bias parameters update to other example

example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_forward_v2.cpp

@@ -135,6 +135,7 @@ using DeviceGemmInstance =
         8,
         8,
         true,
+        4,
         S<16, 16, 1>, // B1BlockTransfer
         S<0, 2, 1>,
         S<0, 2, 1>,
@@ -146,6 +147,7 @@ using DeviceGemmInstance =
         1,              // CShuffleNXdlPerWavePerShuffle
         S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        4,
         MaskingSpec, // MaskingSpecialization
         Deterministic>;
 #elif(DIM <= 64)
@@ -206,6 +208,7 @@ using DeviceGemmInstance =
         8,
         8,
         true,
+        4,
         S<16, 16, 1>, // B1BlockTransfer
         S<0, 2, 1>,
         S<0, 2, 1>,
@@ -217,6 +220,7 @@ using DeviceGemmInstance =
         2,              // CShuffleNXdlPerWavePerShuffle
         S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        4,
         MaskingSpec, // MaskingSpecialization
         Deterministic>;
 #elif(DIM <= 128)
@@ -277,6 +281,7 @@ using DeviceGemmInstance =
         8,
         8,
         true,
+        4,
         S<8, 32, 1>, // B1BlockTransfer
         S<0, 2, 1>,
         S<0, 2, 1>,
@@ -288,6 +293,7 @@ using DeviceGemmInstance =
         2,              // CShuffleNXdlPerWavePerShuffle
         S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        4,
         MaskingSpec, // MaskingSpecialization
         Deterministic>;
 #endif

example/32_batched_gemm_scale_softmax_gemm/grouped_multihead_attention_forward_v2.cpp

@@ -135,6 +135,7 @@ using DeviceGemmInstance =
         8,
         8,
         true,
+        1,
         S<16, 16, 1>, // B1BlockTransfer
         S<0, 2, 1>,
         S<0, 2, 1>,
@@ -146,6 +147,7 @@ using DeviceGemmInstance =
         1,              // CShuffleNXdlPerWavePerShuffle
         S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        1,
         MaskingSpec, // MaskingSpecialization
         Deterministic>;
 #elif(DIM <= 64)
@@ -206,6 +208,7 @@ using DeviceGemmInstance =
         8,
         8,
         true,
+        1,
         S<16, 16, 1>, // B1BlockTransfer
         S<0, 2, 1>,
         S<0, 2, 1>,
@@ -217,6 +220,7 @@ using DeviceGemmInstance =
         2,              // CShuffleNXdlPerWavePerShuffle
         S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        1,
         MaskingSpec, // MaskingSpecialization
         Deterministic>;
 #elif(DIM <= 128)
@@ -277,6 +281,7 @@ using DeviceGemmInstance =
         8,
         8,
         true,
+        1,
         S<8, 32, 1>, // B1BlockTransfer
         S<0, 2, 1>,
         S<0, 2, 1>,
@@ -288,6 +293,7 @@ using DeviceGemmInstance =
         2,              // CShuffleNXdlPerWavePerShuffle
         S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        1,
         MaskingSpec, // MaskingSpecialization
         Deterministic>;
 #endif