[XPU] Enable NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 on Intel XPU backend (#24390)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Yao Matrix <matrix.yao@intel.com>

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]
 

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