fix p_acc1_biases size issue

example/52_flash_atten_bias/grouped_multihead_attention_bias_backward_v2.cpp

@@ -24,7 +24,7 @@ Kernel outputs:
 */
 
 #define USING_MASK 0
-#define DIM 64 // DIM should be a multiple of 8.
+#define DIM 128 // DIM should be a multiple of 8.
 
 #include <iostream>
 #include <numeric>

include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_bwd_xdl_cshuffle_qloop_v2.hpp

@@ -797,16 +797,12 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
                  group_count_ == ck::type_convert<ck::index_t>(p_Vgrads.size()) &&
                  group_count_ == ck::type_convert<ck::index_t>(p_LSEs.size()) &&
                  (group_count_ == ck::type_convert<ck::index_t>(p_acc0_biases.size()) ||
-                  ck::type_convert<ck::index_t>(p_acc0_biases.size() == 0))))
+                  ck::type_convert<ck::index_t>(p_acc0_biases.size() == 0)) &&
+                 0 == p_acc1_biases.size()))
             {
                 throw std::runtime_error("wrong! group_count_ != p_As/b/b1/c.size");
             }
 
-            if(!(p_acc0_biases.size() == p_acc1_biases.size()))
-            {
-                throw std::runtime_error("wrong! acc0_bias_vec.size != acc1_bias_vec.size");
-            }
-
             grid_size_ = 0;
 
             index_t z_random_matrix_offset = 0;