fix: 修改attetion这里的import bug

vllm/v1/attention/backends/flash_attn.py

@@ -566,8 +566,8 @@ class FlashAttentionImpl(AttentionImpl):
                     layer._v_scale,
                 )
             else:
-                from lightop import reshape_and_cache_cuda
                 if envs.VLLM_USE_OPT_RESHAPE_AND_CACHE and key.dtype == value.dtype == "fp16":
+                    from lightop import reshape_and_cache_cuda
                     reshape_and_cache_cuda(
                         key, value,
                         key_cache, value_cache,
@@ -576,6 +576,7 @@ class FlashAttentionImpl(AttentionImpl):
                         layer._k_scale, layer._v_scale
                     )
                 else:
+                    from vllm.attention.utils.fa_utils import reshape_and_cache_cuda
                     reshape_and_cache_cuda(
                         key,
                         value,