diff --git a/.gitignore b/.gitignore
index 306c2d163..57b966665 100644
--- a/.gitignore
+++ b/.gitignore
@@ -270,6 +270,11 @@ sgl-kernel/csrc/**/*_musa/
 # MUSA core dump files
 *.mudmp
 
+# Local-only XPU test: Nemotron-3 Nano 30B-A3B is too large to run in CI
+# (requires tp=4; no 4-GPU XPU runner). Kept on disk for local development,
+# never tracked/pushed.
+test/registered/xpu/test_nvidia_nemotron_3_nano.py
+
 # Others
 # diffusion 3D outputs
 *.glb
diff --git a/python/sglang/srt/layers/attention/mamba/mamba.py b/python/sglang/srt/layers/attention/mamba/mamba.py
index 042cf43dd..0ae2566e5 100644
--- a/python/sglang/srt/layers/attention/mamba/mamba.py
+++ b/python/sglang/srt/layers/attention/mamba/mamba.py
@@ -35,6 +35,7 @@ from sglang.srt.utils import (
     is_cpu,
     is_cuda,
     is_npu,
+    is_xpu,
     set_weight_attrs,
 )
 
@@ -56,6 +57,22 @@ elif is_npu():
     from sgl_kernel_npu.mamba.causal_conv1d import (
         causal_conv1d_update_npu as causal_conv1d_update,
     )
+elif is_xpu():
+    # XPU has no native causal_conv1d kernel yet; use the portable Triton
+    # implementation for both the "native" and the "_triton" entry points so
+    # `causal_conv1d_fn` / `causal_conv1d_fn_triton` are always bound on XPU.
+    from sglang.srt.layers.attention.mamba.causal_conv1d_triton import (
+        causal_conv1d_fn as causal_conv1d_fn,
+    )
+    from sglang.srt.layers.attention.mamba.causal_conv1d_triton import (
+        causal_conv1d_fn as causal_conv1d_fn_triton,
+    )
+    from sglang.srt.layers.attention.mamba.causal_conv1d_triton import (
+        causal_conv1d_update as causal_conv1d_update,
+    )
+    from sglang.srt.layers.attention.mamba.causal_conv1d_triton import (
+        causal_conv1d_update as causal_conv1d_update_triton,
+    )
 
 LoaderFunction = Callable[[torch.Tensor, torch.Tensor], None]
 
diff --git a/python/sglang/srt/layers/quantization/unquant.py b/python/sglang/srt/layers/quantization/unquant.py
index 9aab8f72c..b21972065 100644
--- a/python/sglang/srt/layers/quantization/unquant.py
+++ b/python/sglang/srt/layers/quantization/unquant.py
@@ -635,6 +635,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
         assert moe_runner_config.activation in [
             "silu",
             "gelu",
+            "relu2",  # Nemotron-H (NemotronHForCausalLM) uses squared-ReLU.
         ], f"activation = {moe_runner_config.activation} is not supported."
 
         backend = self.runner.runner_backend