diff --git a/source/tnn/core/const_folder.cc b/source/tnn/core/const_folder.cc
index 4df4451e4..89985561b 100644
--- a/source/tnn/core/const_folder.cc
+++ b/source/tnn/core/const_folder.cc
@@ -161,7 +161,7 @@ Status ConstFolder::Forward() {
                     std::stringstream ss;
                     ss << "<" << blob->GetBlobDesc().name << "> shape:[";
                     for(int i: blob->GetBlobDesc().dims) {ss <<  i << ","; } ss << "]";
-                    LOGD("ConstFolder save const with name: %s\n", ss.str().c_str());
+                    LOGD("ConstFolder save const with name: %s, dtype=%d\n", ss.str().c_str(), blob->GetBlobDesc().data_type);
                 }
 #endif
                 
diff --git a/source/tnn/core/default_network.cc b/source/tnn/core/default_network.cc
index 6294f5be4..d48bdf39a 100644
--- a/source/tnn/core/default_network.cc
+++ b/source/tnn/core/default_network.cc
@@ -234,6 +234,7 @@ Status DefaultNetwork::InitLayers(NetStructure *net_structure, NetResource *net_
 
             cur_layer->InferShapeAhead(inputs, outputs_for_shape, layer_info->param.get(),
                                        net_resource->resource_map[layer_name].get());
+            LOGD("InferShapeAhead Output Shape: [%s]\n", cur_layer->GetOutputBlobs()[0]->GetBlobDesc().description().c_str());
 
             delete cur_layer;
         }
diff --git a/source/tnn/device/arm/acc/arm_expand_layer_acc.cc b/source/tnn/device/arm/acc/arm_expand_layer_acc.cc
index d1012ab1e..2bc608367 100644
--- a/source/tnn/device/arm/acc/arm_expand_layer_acc.cc
+++ b/source/tnn/device/arm/acc/arm_expand_layer_acc.cc
@@ -21,6 +21,10 @@ namespace TNN_NS {
 
 ArmExpandLayerAcc::~ArmExpandLayerAcc() {}
 
+bool ArmExpandLayerAcc::UseNaiveConstantBlobs() {
+    return true;
+}
+
 Status ArmExpandLayerAcc::InferRuntimeOutputShape(const std::vector<Blob *> &inputs, const std::vector<Blob *> &outputs) {
     auto expand_param = dynamic_cast<ExpandLayerParam*>(param_);
     CHECK_PARAM_NULL(expand_param);
diff --git a/source/tnn/device/arm/acc/arm_expand_layer_acc.h b/source/tnn/device/arm/acc/arm_expand_layer_acc.h
index 81d1ecf12..75f8344b2 100644
--- a/source/tnn/device/arm/acc/arm_expand_layer_acc.h
+++ b/source/tnn/device/arm/acc/arm_expand_layer_acc.h
@@ -27,6 +27,9 @@ class ArmExpandLayerAcc : public ArmLayerAcc {
     virtual Status InferRuntimeOutputShape(const std::vector<Blob *> &inputs, const std::vector<Blob *> &outputs) override;
 
     virtual Status DoForward(const std::vector<Blob *> &inputs, const std::vector<Blob *> &outputs) override;
+
+protected:
+    virtual bool UseNaiveConstantBlobs() override;
 };
 
 }  // namespace TNN_NS
diff --git a/source/tnn/device/arm/acc/arm_padv2_layer_acc.cc b/source/tnn/device/arm/acc/arm_padv2_layer_acc.cc
index 57dc0f825..62bae1d7b 100644
--- a/source/tnn/device/arm/acc/arm_padv2_layer_acc.cc
+++ b/source/tnn/device/arm/acc/arm_padv2_layer_acc.cc
@@ -32,7 +32,7 @@ Status ArmPadV2LayerAcc::DoForward(const std::vector<Blob *> &inputs, const std:
     auto input_dims  = input_blob->GetBlobDesc().dims;
     auto output_dims = output_blob->GetBlobDesc().dims;
     PadUtils::PadContext pad_context;
-    if (input_dims.size() < 2 || input_dims.size() > 5) {
+    if (input_dims.size() < 2 || input_dims.size() > 6) {
         LOGE("Error: ArmPadV2LayerAcc layer acc does not support input dims size %lu\n", input_dims.size());
         return Status(TNNERR_MODEL_ERR, "Error: ArmPadV2LayerAcc layer acc does not support;");
     }
@@ -62,7 +62,7 @@ Status ArmPadV2LayerAcc::DoForward(const std::vector<Blob *> &inputs, const std:
         pad_context.output_width = output_dims[4];
     }
     const auto pads = layer_param->pads;
-    if (pads.size() < 2 || pads.size() > 10) {
+    if (pads.size() < 2 || pads.size() > 12) {
         LOGE("Error: ArmPadV2LayerAcc layer acc does not support pas size %lu\n", pads.size());
         return Status(TNNERR_MODEL_ERR, "Error: ArmPadV2LayerAcc layer acc does not support");
     }
@@ -107,6 +107,9 @@ Status ArmPadV2LayerAcc::DoForward(const std::vector<Blob *> &inputs, const std:
             pad_context.pad_r   = pads[9];  // pad width end
             break;
         }
+        case 12: {
+            pad_context.pads = pads;
+        }
     }
 
     pad_context.type  = layer_param->type;
diff --git a/source/tnn/device/arm/acc/arm_scatter_nd_layer_acc.cc b/source/tnn/device/arm/acc/arm_scatter_nd_layer_acc.cc
index 3c6025dab..3a737ebd2 100644
--- a/source/tnn/device/arm/acc/arm_scatter_nd_layer_acc.cc
+++ b/source/tnn/device/arm/acc/arm_scatter_nd_layer_acc.cc
@@ -19,7 +19,11 @@
 
 namespace TNN_NS {
 
-DECLARE_ARM_ACC(ScatterND, LAYER_SCATTER_ND);
+DECLARE_ARM_ACC_WITH_EXTRA(ScatterND, LAYER_SCATTER_ND,
+    virtual bool UseNaiveConstantBlobs() {
+        return true;
+    }
+);
 
 Status ArmScatterNDLayerAcc::DoForward(const std::vector<Blob*>& inputs, const std::vector<Blob*>& outputs) {
     auto resource = dynamic_cast<ScatterNDLayerResource*>(resource_);
diff --git a/source/tnn/device/arm/acc/arm_stride_slice_layer_acc.cc b/source/tnn/device/arm/acc/arm_stride_slice_layer_acc.cc
index f75c1a36a..157eb1ebd 100644
--- a/source/tnn/device/arm/acc/arm_stride_slice_layer_acc.cc
+++ b/source/tnn/device/arm/acc/arm_stride_slice_layer_acc.cc
@@ -161,7 +161,12 @@ Status ArmStrideSliceLayerAcc::DoForward(const std::vector<Blob *> &inputs, cons
 REGISTER_ARM_ACC(StrideSlice, LAYER_STRIDED_SLICE)
 REGISTER_ARM_LAYOUT(LAYER_STRIDED_SLICE, DATA_FORMAT_NC4HW4)
 
-DECLARE_ARM_ACC(StrideSliceV2, LAYER_STRIDED_SLICE_V2);
+
+DECLARE_ARM_ACC_WITH_EXTRA(StrideSliceV2, LAYER_STRIDED_SLICE_V2,
+    virtual bool UseNaiveConstantBlobs() {
+        return true;
+    }
+);
 
 static Status FastSliceForHW(const std::vector<Blob *> &inputs, const std::vector<Blob *> &outputs,
                              StrideSliceV2LayerParam *param) {
diff --git a/source/tnn/device/arm/acc/compute/pad_function.cc b/source/tnn/device/arm/acc/compute/pad_function.cc
index 686b4b084..4e374ecc6 100644
--- a/source/tnn/device/arm/acc/compute/pad_function.cc
+++ b/source/tnn/device/arm/acc/compute/pad_function.cc
@@ -50,6 +50,92 @@ static void CommonPadImpl(float *input_data, float *output_data, int batch_c_r4,
     }
 }
 
+// Common pad in height and width directions
+static Status CommonPadDimSixImpl(float *input_data, float *output_data, int batch_c_r4, DimsVector input_dims, DimsVector output_dims, DimsVector pads, Float4 &vvalue) {
+    if (input_dims[2] != output_dims[2] || input_dims[4] != output_dims[4]) {
+        LOGE("CommonPadDimSixImpl(const type) only support 3 & 5 dims\n");
+        return Status(TNNERR_UNKNOWN_LAYER, "CommonPadDimSixImpl(const type) only support 3 & 5 dims");
+    }
+
+   int height = input_dims[2];
+   int width  = input_dims[4];
+
+   int in_h  = input_dims[3];
+   int in_w  = input_dims[5];
+   int out_h = output_dims[3];
+   int out_w = output_dims[5];
+
+   int pad_t = pads[3];
+   int pad_b = pads[9];
+   int pad_l = pads[5];
+   int pad_r = pads[11];
+
+    /*
+    printf("input_dims: ");
+    for (auto dim : input_dims) {
+        printf("%d, ", dim);
+    }
+    printf("\n");
+
+    printf("output_dims:");
+    for (auto dim : output_dims) {
+        printf("%d, ", dim);
+    }
+    printf("\n");
+
+    printf("pads:       ");
+    for (auto dim : pads) {
+        printf("%d, ", dim);
+    }
+    printf("\n");
+
+    input_dims: 1, 512, 32, 1, 32, 1, 
+    output_dims:1, 512, 32, 2, 32, 2, 
+    pads:       0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1,
+    */
+
+    for (int c = 0; c < batch_c_r4; c += 4) {
+        auto input_ptr_c  = input_data + c * height * in_h * width * in_w;
+        auto output_ptr_c = output_data + c * height * out_h * width * out_w;
+
+        for (int hi = 0; hi < height; ++hi) {
+            if (pad_t)
+                for (int i = 0; i < width * out_w * pad_t; ++i)
+                    Float4::save(output_ptr_c + i * 4, vvalue);
+
+            for (int h = 0; h < in_h; ++h) {
+                auto output_ptr_h = output_ptr_c + width * out_w * (h + pad_t) * 4;
+                auto input_ptr_h  = input_ptr_c + width * in_w * h * 4;
+
+                for (int w = 0; w < width; ++w) {
+                    auto output_ptr_w = output_ptr_h + w * out_w * 4;
+                    auto input_ptr_w  = input_ptr_h + w * in_w * 4;
+
+                    for (int i = 0; i < pad_l; i++)
+                        Float4::save(output_ptr_w + i * 4, vvalue);
+
+                    memcpy(output_ptr_w + pad_l * 4, input_ptr_w, in_w * sizeof(float) * 4);
+
+                    for (int i = in_w + pad_l; i < out_w; i++)
+                        Float4::save(output_ptr_w + i * 4, vvalue);
+                }
+            }
+
+            if (pad_b) {
+                auto output_ptr_h = output_ptr_c + width * out_w * (in_h + pad_t) * 4;
+                for (int i = 0; i < width * out_w * pad_b; ++i)
+                    Float4::save(output_ptr_h + i * 4, vvalue);
+            }
+
+            input_ptr_c += in_h * width * in_w * 4;
+            output_ptr_c += out_h * width * out_w * 4;
+        }
+
+    }
+
+    return TNN_OK;
+}
+
 static void CalculatePad(Float4 &src, const Float4 &vvalue, const int padded_zero) {
     if (padded_zero)
         src = Float4::pad(src, vvalue, padded_zero);
@@ -218,9 +304,9 @@ static void ChannelPadImpl(float *input_data, float *output_data, int batch, int
 
 Status PadUtils::ConstPadV2(float *input_data, float *output_data, DimsVector input_dims, DimsVector output_dims,
                             PadContext context) {
-    if (input_dims.size() < 2 || input_dims.size() > 5) {
-        LOGE("Arm PadV2(const type) only support 2 - 5 dims\n");
-        return Status(TNNERR_UNKNOWN_LAYER, "Arm PadV2 only support 2 - 5 dims");
+    if (input_dims.size() < 2 || input_dims.size() > 6) {
+        LOGE("Arm PadV2(const type) only support 2 - 6 dims\n");
+        return Status(TNNERR_UNKNOWN_LAYER, "Arm PadV2 only support 2 - 6 dims");
     }
     const int batch    = context.output_batch;
     const int oc_r4    = context.output_channel_r4;
@@ -237,7 +323,16 @@ Status PadUtils::ConstPadV2(float *input_data, float *output_data, DimsVector in
     const int pad_b    = context.pad_b;
     const int pad_r    = context.pad_r;
     Float4 value_v     = Float4(context.value);
-    
+
+    if (input_dims.size() == 6) {
+        if (input_dims[0] == output_dims[0] && context.pads[1] == context.pads[7]) {
+            return CommonPadDimSixImpl(input_data, output_data, batch * oc_r4, input_dims, output_dims, context.pads, value_v);
+        } else {
+            LOGE("Arm PadV2(const type) with 6 dims do not support channel dim\n");
+            return Status(TNNERR_UNKNOWN_LAYER, "Arm PadV2(const type) with 6 dims do not support channel dim");
+        }
+    }
+
     //ncdhw, extend dim except the batch n
     if (context.input_batch == context.output_batch) {
     if (pad_c_b == 0 && pad_c_e == 0) {
diff --git a/source/tnn/device/arm/acc/compute/pad_function.h b/source/tnn/device/arm/acc/compute/pad_function.h
index c76e74041..3f56a3e75 100644
--- a/source/tnn/device/arm/acc/compute/pad_function.h
+++ b/source/tnn/device/arm/acc/compute/pad_function.h
@@ -47,6 +47,7 @@ class PadUtils {
         int32_t pad_b   = 0;
         int32_t pad_l   = 0;
         int32_t pad_r   = 0;
+        DimsVector pads;
         int32_t type    = 0;
         float value     = 0.0f;
     };
diff --git a/source/tnn/device/cpu/acc/cpu_layer_acc.cc b/source/tnn/device/cpu/acc/cpu_layer_acc.cc
index 1ad4399a7..d77564d07 100644
--- a/source/tnn/device/cpu/acc/cpu_layer_acc.cc
+++ b/source/tnn/device/cpu/acc/cpu_layer_acc.cc
@@ -55,13 +55,22 @@ Status CpuLayerAcc::ReloadConstantBlobs(const std::vector<Blob *> &inputs, bool
         if (const_blob_map.find(name) != const_blob_map.end()) {
             blob = const_blob_map[name];
         }
-        auto status = RawBuffer2Blob(buffer.get(), blob);
+
+        auto buffer_cvt = RawBuffer();
+        if (buffer->GetDataType() == DATA_TYPE_HALF) {
+            buffer_cvt = ConvertHalfHandle(*(buffer.get()));
+        } else {
+            buffer_cvt = *(buffer.get());
+        }
+
+        auto status = RawBuffer2Blob(&buffer_cvt, blob);
         RETURN_ON_NEQ(status, TNN_OK);
 
         blob->SetFlag(DATA_FLAG_CHANGE_NEVER);
         const_blob_map[name] = blob;
         iter->SetHandle(blob->GetHandle());
-        LOGD("Reload constant blob: %s %p\n", name.c_str(), &blob);
+        iter->SetBlobDesc(blob->GetBlobDesc());
+        LOGD("Reload constant blob: %s %p, dtype=%d\n", name.c_str(), &blob, blob->GetBlobDesc().data_type);
     }
     const_blob_map_ = const_blob_map;
     return TNN_OK;
diff --git a/source/tnn/layer/conv_layer.cc b/source/tnn/layer/conv_layer.cc
index f5cd2ade5..1d304ab41 100644
--- a/source/tnn/layer/conv_layer.cc
+++ b/source/tnn/layer/conv_layer.cc
@@ -32,9 +32,14 @@ Status ConvLayer::InferOutputShape(bool ignore_error) {
     ConvLayerParam* conv_param = dynamic_cast<ConvLayerParam*>(param_);
     CHECK_PARAM_NULL(conv_param);
 
-    int num    = input_blob->GetBlobDesc().dims[0];
-    int height = input_blob->GetBlobDesc().dims[2];
-    int width  = input_blob->GetBlobDesc().dims[3];
+    if (input_blob->GetBlobDesc().dims.size() == 0) {
+        LOGE_IF(!ignore_error, "Error: dims not supported\n");
+        return Status(TNNERR_PARAM_ERR, "Error: dims not supported");
+    }
+
+    int num    = DimsVectorUtils::Count(input_blob->GetBlobDesc().dims, 0, 1);
+    int height = DimsVectorUtils::Count(input_blob->GetBlobDesc().dims, 2, 3);
+    int width  = DimsVectorUtils::Count(input_blob->GetBlobDesc().dims, 3, 4);
 
     const int pad_w_begin = conv_param->pads[0];
     const int pad_h_begin = conv_param->pads[2];
diff --git a/source/tnn/layer/pad_layer.cc b/source/tnn/layer/pad_layer.cc
index 76dae2e06..5d413ca91 100644
--- a/source/tnn/layer/pad_layer.cc
+++ b/source/tnn/layer/pad_layer.cc
@@ -36,6 +36,11 @@ Status PadLayer::InferOutputShape(bool ignore_error) {
     Blob* input_blob  = input_blobs_[0];
     Blob* output_blob = output_blobs_[0];
     auto dims         = input_blob->GetBlobDesc().dims;
+    if (dims.size() < 4) {
+        LOGE_IF(!ignore_error, "Error: dims not supported\n");
+        return Status(TNNERR_PARAM_ERR, "Error: dims not supported");
+    }
+
     dims[3] += layer_param->pads[0] + layer_param->pads[1];
     dims[2] += layer_param->pads[2] + layer_param->pads[3];
     dims[1] += layer_param->pads[4] + layer_param->pads[5];
diff --git a/source/tnn/optimizer/graph_matcher/ir.cc b/source/tnn/optimizer/graph_matcher/ir.cc
index 264b760f4..8703f6164 100644
--- a/source/tnn/optimizer/graph_matcher/ir.cc
+++ b/source/tnn/optimizer/graph_matcher/ir.cc
@@ -581,7 +581,7 @@ namespace TNN_NS {
             RETURN_ON_FAIL(IsConnectedGraph(this, connected));
         }
         if (!connected) {
-            ERRORV("the graph is not connected.", msg);
+            DEBUGV("the graph is not connected.", msg);
             return Status(TNNERR_COMMON_ERROR, msg);
         }
         return TNN_OK;
diff --git a/source/tnn/optimizer/graph_matcher/logger.h b/source/tnn/optimizer/graph_matcher/logger.h
index be2df52b1..f414c8f60 100644
--- a/source/tnn/optimizer/graph_matcher/logger.h
+++ b/source/tnn/optimizer/graph_matcher/logger.h
@@ -144,6 +144,12 @@ class Logger {
     ::TNN_NS::Logger::instance().log(std::string(__ss), ::TNN_NS::Logger::kLogLevelDebug());\
   } while(0)
 
+#define DEBUGV(f_, vname_, ...) \
+  __LOG_FORMAT(vname_, 2000, f_, ##__VA_ARGS__); \
+  do { \
+    ::TNN_NS::Logger::instance().log(std::string(vname_), ::TNN_NS::Logger::kLogLevelDebug());\
+  } while(0)
+
 #define INFO(f_, ...) \
   do { \
     __LOG_FORMAT(__ss, 2000, f_, ##__VA_ARGS__); \
diff --git a/source/tnn/optimizer/net_optimizer_convert_matmul_to_conv.cc b/source/tnn/optimizer/net_optimizer_convert_matmul_to_conv.cc
index ab3c14307..ba896cecb 100644
--- a/source/tnn/optimizer/net_optimizer_convert_matmul_to_conv.cc
+++ b/source/tnn/optimizer/net_optimizer_convert_matmul_to_conv.cc
@@ -74,7 +74,7 @@ namespace optimizer {
         std::shared_ptr<Graph> graph = std::make_shared<Graph>();
         auto status = graph->fromInterpreted(structure, resource);
         if (status != TNN_OK) {
-            LOGE("%s", status.description().c_str());
+            LOGD("%s", status.description().c_str());
             return TNN_OK;
         }
 
diff --git a/source/tnn/utils/dims_function_utils.cc b/source/tnn/utils/dims_function_utils.cc
index dbc7954bc..f6bbe3ed2 100644
--- a/source/tnn/utils/dims_function_utils.cc
+++ b/source/tnn/utils/dims_function_utils.cc
@@ -51,6 +51,13 @@ DimsVector DimsFunctionUtils::Expand(DimsVector dims0, DimsVector dims1, Status
 
 DimsVector DimsFunctionUtils::Upsample(const DimsVector input_dims,
                                      std::vector<float> scales, std::vector<int> sizes, int mode, Status *status) {
+    if (input_dims.size() == 0) {
+        if (status) {
+            *status = Status(TNNERR_PARAM_ERR, "unsupport upsample input dims");
+        }
+        return DimsVector();
+    }
+
     int num          = input_dims[0];
     int channels   = input_dims[1];
     int height       = input_dims[2];
@@ -108,6 +115,13 @@ DimsVector DimsFunctionUtils::Reshape(const DimsVector input_dims, const DimsVec
     int output_size = shape.size() + axis;
     DimsVector output_dims(output_size, 1);
 
+    if (input_dims.size() == 0) {
+        if (status) {
+            *status = Status(TNNERR_LAYER_ERR, "reshape input_dims size error");
+        }
+        return DimsVector();
+    }
+
     for(int i = 0; i < axis; ++i) {
         output_dims[i] = input_dims[i];
     }
diff --git a/test/unit_test/layer_test/test_padv2_layer.cc b/test/unit_test/layer_test/test_padv2_layer.cc
index a8b8a8769..02d1edb45 100644
--- a/test/unit_test/layer_test/test_padv2_layer.cc
+++ b/test/unit_test/layer_test/test_padv2_layer.cc
@@ -34,7 +34,7 @@ INSTANTIATE_TEST_SUITE_P(LayerTest, PadV2LayerTest,
                                             // pad_type
                                             testing::Values(0, 1),
                                             // dim size
-                                            testing::Values(3, 4, 5),
+                                            testing::Values(3, 4, 5, 6),
                                             // pad value
                                             testing::Values(-FLT_MAX, 0, 2, FLT_MAX)));
 
@@ -67,8 +67,8 @@ TEST_P(PadV2LayerTest, PadV2Layer) {
     if (!(DEVICE_CUDA == dev || DEVICE_ARM == dev || DEVICE_OPENCL == dev || DEVICE_METAL == dev)) {
         GTEST_SKIP();
     }
-    // arm only support dims size 4
-    if (DEVICE_ARM == dev && dim_count != 4) {
+    // arm only support dims size 4 and 6
+    if (DEVICE_ARM == dev && (dim_count != 4 && dim_count != 6)) {
         GTEST_SKIP();
     }
     // opnecl only support dims size 4
@@ -81,6 +81,20 @@ TEST_P(PadV2LayerTest, PadV2Layer) {
         GTEST_SKIP();
     }
 
+    if (dim_count == 6) {
+        // only arm support dims size 6
+        if (DEVICE_ARM != dev) {
+            GTEST_SKIP();
+        }
+        if (pad_type != 0) {
+            GTEST_SKIP();
+        }
+        // reduce test time
+        if (input_size >= 10) {
+            GTEST_SKIP();
+        }
+    }
+
     // param
     std::shared_ptr<PadLayerParam> param(new PadLayerParam());
     param->name = "PadV2";
@@ -93,6 +107,8 @@ TEST_P(PadV2LayerTest, PadV2Layer) {
         param->pads = {0, pad_c, pad_h, pad_w, 0, pad_c, pad_h, pad_w};
     } else if (dim_count == 5) {
         param->pads = {0, pad_c, pad_h, pad_w, pad_w, 0, pad_c, pad_h, pad_w, pad_w};
+    } else if (dim_count == 6) {
+        param->pads = {0, 0, 0, pad_h, 0, pad_w, 0, 0, 0, pad_h, 0, pad_w};
     }
     param->value = value;