diff --git a/tools/modelwriter.h b/tools/modelwriter.h
index 9fc24369368..8707fe9eb81 100644
--- a/tools/modelwriter.h
+++ b/tools/modelwriter.h
@@ -796,7 +796,7 @@ int ModelWriter::save(const char* parampath, const char* binpath)
             continue;
         }
 
-        ncnn::Layer* layer_default = ncnn::create_layer(layer->typeindex);
+        ncnn::Layer* layer_default = ncnn::create_layer_cpu(layer->typeindex);
 
         ncnn::ParamDict pd;
         layer_default->load_param(pd);
@@ -1764,6 +1764,19 @@ int ModelWriter::save(const char* parampath, const char* binpath)
             fprintf_param_value(" 20=%d", constant_TILE_M)
             fprintf_param_value(" 21=%d", constant_TILE_N)
             fprintf_param_value(" 22=%d", constant_TILE_K)
+
+            if (op->constantA == 1)
+            {
+                fwrite_weight_tag_data(op->A_data, bp);
+            }
+            if (op->constantB == 1)
+            {
+                fwrite_weight_tag_data(op->B_data, bp);
+            }
+            if (op->constantC == 1 && op->constant_broadcast_type_C != -1)
+            {
+                fwrite_weight_tag_data(op->C_data, bp);
+            }
         }
         else if (layer->type == "GLU")
         {
diff --git a/tools/ncnnoptimize.cpp b/tools/ncnnoptimize.cpp
index d1a0cf63633..001e62a3445 100644
--- a/tools/ncnnoptimize.cpp
+++ b/tools/ncnnoptimize.cpp
@@ -1993,7 +1993,7 @@ int NetOptimize::fuse_binaryop_eltwise()
 
         fprintf(stderr, "fuse_binaryop_eltwise %s %s %s\n", binaryop0->name.c_str(), binaryop1->name.c_str(), binaryop->name.c_str());
 
-        ncnn::Eltwise* eltwise = (ncnn::Eltwise*)ncnn::create_layer("Eltwise");
+        ncnn::Eltwise* eltwise = (ncnn::Eltwise*)ncnn::create_layer_cpu("Eltwise");
 
         eltwise->type = "Eltwise";
         eltwise->name = binaryop->name;
@@ -2554,7 +2554,7 @@ int NetOptimize::replace_reduction_with_global_pooling()
 
         fprintf(stderr, "replace_reduction_with_global_pooling %s %s\n", reduction1->name.c_str(), reduction2->name.c_str());
 
-        ncnn::Pooling* pooling = (ncnn::Pooling*)ncnn::create_layer("Pooling");
+        ncnn::Pooling* pooling = (ncnn::Pooling*)ncnn::create_layer_cpu("Pooling");
 
         pooling->type = "Pooling";
         pooling->name = reduction2->name;
@@ -2593,7 +2593,7 @@ int NetOptimize::replace_prelu_with_leaky_relu()
 
         fprintf(stderr, "replace_prelu_with_leaky_relu %s\n", prelu->name.c_str());
 
-        ncnn::ReLU* relu = (ncnn::ReLU*)ncnn::create_layer("ReLU");
+        ncnn::ReLU* relu = (ncnn::ReLU*)ncnn::create_layer_cpu("ReLU");
 
         relu->type = "ReLU";
         relu->name = prelu->name;
@@ -2647,7 +2647,7 @@ int NetOptimize::replace_convolution_with_innerproduct_after_global_pooling()
 
         fprintf(stderr, "replace_convolution_with_innerproduct_after_global_pooling %s %s\n", pooling->name.c_str(), convolution->name.c_str());
 
-        ncnn::InnerProduct* innerproduct = (ncnn::InnerProduct*)ncnn::create_layer("InnerProduct");
+        ncnn::InnerProduct* innerproduct = (ncnn::InnerProduct*)ncnn::create_layer_cpu("InnerProduct");
 
         innerproduct->type = "InnerProduct";
         innerproduct->name = convolution->name;
@@ -2715,7 +2715,7 @@ int NetOptimize::replace_convolution_with_innerproduct_after_innerproduct()
 
             fprintf(stderr, "replace_convolution_with_innerproduct_after_innerproduct %s %s\n", innerproduct->name.c_str(), convolution->name.c_str());
 
-            ncnn::InnerProduct* innerproduct2 = (ncnn::InnerProduct*)ncnn::create_layer("InnerProduct");
+            ncnn::InnerProduct* innerproduct2 = (ncnn::InnerProduct*)ncnn::create_layer_cpu("InnerProduct");
 
             innerproduct2->type = "InnerProduct";
             innerproduct2->name = convolution->name;
diff --git a/tools/quantize/ncnn2table.cpp b/tools/quantize/ncnn2table.cpp
index 9fbafa2d181..386bfff2a09 100644
--- a/tools/quantize/ncnn2table.cpp
+++ b/tools/quantize/ncnn2table.cpp
@@ -1112,7 +1112,7 @@ int QuantNet::quantize_EQ()
                 ncnn::Mat out;
                 ex.extract(conv_top_blobs[i], out);
 
-                ncnn::Layer* layer_int8 = ncnn::create_layer(layer->typeindex);
+                ncnn::Layer* layer_int8 = ncnn::create_layer_cpu(layer->typeindex);
 
                 ncnn::ParamDict pd;
                 get_layer_param(layer, pd);
@@ -1222,7 +1222,7 @@ int QuantNet::quantize_EQ()
                 ncnn::Mat out;
                 ex.extract(conv_top_blobs[i], out);
 
-                ncnn::Layer* layer_int8 = ncnn::create_layer(layer->typeindex);
+                ncnn::Layer* layer_int8 = ncnn::create_layer_cpu(layer->typeindex);
 
                 ncnn::ParamDict pd;
                 get_layer_param(layer, pd);