diff --git a/library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_bf16_instances.hpp b/library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_bf16_instances.hpp
index 8fe58fd404..a2d9fc358e 100644
--- a/library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_bf16_instances.hpp
+++ b/library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_bf16_instances.hpp
@@ -37,13 +37,10 @@ using tile_grouped_conv_bwd_weight_bf16_instances = std::tuple<
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     32,     32,      8,      2,     4,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      2,     4,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      4,     4,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,      8,      2,     2,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,     16,      2,     2,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     16,     16,     32,      2,     2,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
 
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
@@ -53,19 +50,37 @@ using tile_grouped_conv_bwd_weight_bf16_instances = std::tuple<
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     32,     32,      8,      2,     4,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      2,     4,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      4,     4,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      8,     8,       8,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,      8,      2,     2,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,     16,      2,     2,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     16,     16,     32,      2,     2,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
 
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      2,     2,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
-#endif    
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>
+#endif
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      2,      2,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      1,      1,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      32,      32,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      32,      32,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      8,     8,       8,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        1,      32,      32,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,  PassThrough, PassThrough, PassThrough,        2,      32,      32,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>
     // clang-format on
 >;
 
diff --git a/library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_fp16_instances.hpp b/library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_fp16_instances.hpp
index 1d4dc8c49f..29fa86120f 100644
--- a/library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_fp16_instances.hpp
+++ b/library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_fp16_instances.hpp
@@ -37,13 +37,10 @@ using tile_grouped_conv_bwd_weight_f16_instances = std::tuple<
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     32,     32,      8,      2,     4,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      2,     4,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      4,     4,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,      8,      2,     2,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,     16,      2,     2,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     16,     16,     32,      2,     2,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
 
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
@@ -53,19 +50,37 @@ using tile_grouped_conv_bwd_weight_f16_instances = std::tuple<
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     32,     32,      8,      2,     4,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      2,     4,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      4,     4,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      8,     8,       8,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,      8,      2,     2,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,     16,      2,     2,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     16,     16,     32,      2,     2,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
 
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
     GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      4,     4,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
 #endif
-    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      2,      2,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      64,      64,      64,      1,      1,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      1,      1,      1,     32,     32,     16,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      8,     8,       8,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      32,      32,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      32,      32,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,   false,  CK_TILE_PIPELINE_COMPUTE_V3>,
+    
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       2,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      64,      64,      64,      2,      2,      1,     16,     16,     32,      8,     8,       8,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      16,      16,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        1,      32,      32,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>,
+    GroupedConvolutionBackwardWeightInvoker<NDimSpatial,   ALayout,   BLayout,   ELayout,    F16,     F16,     F16,  PassThrough, PassThrough, PassThrough,        2,      32,      32,      64,      1,      1,      1,     16,     16,     32,      8,     8,       4,    true,  CK_TILE_PIPELINE_COMPUTE_V4>
 // clang-format on
 >;