fix:yygrad_threadwise_copy InvalidElementAsNaN to false

example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_backward_fp16.cpp

@@ -30,7 +30,6 @@ Kernel outputs:
 #include <numeric>
 #include <initializer_list>
 #include <cstdlib>
-#include <fstream>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
@@ -611,9 +610,8 @@ int run(int argc, char* argv[])
               << gemm.GetTypeString() << std::endl;
 
     // copy z matirx data form device
-    std::ofstream file("./z_matrix_txt");
     z_device_buf.FromDevice(z_g_m_n.mData.data());
-    file << z_g_m_n << std::endl;
+
     //       std::cout << "z_g_m_n ref:\n" << z_g_m_n;
     bool pass = true;
     if(do_verification)

include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_v1.hpp

@@ -1649,8 +1649,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
             YDotYGrad_M_O::SrcScalarPerVector, // SrcScalarPerVector
             1,                                 // SrcScalarStrideInVector
             true /* ResetCoordAfterRun */,
-            true /* InvalidElementAsNaN */>(y_grid_desc_mblock_mperblock_oblock_operblock,
-                                            y_thread_data_on_grid_idx);
+            false /* InvalidElementAsNaN */>(y_grid_desc_mblock_mperblock_oblock_operblock,
+                                             y_thread_data_on_grid_idx);
 
         auto y_thread_buf                 = typename YDotYGrad_M_O::SrcBufType{};
         auto ygrad_thread_buf             = typename YDotYGrad_M_O::SrcBufType{};