refactor: raise error if flash attention is not installed

rotary.py

@@ -6,7 +6,11 @@ from typing import Optional, Tuple, Union
 
 import torch
 from einops import rearrange, repeat
-from flash_attn.ops.triton.rotary import apply_rotary
+try:
+    from flash_attn.ops.triton.rotary import apply_rotary
+except ImportError:
+    def apply_rotary(*args, **kwargs):
+        raise RuntimeError('RoPE requires flash-attention to be installed')
 
 
 def rotate_half(x, interleaved=False):