Tiny fix DeepseekScalingRotaryEmbedding always use forward_native (#5406)

Author: fzyzcjy

python/sglang/srt/layers/rotary_embedding.py

@@ -649,7 +649,7 @@ def forward_hip(self, *args, **kwargs):
         return self.forward_native(*args, **kwargs)
 
     def forward(self, *args, **kwargs):
-        if torch._dynamo.is_compiling:
+        if torch.compiler.is_compiling():
             return self.forward_native(*args, **kwargs)
         if _is_cuda_available:
             return self.forward_cuda(*args, **kwargs)