skip apply_rotary_emb from vllm_fa

vllm/model_executor/layers/rotary_embedding.py

@@ -77,7 +77,7 @@ def _apply_rotary_emb(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor,
         is_neox_style: Whether to use the Neox-style or GPT-J-style rotary
             positional embeddings.
     """
-    if current_platform.is_cuda_alike():
+    if current_platform.is_cuda_alike() and not current_platform.is_rocm():
         from vllm.vllm_flash_attn.layers.rotary import apply_rotary_emb
         return apply_rotary_emb(x.unsqueeze(0), cos, sin,
                                 not is_neox_style).squeeze(0)