fix: memory leak of multi initialization.

When initializing desc the memory will be allocated. But it will allocate
again when loading the quantized weights. So one piece of memory will
lost which will make the app memory leaking.

bigquant/native/c_api.cc

@@ -112,10 +112,6 @@ void InternalQuantizedConvKernelInit(QuantizedTensorDesc *quantized_tensor, floa
 void InternalQuantizedConvKernelLoadFromModel(QuantizedTensorDesc *quantized_tensor, int8_t *src, float *min,
                                               float *max, size_t c_out, size_t c_in, size_t kernel_h, size_t kernel_w,
                                               float threshold, LAYOUT layout) {
-  aligned_malloc(&(quantized_tensor->min), 64, quantized_tensor->workspace_size_per_meta_info);
-  aligned_malloc(&(quantized_tensor->max), 64, quantized_tensor->workspace_size_per_meta_info);
-  aligned_malloc(&(quantized_tensor->ratio), 64, quantized_tensor->workspace_size_per_meta_info);
-  aligned_malloc(&(quantized_tensor->data), 64, quantized_tensor->workspace_size);
   std::vector<float> fp_model(c_out * c_in * kernel_h * kernel_w);
   DequantizeModel(fp_model.data(), src, min, max, c_out, c_in, kernel_h, kernel_w);
   float *tmp;
@@ -230,10 +226,6 @@ void InternalQuantizedFCKernelInit(QuantizedTensorDesc *quantized_tensor, float
 void InternalQuantizedFCKernelLoadFromModel(QuantizedTensorDesc *quantized_tensor, int8_t *src, float *min, float *max,
                                             size_t c_out, size_t c_in, float threshold, LAYOUT layout) {
   assert((layout == NCHW) || (layout == NHWC));
-  aligned_malloc(&(quantized_tensor->min), 64, quantized_tensor->workspace_size_per_meta_info);
-  aligned_malloc(&(quantized_tensor->max), 64, quantized_tensor->workspace_size_per_meta_info);
-  aligned_malloc(&(quantized_tensor->ratio), 64, quantized_tensor->workspace_size_per_meta_info);
-  aligned_malloc(&(quantized_tensor->data), 64, quantized_tensor->workspace_size);
   std::vector<float> fp_model(c_out * c_in);
   DequantizeModel(fp_model.data(), src, min, max, c_out, c_in, 1, 1);
   shuffle::PadQuantizeShuffle2D<float, FC_SHUFFLE_KERNEL_M, FC_SHUFFLE_KERNEL_K>(