diff --git a/python/sglang/srt/models/llama4.py b/python/sglang/srt/models/llama4.py
index 2d3a4fc10..22749a163 100644
--- a/python/sglang/srt/models/llama4.py
+++ b/python/sglang/srt/models/llama4.py
@@ -56,11 +56,13 @@ from sglang.srt.utils import (
     fast_topk,
     get_compiler_backend,
     is_cuda,
+    is_npu,
     make_layers,
 )
 from sglang.srt.utils.common import get_current_device_stream_fast
 
 _is_cuda = is_cuda()
+_is_npu = is_npu()
 
 logger = logging.getLogger(__name__)
 
@@ -329,6 +331,8 @@ class Llama4Attention(nn.Module):
         if self.rotary_emb is not None:
             q_view, k_view = qk.split([self.q_size, self.kv_size], dim=-1)
             q_out_unused, k_out_unused = self.rotary_emb(positions, q_view, k_view)
+            if _is_npu:
+                qk = torch.cat([q_out_unused, k_out_unused], dim=-1)
             del q_view, k_view, q_out_unused, k_out_unused
 
         if self.qk_norm is not None:
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index a250cf1fe..c350bb661 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -1480,9 +1480,10 @@ class ServerArgs:
                 "fa3",
                 "aiter",
                 "triton",
+                "ascend",
                 "trtllm_mha",
                 "intel_xpu",
-            }, f"fa3, aiter, triton, trtllm_mha or intel_xpu is required for Llama4 model but got {self.attention_backend}"
+            }, f"fa3, aiter, triton, ascend, trtllm_mha or intel_xpu is required for Llama4 model but got {self.attention_backend}"
             if is_sm100_supported() and self.moe_runner_backend == "auto":
                 if self.quantization in {"fp8", "modelopt_fp8"}:
                     self.moe_runner_backend = "flashinfer_trtllm"
diff --git a/test/registered/ascend/llm_models/test_ascend_llama4_scount_17b_16e.py b/test/registered/ascend/llm_models/test_ascend_llama4_scount_17b_16e.py
new file mode 100644
index 000000000..ee6be144e
--- /dev/null
+++ b/test/registered/ascend/llm_models/test_ascend_llama4_scount_17b_16e.py
@@ -0,0 +1,31 @@
+import unittest
+
+from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
+from sglang.test.ci.ci_register import register_npu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
+
+
+class TestLlama4(GSM8KAscendMixin, CustomTestCase):
+    model = (
+        "/root/.cache/modelscope/hub/models/meta-llama/Llama-4-Scout-17B-16E-Instruct"
+    )
+    accuracy = 0.9
+    other_args = [
+        "--chat-template",
+        "llama-4",
+        "--tp-size",
+        4,
+        "--mem-fraction-static",
+        "0.9",
+        "--context-length",
+        "8192",
+        "--attention-backend",
+        "ascend",
+        "--disable-cuda-graph",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()