[AMD] Fix RotaryEmbedding crash on AMD/ROCm (regression from #17934) (#18903)

Co-authored-by: michaelzhang-ai <michaelzhang-ai@users.noreply.github.com>

python/sglang/srt/layers/rotary_embedding.py

@@ -389,6 +389,19 @@ class RotaryEmbedding(MultiPlatformOp):
             )
         return query, key
 
+    def forward_hip(self, *args, **kwargs):
+        """HIP/ROCm implementation.
+
+        The JIT kernels (sglang.jit_kernel.pos_enc) used in forward_cuda's
+        fallback path depend on tvm_ffi which invokes nvidia-smi to detect
+        CUDA compute capability. This fails on AMD GPUs, so we use the
+        pure-PyTorch native implementation instead.
+
+        Uses *args/**kwargs because subclasses (MRotaryEmbedding, etc.)
+        have different forward_native() signatures.
+        """
+        return self.forward_native(*args, **kwargs)
+
     def extra_repr(self) -> str:
         s = f"head_size={self.head_size}, rotary_dim={self.rotary_dim}"
         s += f", max_position_embeddings={self.max_position_embeddings}"