【NPU】 fix CANN_930 question (#1419)

Author: xuanyuanminzheng

backends/npu/kernels/clip_by_norm_kernel.cc

@@ -58,12 +58,15 @@ void NormKernel(const Context& dev_ctx,
 
   phi::Scalar p = 2.0f;
   const auto& x_dims = x.dims();
-  std::vector<int64_t> axis;
+  std::vector<int64_t> axis, resize_list;
   for (int64_t i = 0; i < x_dims.size(); ++i) {
     axis.push_back(i);
+    resize_list.push_back(1);
   }
-  bool keepdim = false;
+  bool keepdim = true;
+  x_norm->Resize(phi::make_ddim(resize_list));
   EXEC_NPU_CMD(aclnnNorm, dev_ctx, x, p, axis, keepdim, *x_norm);
+  x_norm->Resize({1});
 }
 
 template <typename T, typename Context>

backends/npu/kernels/concat_kernel.cc

@@ -215,7 +215,7 @@ void ConcatGradKernel(const Context& dev_ctx,
       ends_array.push_back(ins[j]->dims()[axis] + offset);
 
       std::vector<int64_t> steps;
-      for (int i = 0; i < outs[j]->dims().size(); i++) {
+      for (int i = 0; i < axes_t.size(); i++) {
         steps.push_back(1.0);
       }
 

backends/npu/kernels/slice_kernel.cc

@@ -184,7 +184,7 @@ void SliceRawKernel(const Context& dev_ctx,
   out->Resize(phi::make_ddim(size));
 
   std::vector<int64_t> steps;
-  for (int i = 0; i < out->dims().size(); i++) {
+  for (int i = 0; i < axes_t.size(); i++) {
     steps.push_back(1.0);
   }
 

backends/npu/kernels/tile_kernel.cc

@@ -316,7 +316,7 @@ void TileGradKernel(const Context& dev_ctx,
       for (int i = 0; i < out_grad.dims().size(); i++) {
         axes.push_back(i);
       }
-      std::vector<int64_t> steps(out_grad.dims().size(), 1);
+      std::vector<int64_t> steps(axes.size(), 1);
       static const auto aclCreateIntArray = GET_OP_API_FUNC(aclCreateIntArray);
       auto starts_acl = aclCreateIntArray(starts.data(), starts.size());
       auto ends_acl = aclCreateIntArray(x_grad_dims.data(), x_grad_dims.size());