readme中增加llama.c

README.md

@@ -16,6 +16,15 @@
 
 **视频课程链接：**[https://space.bilibili.com/1822828582](https://space.bilibili.com/1822828582)
 
+## 对llama的支持
+> 我们将[llama.c](https://github.com/karpathy/llama2.c )中的算子替换为了KuiperInfer中的实现
+
+模型下载地址
+
+链接:https://pan.baidu.com/s/1PF5KqvIvNFR8yDIY1HmTYA?pwd=ma8r 
+
+![a.gif](imgs/a.gif)
+
 ## 第二次课程大纲
 
 第二次课程是第一次课程的重置版，内容更加充实和完善，第一次课程大纲见下方章节。

demos/llama2/CMakeLists.txt

@@ -1,12 +1,12 @@
 cmake_minimum_required(VERSION 3.16)
 set(CMAKE_CXX_STANDARD 17)
 
-#if (MSVC)
-#    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /O2")
-#else ()
-#    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -fopenmp -march=native")
-#endif ()
-#set(CMAKE_BUILD_TYPE Release)
+if (MSVC)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /O2")
+else ()
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -fopenmp -march=native")
+endif ()
+set(CMAKE_BUILD_TYPE Release)
 add_executable(llama_chat llama_chat.cpp main.cpp)
 target_link_directories(llama_chat PUBLIC ${PROJECT_SOURCE_DIR}/lib)
 target_link_libraries(llama_chat ${OpenCV_LIBS} kuiper)

source/layer/details/matmul.cpp

@@ -117,8 +117,7 @@ StatusCode LLamaMatmulLayer::Forward(const std::vector<std::shared_ptr<Tensor<fl
 #pragma omp parallel for
       for (int j = 0; j < weight_dim0_; ++j) {
         arma::fmat sub_weight(weight_ptr + j * weight_dim1_, weight_dim1_, 1, false, true);
-        arma::fmat output_mat(output_ptr + j, 1, 1, false, true);
-        output_mat = input_vec * sub_weight;
+        *(output_ptr + j) = as_scalar(input_vec * sub_weight);
       }
     } else if (weight_dim0_ == 1) {
       arma::fmat output_mat(output->raw_ptr(), input_dim1, weight_dim0_, false, true);

test/test_layer/test_softmax.cpp

@@ -211,6 +211,32 @@ TEST(test_layer, forward_softmax_dim01) {
   }
 }
 
+
+TEST(test_layer, forward_softmax_dim011) {
+  using namespace kuiper_infer;
+  SoftmaxLayer softmax_layer(0);
+  uint32_t size = 97;
+  std::vector<float> values;
+  for (uint32_t i = 0; i < size; ++i) {
+    values.push_back(float(i));
+  }
+
+  const uint32_t batch_size = 1;
+  std::vector<sftensor> inputs;
+
+  sftensor input = std::make_shared<ftensor>(97);
+  input->Fill(values);
+
+  std::vector<sftensor> outputs(1);
+  inputs.push_back(input);
+  softmax_layer.Forward(inputs, outputs);
+  softmax1(values.data(), values.size());
+  for (int i = 0; i < size; ++i) {
+    ASSERT_EQ(input->index(i), float(i));
+    ASSERT_NEAR(values.at(i), outputs.front()->index(i), 1e-3f);
+  }
+}
+
 TEST(test_layer, forward_softmax_dim02) {
   using namespace kuiper_infer;
   SoftmaxLayer softmax_layer(0);