fix: 取消对sparse attention的依赖

README.md

+flash_c_api 放在third_party目录下
+
+编译指令：
 source /opt/dtk/env.sh
 
 source /usr/local/bin/fastpt -T

setup.py

@@ -100,7 +100,7 @@ if __name__ == "__main__":
         "third_party/json/include",
         "third_party/mio/include",
         "third_party/spdlog/include",
-        "third_party/Block-Sparse-Attention/csrc/block_sparse_attn",
+        # "third_party/Block-Sparse-Attention/csrc/block_sparse_attn",
         "third_party/flash_c_api/include",
     ]
 

src/FluxModel.cpp

@@ -278,6 +278,7 @@ Tensor Attention::forward(Tensor qkv, Tensor pool_qkv, float sparsityRatio) {
     //                                        -1,
     //                                        -1)
     //                              .front();
+    Tensor raw_attn_output = Tensor::ones({batch_size * num_tokens, num_heads, dim_head}, Tensor::FP16, Device::cuda());
     std::cout << "mha_fwd_block not support !!!" << std::endl;
     debug("raw_attn_output", raw_attn_output);