vector load

example/52_flash_atten_bias/batched_multihead_attention_bias_backward_v2.cpp

@@ -420,7 +420,6 @@ int run(int argc, char* argv[])
         v_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<InputDataType>{-2, 2});
         ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_2<InputDataType>{-2, 2});
         d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<Acc0BiasDataType>{-2, 2});
-        //d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<Acc0BiasDataType>{1});
         break;
     case 2:
         q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<InputDataType>{0.0, 1.0});

include/ck/tensor_operation/gpu/grid/gridwise_batched_mha_bwd_xdl_cshuffle_qloop_b2t_v2.hpp

@@ -1226,8 +1226,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
             Sequence<1, 0, 2>,                         // DstDimAccessOrder
             1,                                         // SrcVectorDim
             2,                                         // DstVectorDim
-            1,                                         // SrcScalarPerVector
-            1,                                         // DstScalarPerVector
+            4,                                         // SrcScalarPerVector
+            4,                                         // DstScalarPerVector
             1,
             1,
             true,
@@ -1242,8 +1242,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
                                              Sequence<1, 1, 8, 1, 4>, // SliceLengths
                                              Sequence<0, 1, 2, 3, 4>, // DimAccessOrder
                                              4,                       // SrcVectorDim
-                                             1,              // SrcScalarPerVector
-                                             1>;
+                                             4,              // SrcScalarPerVector
+                                             4>;
     };
 
     template <bool HasMainKBlockLoop,