add sinks of vllm_flash_attn_varlen_func

vllm/v1/attention/backends/flash_attn.py

@@ -644,7 +644,7 @@ class FlashAttentionImpl(AttentionImpl):
                     # k_descale=layer._k_scale.expand(descale_shape),
                     # v_descale=layer._v_scale.expand(descale_shape),
                     # num_splits=attn_metadata.max_num_splits,
-                    # s_aux=self.sinks,
+                    s_aux=self.sinks,
                     is_prefix_cache=True,
                 )
             return output