fix：取消对saprse attention 的依赖

src/SanaModel.cpp

@@ -2,7 +2,7 @@
 
 #include "SanaModel.h"
 #include "kernels/zgemm/zgemm.h"
-#include "flash_api.h"
+// #include "flash_api.h"
 #include "kernels/misc_kernels.h"
 
 // #include <nvtx3/nvToolsExt.h>
@@ -180,6 +180,7 @@ Tensor MultiHeadCrossAttention::forward(Tensor x, Tensor cond, Tensor cu_seqlens
     //                                     false)
     //                          .front()
     //                          .view({batch_size, num_tokens_img, num_heads * head_dim});
+    Tensor attn_output = Tensor::ones({batch_size * num_tokens, num_heads, dim_head}, Tensor::FP16, Device::cuda());    
     std::cout << "mha_varlen_fwd not support !!!" << std::endl;
     // Tensor attn_output = mha_fwd(q, k, v,
     //     0.0f,