npu fused op (#7386)

Co-authored-by: Li Junwen <lijunwen13@hisilicon.com>

python/sglang/srt/custom_op.py

@@ -1,11 +1,12 @@
 from torch import nn
 
-from sglang.srt.utils import cpu_has_amx_support, is_cpu, is_cuda, is_hip
+from sglang.srt.utils import cpu_has_amx_support, is_cpu, is_cuda, is_hip, is_npu
 
 _is_cuda = is_cuda()
 _is_hip = is_hip()
 _is_cpu = is_cpu()
 _is_cpu_amx_available = cpu_has_amx_support()
+_is_npu = is_npu()
 
 
 class CustomOp(nn.Module):
@@ -60,6 +61,9 @@ class CustomOp(nn.Module):
     def forward_cuda(self, *args, **kwargs):
         raise NotImplementedError
 
+    def forward_npu(self, *args, **kwargs):
+        raise NotImplementedError
+
     def forward_hip(self, *args, **kwargs):
         return self.forward_cuda(*args, **kwargs)
 
@@ -79,5 +83,7 @@ class CustomOp(nn.Module):
             return self.forward_hip
         elif _is_cpu and _is_cpu_amx_available:
             return self.forward_cpu
+        elif _is_npu:
+            return self.forward_npu
         else:
             return self.forward_native