Merge pull request #3130 from alibaba/feature/sync

MNN:Sync: Sync Interal 3.0.2

.gitignore

@@ -361,10 +361,3 @@ pymnn_build/
 
 # mnncompress generated
 MNN_compression_pb2.py
-
-# model path
-model/
-
-# datasets
-datasets/*
-!datasets/*.sh

3rd_party/OpenCLHeaders/CL/cl2.hpp

@@ -3810,6 +3810,24 @@ class Buffer : public Memory
         }
     }
 
+    Buffer(
+        const Context& context,
+        cl_mem_flags flags,
+        const cl_import_properties_arm *properties,
+        void *memory,
+        size_type size,
+        cl_int* err = NULL)
+    {
+        cl_int error;
+        object_ = ::clImportMemoryARM(context(), flags, properties, memory, size, &error);
+
+        detail::errHandler(error, __CREATE_BUFFER_ERR);
+        if (err != NULL) {
+            *err = error;
+        }
+    }
+
+
     /*!
      * \brief Construct a Buffer from a host container via iterators using a specified context.
      * IteratorType must be random access.

3rd_party/OpenCLHeaders/CL/cl_ext.h

@@ -430,6 +430,23 @@ typedef struct _cl_mem_android_native_buffer_host_ptr
 } cl_mem_android_native_buffer_host_ptr;
 
 
+/*********************************
+* cl_qcom_ahardwarebuffer_host_ptr extension
+*********************************/
+
+#define CL_MEM_ANDROID_AHARDWAREBUFFER_HOST_PTR_QCOM                0x4119
+
+typedef struct _cl_mem_ahardwarebuffer_host_ptr
+{
+    /* Type of external memory allocation. */
+    /* Must be CL_MEM_ANDROID_AHARDWAREBUFFER_HOST_PTR_QCOM for Android Hardware buffers. */
+    cl_mem_ext_host_ptr  ext_host_ptr;
+
+    /* Virtual pointer to the android hardware buffer */
+    void*                ahb_ptr;
+
+} cl_mem_ahardwarebuffer_host_ptr;
+
 /******************************************
  * cl_img_yuv_image extension *
  ******************************************/
@@ -583,6 +600,11 @@ typedef intptr_t cl_import_properties_arm;
 
 /* Protected DMA BUF memory type value for CL_IMPORT_TYPE_ARM property */
 #define CL_IMPORT_TYPE_PROTECTED_ARM              0x40B5
+#define CL_IMPORT_TYPE_ANDROID_HARDWARE_BUFFER_ARM          0x41E2
+#define CL_IMPORT_DMA_BUF_DATA_CONSISTENCY_WITH_HOST_ARM    0x41E3
+#define CL_IMPORT_MEMORY_WHOLE_ALLOCATION_ARM               SIZE_MAX
+#define CL_IMPORT_ANDROID_HARDWARE_BUFFER_PLANE_INDEX_ARM   0x41EF
+#define CL_IMPORT_ANDROID_HARDWARE_BUFFER_LAYER_INDEX_ARM   0x41F0
 
 /* This extension adds a new function that allows for direct memory import into
  * OpenCL via the clImportMemoryARM function.

CMakeLists.txt

@@ -20,9 +20,7 @@ endif()
 project(MNN VERSION ${MNN_VERSION} LANGUAGES C CXX ASM)
 # complier options
 set(CMAKE_C_STANDARD 99)
-IF (NOT (CMAKE_CXX_STANDARD EQUAL 17))
-  set(CMAKE_CXX_STANDARD 11)
-ENDIF()
+set(CMAKE_CXX_STANDARD 11)
 set(CMAKE_MODULE_PATH
   ${CMAKE_MODULE_PATH}
   "${CMAKE_CURRENT_LIST_DIR}/cmake"
@@ -49,7 +47,7 @@ option(MNN_BUILD_TOOLS "Build tools/cpp or not" ON)
 option(MNN_BUILD_QUANTOOLS "Build Quantized Tools or not" OFF)
 option(MNN_EVALUATION "Build Evaluation Tools or not" OFF)
 option(MNN_BUILD_CONVERTER "Build Converter" OFF)
-option(MNN_SUPPORT_DEPRECATED_OP "Enable MNN's tflite quantized op" ON)
+option(MNN_SUPPORT_DEPRECATED_OP "Enable MNN's tflite quantized op" OFF)
 option(MNN_DEBUG_MEMORY "MNN Debug Memory Access" OFF)
 option(MNN_DEBUG_TENSOR_SIZE "Enable Tensor Size" OFF)
 option(MNN_GPU_TRACE "Enable MNN Gpu Debug" OFF)
@@ -74,6 +72,7 @@ option(MNN_JNI "Build MNN Jni for java to use" OFF)
 option(MNN_SUPPORT_BF16 "Enable MNN's bf16 op" OFF)
 option(MNN_LOW_MEMORY "Build MNN support low memory for weight quant model." OFF)
 option(MNN_CPU_WEIGHT_DEQUANT_GEMM "Build MNN CPU weight dequant related gemm kernels." OFF)
+option(MNN_BUILD_AUDIO "Build audio api in MNN." OFF)
 
 IF (OHOS AND MNN_INTERNAL)
   include($ENV{NODE_PATH}/@ali/tcpkg/tcpkg.cmake)
@@ -192,6 +191,9 @@ endif()
 if(MNN_SUPPORT_TRANSFORMER_FUSE)
     add_definitions(-DMNN_SUPPORT_TRANSFORMER_FUSE)
 endif()
+if(MNN_BUILD_AUDIO)
+    add_definitions(-DMNN_BUILD_AUDIO)
+endif()
 # debug options
 if(MNN_DEBUG_MEMORY)
     add_definitions(-DMNN_DEBUG_MEMORY)
@@ -287,7 +289,7 @@ if(CMAKE_SYSTEM_NAME MATCHES "^Android")
 endif()
 option(MNN_USE_CPP11 "Enable MNN use c++11" ON)
 if (NOT MSVC)
-    if((MNN_CUDA AND MNN_SUPPORT_TRANSFORMER_FUSE) OR (CMAKE_CXX_STANDARD EQUAL 17))
+    if(MNN_CUDA AND MNN_SUPPORT_TRANSFORMER_FUSE)
         set(CMAKE_CXX_STANDARD 17)
         set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -std=gnu99")
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++17")
@@ -463,6 +465,10 @@ IF(MNN_BUILD_OPENCV)
   list(APPEND MNN_EXTRA_HEADERS ${MNN_CV_HDRS})
   list(APPEND MNN_EXTRA_HEADERS ${MNN_CV_IMGHDRS})
 ENDIF()
+IF(MNN_BUILD_AUDIO)
+  file(GLOB MNN_AUDIO_HDRS ${CMAKE_CURRENT_SOURCE_DIR}/tools/audio/include/audio/*.hpp PARENT_SCOPE)
+  list(APPEND MNN_EXTRA_HEADERS ${MNN_AUDIO_HDRS})
+ENDIF()
 IF(MNN_BUILD_LLM)
   file(GLOB MNN_LLM_HDRS ${CMAKE_CURRENT_SOURCE_DIR}/transformers/llm/engine/include/llm/*)
   list(APPEND MNN_EXTRA_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/transformers/llm/engine/include/llm/llm.hpp)
@@ -775,6 +781,14 @@ IF(MNN_BUILD_OPENCV AND NOT MNN_SEP_BUILD)
   ENDIF()
   target_sources(MNN PRIVATE $<TARGET_OBJECTS:MNNOpenCV>)
 ENDIF()
+add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/tools/audio)
+IF(MNN_BUILD_AUDIO AND NOT MNN_SEP_BUILD)
+  IF(MSVC)
+    target_compile_definitions(MNNAudio PRIVATE "-DBUILDING_MNN_DLL" INTERFACE "-DUSING_MNN_DLL")
+  ENDIF()
+  message(STATUC "### build MNNAudio into MNN")
+  target_sources(MNN PRIVATE $<TARGET_OBJECTS:MNNAudio>)
+ENDIF()
 
 
 if(CMAKE_SYSTEM_NAME MATCHES "^Linux")
@@ -884,6 +898,14 @@ ELSE()
       SET_SOURCE_FILES_PROPERTIES(${HDR} PROPERTIES MACOSX_PACKAGE_LOCATION Headers/cv/imgproc )
     ENDFOREACH()
   ENDIF()
+  IF(MNN_BUILD_AUDIO)
+    if (NOT MNN_AAPL_FMWK)
+      INSTALL(FILES ${MNN_AUDIO_HDRS} DESTINATION include/MNN/audio)
+    endif()
+    FOREACH(HDR ${MNN_AUDIO_HDRS})
+      SET_SOURCE_FILES_PROPERTIES(${HDR} PROPERTIES MACOSX_PACKAGE_LOCATION Headers/audio/ )
+    ENDFOREACH()
+  ENDIF()
   IF(MNN_BUILD_LLM)
     if (NOT MNN_AAPL_FMWK)
         INSTALL(FILES ${MNN_LLM_HDRS} DESTINATION include/MNN/llm)

docs/compile/cmake.md

@@ -16,7 +16,7 @@ MNN使用CMake构建项目，CMake中的宏定义列表如下：
 | MNN_BUILD_QUANTOOLS  | 是否构建MNN的量化工具，默认为`OFF` |
 | MNN_EVALUATION       | 是否构建MNN的评估工具，默认为`OFF` |
 | MNN_BUILD_CONVERTER  | 是否构建MNN的转换工具，默认为`OFF` |
-| MNN_SUPPORT_DEPRECATED_OP | 是否支持Tflite的量化算子，默认为`ON` |
+| MNN_SUPPORT_DEPRECATED_OP | 是否支持Tflite的量化算子等已经废弃的算子，用于兼容历史模型(1.1.0版本之前)，默认为`OFF` |
 | MNN_DEBUG_MEMORY     | 是否开启MNN内存调试，默认为`OFF` |
 | MNN_DEBUG_TENSOR_SIZE | 是否开启MNN tensor size调试，默认为`OFF` |
 | MNN_GPU_TRACE        | 是否开启MNN GPU调试，默认为`OFF` |
@@ -32,6 +32,7 @@ MNN使用CMake构建项目，CMake中的宏定义列表如下：
 | MNN_ENABLE_COVERAGE  | 是否开启MNN的代码覆盖率，默认为`OFF` |
 | MNN_BUILD_PROTOBUFFER | 是否使用MNN中的`protobuffer`，默认为`ON` |
 | MNN_BUILD_OPENCV     | 是否构建MNN的OpenCV功能，默认为`OFF` |
+| MNN_BUILD_AUDIO      | 是否构建MNN的Audio功能，默认为`OFF` |
 | MNN_INTERNAL         | 是否构建MNN的一些内部功能，如：日志；默认为`OFF` |
 | MNN_JNI              | 是否构建MNN的JNI支持，默认为`OFF` |
 | MNN_METAL            | 是否构建`Metal`后端，默认为`OFF` |
@@ -79,6 +80,7 @@ MNN使用CMake构建项目，CMake中的宏定义列表如下：
 | MNN_CVCORE           | 构建MNN的OpenCV功能是否开启`core`功能，默认为`ON` |
 | MNN_OPENCV_TEST      | 构建MNN的OpenCV功能是否开启单元测试，默认为`OFF` |
 | MNN_OPENCV_BENCH     | 构建MNN的OpenCV功能是否开启性能benchmark，默认为`OFF` |
+| MNN_AUDIO_TEST       | 构建MNN的Audio功能是否开启单元测试，默认为`OFF` |
 | MNN_VULKAN_IMAGE     | 构建MNN的Vulkan后端时采用Image内存模式，以便支持FP16和部分移动端上GPU的加速，默认为`ON` |
 | MNN_LOW_MEMORY       | 是否支持低内存模式，支持低内存模式使用权值量化模型并设置`low_memory`则会使用计算时反量化，默认为`OFF` |
 | MNN_CPU_WEIGHT_DEQUANT_GEMM       | 是否编译CPU权重反量化的矩阵乘Kernel， 如果打开该编译宏并且在CPU推理时设置MNN::BackendConfig::MemoryMode=Memory_Normal，就会使用权重反量化算子进行权重量化模型的推理，默认为`OFF` |

docs/compile/other.md

@@ -133,6 +133,19 @@
   - `libMNNOpenCV.so` MNN OpenCV函数库
   - `opencv_test` MNN OpenCV单元测试
   - `opencv_bench` MNN OpenCV性能测试
+## MNN Audio库
+- 相关编译选项
+  - `MNN_BUILD_AUDIO` 是否编译Audio函数接口
+  - `MNN_AUDIO_TEST` 是否编译Audio单元测试
+- 编译命令
+    ```bash
+    mkdir build && cd build
+    cmake .. -MNN_BUILD_AUDIO=ON -MNN_AUDIO_TEST=ON
+    make -j4
+    ```
+- 编译产物
+  - `libMNNAudio.so` MNN Audio函数库
+  - `audio_test` MNN Audio单元测试
 
 ## 示例工程
 - 相关编译选项

docs/transformers/llm.md

@@ -49,7 +49,7 @@ python llmexport.py \
 
 ### 功能
 - 直接转为mnn模型，使用`--export mnn`，注意，你需要先安装pymnn或者通过`--mnnconvert`选项指定MNNConvert工具的地址，两种条件必须满足其中一个。如果没有安装pymnn并且没有通过`--mnnconvert`指定MNNConvert工具的地址，那么llmexport.py脚本会在目录"../../../build/"下寻找MNNConvert工具，需保证该目录下存在MNNConvert文件。此方案目前支持导出4bit和8bit模型
-- 如果直接转为mnn模型遇到问题，或者需要其他bits数的量化（如5bit/6bit），可以先将模型先转为onnx模型，使用`--export onnx`，然后使用./MNNConvert工具将onnx模型转为mnn模型: 
+- 如果直接转为mnn模型遇到问题，或者需要其他bits数的量化（如5bit/6bit），可以先将模型先转为onnx模型，使用`--export onnx`，然后使用./MNNConvert工具将onnx模型转为mnn模型:
 
 ```
 ./MNNConvert --modelFile ../transformers/llm/export/model/onnx/llm.onnx --MNNModel llm.mnn --keepInputFormat --weightQuantBits=4 --weightQuantBlock=128 -f ONNX --transformerFuse=1 --allowCustomOp --saveExternalData
@@ -98,13 +98,17 @@ options:
 [从源码编译](../compile/other.html#id4)
 在原有编译过程中增加必需编译宏即可：
 ```
--DMNN_LOW_MEMORY=true -DMNN_CPU_WEIGHT_DEQUANT_GEMM=true -DMNN_BUILD_LLM=true -DMNN_SUPPORT_TRANSFORMER_FUSE=true 
+-DMNN_LOW_MEMORY=true -DMNN_CPU_WEIGHT_DEQUANT_GEMM=true -DMNN_BUILD_LLM=true -DMNN_SUPPORT_TRANSFORMER_FUSE=true
 ```
 
 - 需要开启视觉功能时，增加相关编译宏
 ```
 -DLLM_SUPPORT_VISION=true -DMNN_BUILD_OPENCV=true -DMNN_IMGCODECS=true
 ```
+- 需要开启音频功能时，增加相关编译宏
+```
+-DLLM_SUPPORT_AUDIO=true
+```
 
 #### mac / linux / windows
 
@@ -137,7 +141,7 @@ sh package_scripts/ios/buildiOS.sh "-DMNN_ARM82=true -DMNN_LOW_MEMORY=true -DMNN
 ```
 
 #### Web
-环境配置参考 https://mnn-docs.readthedocs.io/en/latest/compile/engine.html#web 
+环境配置参考 https://mnn-docs.readthedocs.io/en/latest/compile/engine.html#web
 
 - 编译库，产出 `libMNN.a`，`libMNN_Express.a`，`libllm.a`
 
@@ -189,7 +193,7 @@ node llm_demo.js ~/qwen2.0_1.5b/config.json ~/qwen2.0_1.5b/prompt.txt
   - visual_model: 当使用VL模型时，visual_model的实际路径为`base_dir + visual_model`，默认为`base_dir + 'visual.mnn'`
 - 推理配置
   - max_new_tokens: 生成时最大token数，默认为`512`
-  - reuse_kv: 多轮对话时是否复用之前对话的`kv cache`，默认为`false`, 目前只有CPU后端支持设置为`true`.
+  - reuse_kv: 多轮对话时是否复用之前对话的`kv cache`，默认为`false`
   - quant_qkv: CPU attention 算子中`query, key, value`是否量化，可选为：`0, 1, 2, 3, 4`，默认为`0`，含义如下：
     - 0: key和value都不量化
     - 1: 使用非对称8bit量化存储key
@@ -205,19 +209,6 @@ node llm_demo.js ~/qwen2.0_1.5b/config.json ~/qwen2.0_1.5b/prompt.txt
   - thread_num: CPU推理使用硬件线程数，默认为：`4`; OpenCL推理时使用`68`
   - precision: 推理使用精度策略，默认为：`"low"`，尽量使用`fp16`
   - memory: 推理使用内存策略，默认为：`"low"`，开启运行时量化
-- Sampler配置
-  - sampler_type: 使用的sampler种类，目前支持`greedy`, `temperature`, `topK`, `topP`, `minP`, `tfs`, `typical`, `penalty`8种基本sampler，外加`mixed`(混合sampler)。当选择`mixed`时，依次执行mixed_samplers中的sampler。默认为`mixed`。
-  - mixed_samplers: 当`sampler_type`为`mixed`时有效，默认为`["topK", "tfs", "typical", "topP", "min_p", "temperature"]`
-  - temperature: `temperature`, `topP`, `minP`, `tfsZ`, `typical`中temerature值，默认为1.0
-  - topK: `topK`中top K 个的个数，默认为40
-  - topP: `topP`中top P的值，默认为0.9
-  - minP: `minP`中min P的值，默认为0.1
-  - tfsZ: `tfs`中Z的值，默认为1.0，即不使用tfs算法
-  - typical: `typical`中p的值，默认为1.0，即不使用typical算法
-  - penalty: `penalty`中对于logits的惩罚项，默认为0.0，即不惩罚
-  - n_gram: `penalty`中最大存储的ngram大小，默认为8
-  - ngram_factor: `penalty`中对于重复ngram的额外惩罚，默认为1.0，即没有额外惩罚
-  - penalty_sampler: `penalty`中最后一步采用的sampling策略，可选"greedy"或"temperature"，默认greedy.
 
 ##### 配置文件示例
 - `config.json`
@@ -229,15 +220,7 @@ node llm_demo.js ~/qwen2.0_1.5b/config.json ~/qwen2.0_1.5b/prompt.txt
       "backend_type": "cpu",
       "thread_num": 4,
       "precision": "low",
-      "memory": "low",
-      "sampler_type": "mixed",
-      "mixed_samplers": ["topK", "tfs", "typical", "topP", "min_p", "temperature"],
-      "temperature": 1.0,
-      "topK": 40,
-      "topP": 0.9,
-      "tfsZ": 1.0,
-      "minP": 0.1,
-      "reuse_kv": true
+      "memory": "low"
   }
   ```
 - `llm_config.json`
@@ -261,8 +244,7 @@ node llm_demo.js ~/qwen2.0_1.5b/config.json ~/qwen2.0_1.5b/prompt.txt
 
 #### 推理用法
 `llm_demo`的用法如下：
-pc端直接推理
-```bash
+```
 # 使用config.json
 ## 交互式聊天
 ./llm_demo model_dir/config.json
@@ -276,16 +258,15 @@ pc端直接推理
 ./llm_demo model_dir/llm.mnn prompt.txt
 ```
 
-android手机端adb推理用法：
-```bash
-# 利用adb push将链接库push到手机上
-adb shell mkdir /data/local/tmp/llm
-adb push llm_demo ppl_demo libllm.so libMNN_CL.so libMNN_Express.so libMNN.so tools/cv/libMNNOpenCV.so /data/local/tmp/llm
-```
-
 - 对于视觉大模型，在prompt中嵌入图片输入
 ```
 <img>https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg</img>介绍一下图片里的内容
+# 指定图片大小
+<img><hw>280, 420</hw>https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg</img>介绍一下图片里的内容
+```
+- 对于音频大模型，在prompt中嵌入音频输入
+```
+<audio>https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/translate_to_chinese.wav</audio>介绍一下音频里的内容
 ```
 
 #### GPTQ权重加载

docs/transformers/models.md

@@ -47,4 +47,5 @@
 | [reader-lm-0.5b](https://huggingface.co/jinaai/reader-lm-0.5b) | [Q4_1](https://modelscope.cn/models/MNN/reader-lm-0.5b-MNN) | [Q4_1](https://huggingface.co/taobao-mnn/reader-lm-0.5b-MNN) |
 | [reader-lm-1.5b](https://huggingface.co/jinaai/reader-lm-1.5b) | [Q4_1](https://modelscope.cn/models/MNN/reader-lm-1.5b-MNN) | [Q4_1](https://huggingface.co/taobao-mnn/reader-lm-1.5b-MNN) |
 | [TinyLlama-1.1B-Chat-v1.0](https://modelscope.cn/models/AI-ModelScope/TinyLlama-1.1B-Chat-v1.0/summary) | [Q4_1](https://modelscope.cn/models/MNN/TinyLlama-1.1B-Chat-MNN) | [Q4_1](https://huggingface.co/taobao-mnn/TinyLlama-1.1B-Chat-MNN) |
-| [Yi-6B-Chat](https://modelscope.cn/models/01ai/Yi-6B-Chat/summary) | [Q4_1](https://modelscope.cn/models/MNN/Yi-6B-Chat-MNN) | [Q4_1](https://huggingface.co/taobao-mnn/Yi-6B-Chat-MNN) |
+| [Yi-6B-Chat](https://modelscope.cn/models/01ai/Yi-6B-Chat/summary) | [Q4_1](https://modelscope.cn/models/MNN/Yi-6B-Chat-MNN) | [Q4_1](https://huggingface.co/taobao-mnn/Yi-6B-Chat-MNN) |
+| [QwQ-32B-Preview](https://modelscope.cn/models/Qwen/QwQ-32B-Preview/summary) | [Q4_1](https://modelscope.cn/models/MNN/QwQ-32B-Preview-MNN) | [Q4_1](https://huggingface.co/taobao-mnn/QwQ-32B-Preview-MNN) |

express/MathOp.cpp

@@ -1208,7 +1208,7 @@ VARP _LinSpace(VARP start, VARP stop, VARP num) {
     return (Variable::create(Expr::create(std::move(op), {start, stop, num})));
 }
 
-VARP _EltwiseProdInt8(VARP x, VARP y, 
+VARP _EltwiseProdInt8(VARP x, VARP y,
                     std::vector<int8_t> x_weight, std::vector<int32_t> x_bias, std::vector<float> x_scale, std::vector<float> x_tensorScale,
                     std::vector<int8_t> y_weight, std::vector<int32_t> y_bias, std::vector<float> y_scale, std::vector<float> y_tensorScale,
                     std::vector<int8_t> output_weight, std::vector<int32_t> output_bias, std::vector<float> output_scale, std::vector<float> output_tensorScale)
@@ -1219,7 +1219,7 @@ VARP _EltwiseProdInt8(VARP x, VARP y,
                         output_weight, output_bias, output_scale, output_tensorScale);
 }
 
-VARP _EltwiseSumInt8(VARP x, VARP y, 
+VARP _EltwiseSumInt8(VARP x, VARP y,
                     std::vector<int8_t> x_weight, std::vector<int32_t> x_bias, std::vector<float> x_scale, std::vector<float> x_tensorScale,
                     std::vector<int8_t> y_weight, std::vector<int32_t> y_bias, std::vector<float> y_scale, std::vector<float> y_tensorScale,
                     std::vector<int8_t> output_weight, std::vector<int32_t> output_bias, std::vector<float> output_scale, std::vector<float> output_tensorScale)
@@ -1230,7 +1230,7 @@ VARP _EltwiseSumInt8(VARP x, VARP y,
                         output_weight, output_bias, output_scale, output_tensorScale);
 }
 
-VARP _EltwiseSubInt8(VARP x, VARP y, 
+VARP _EltwiseSubInt8(VARP x, VARP y,
                     std::vector<int8_t> x_weight, std::vector<int32_t> x_bias, std::vector<float> x_scale, std::vector<float> x_tensorScale,
                     std::vector<int8_t> y_weight, std::vector<int32_t> y_bias, std::vector<float> y_scale, std::vector<float> y_tensorScale,
                     std::vector<int8_t> output_weight, std::vector<int32_t> output_bias, std::vector<float> output_scale, std::vector<float> output_tensorScale)
@@ -1241,7 +1241,7 @@ VARP _EltwiseSubInt8(VARP x, VARP y,
                         output_weight, output_bias, output_scale, output_tensorScale);
 }
 
-VARP _EltwiseMaxInt8(VARP x, VARP y, 
+VARP _EltwiseMaxInt8(VARP x, VARP y,
                     std::vector<int8_t> x_weight, std::vector<int32_t> x_bias, std::vector<float> x_scale, std::vector<float> x_tensorScale,
                     std::vector<int8_t> y_weight, std::vector<int32_t> y_bias, std::vector<float> y_scale, std::vector<float> y_tensorScale,
                     std::vector<int8_t> output_weight, std::vector<int32_t> output_bias, std::vector<float> output_scale, std::vector<float> output_tensorScale)
@@ -1320,5 +1320,20 @@ VARP _Histogram(VARP x, int bin, int min, int max, int channel) {
     return (Variable::create(Expr::create(std::move(op), {x})));
 }
 
+#ifdef MNN_BUILD_AUDIO
+VARP _Stft(VARP sample, VARP window, int n_fft, int hop_length, bool abs) {
+    std::unique_ptr<OpT> op(new OpT);
+    op->type      = OpType_Stft;
+    op->main.type = OpParameter_StftParam;
+    auto param = new StftParamT;
+    param->n_fft = n_fft;
+    param->hop_length = hop_length;
+    param->abs = abs;
+    op->main.value = param;
+    EXPRP expr = Expr::create(std::move(op), {sample, window});
+    return Variable::create(expr);
+}
+#endif
+
 } // namespace Express
 } // namespace MNN