[ROCm] AITER fused RoPE+KVCache (#33443)

Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Signed-off-by: charlifu <charlifu@amd.com>
Signed-off-by: Rohan Potdar <66227218+Rohan138@users.noreply.github.com>
Co-authored-by: charlifu <charlifu@amd.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>

vllm/v1/attention/backend.py

@@ -723,6 +723,33 @@ class AttentionImpl(AttentionImplBase[T], Generic[T]):
         """
         return False
 
+    def fused_rope_kvcache_supported(self):
+        """
+        Does this attention implementation support RoPE+KVCache fusion.
+        This is used by the RopeKVCacheFusionPass to only fuse the RoPE ops
+        with the KV cache update for implementations that support it.
+        """
+        return False
+
+    def do_rope_and_kv_cache_update(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        positions: torch.Tensor,
+        cos_sin_cache: torch.Tensor,
+        is_neox: bool,
+        kv_cache: torch.Tensor,
+        layer_slot_mapping: torch.Tensor,
+    ):
+        """
+        If `fused_rope_kvcache_supported` returns True, this method will be called
+        by torch.ops.vllm.fused_rope_and_unified_kv_cache_update
+        to perform the inplace RoPE and KV cache update.
+        """
+        raise NotImplementedError
+
 
 class MLAAttentionImpl(AttentionImplBase[T], Generic[T]):
     """MLA attention implementation with forward_mqa and forward_mha methods."""