP4 = "nvidia/Qwen3-Next-80B-A3B-Instruct-NVFP4"
+
+ACC_THRESHOLDS = {
+    QWEN3_NEXT_MODEL_FP4: {"kl_div": 0.0025, "gsm8k": 0.93},
+}
+
+
+@unittest.skipIf(
+    get_device_sm() < 100, "Test requires CUDA SM 100 or higher (Blackwell)"
+)
+class TestQwen3NextFp4(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = QWEN3_NEXT_MODEL_FP4
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--tp-size",
+                "4",
+                "--chunked-prefill-size",
+                "2048",
+                "--quantization",
+                "modelopt_fp4",
+                "--mamba-scheduler-strategy",
+                "extra_buffer",
+                "--mamba-track-interval",
+                "128",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval(args)
+        print(f"{metrics=}")
+        self.assertGreaterEqual(
+            metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()