[LLM:Bugfix] Bugfix of Omni audio.

docs/transformers/llm.md

@@ -487,7 +487,9 @@ python llmexport.py --path /path/to/Qwen2.5-0.5B-Instruct --lora_path /path/to/l
   ```
 
 #### 获取语音输出
-使用Omni模型时，可以使用接口`setWavformCallback`获取语音输出，示例如下：
+使用Omni模型时，可以使用接口`setWavformCallback`获取语音输出，使用接口`generateWavform`开始输出语音。
+注意`setWavformCallback`需要在文本生成前调用， `generateWavform`在文本生成结束后调用，示例如下：
+
 1. 保存语音到文件中
 ```cpp
 #include <audio/audio.hpp>
@@ -504,6 +506,9 @@ int main() {
       }
       return true;
   });
+  llm->response("Hello");
+  // generate wavform
+  llm->generateWavform();
   return 0;
 }
 
@@ -600,12 +605,15 @@ bool AudioPlayer::play(const float* ptr, size_t size, bool last_chunk) {
 }
 
 int main() {
-  //....
-  AudioPlayer audio_player;
+    //....
+    AudioPlayer audio_player;
     llm->setWavformCallback([&](const float* ptr, size_t size, bool last_chunk) {
         return audio_player.play(ptr, size, last_chunk);
     });
-  //....
-  return 0;
+    //....
+    llm->response("Hello");
+    // generate wavform
+    llm->generateWavform();
+    return 0;
 }
 ```

transformers/llm/engine/include/llm/llm.hpp

@@ -112,6 +112,7 @@ public:
         return mContext.get();
     }
     virtual void setWavformCallback(std::function<bool(const float*, size_t, bool)> callback) {}
+    virtual void generateWavform() {}
 protected:
     void initRuntime();
     std::shared_ptr<LlmContext> mContext;

transformers/llm/engine/src/omni.cpp

@@ -299,7 +299,7 @@ std::vector<int> Omni::audioProcess(const std::string& file) {
     auto input_features  = MNN::AUDIO::whisper_fbank(waveform);
     VARP audio_embedding;
     if (mAudioModule->getInfo()->inputNames.size() > 1) {
-        int seqlen = input_features->getInfo()->dim[2] / 2;
+        int seqlen = UP_DIV(input_features->getInfo()->dim[2], 2);
         constexpr int n_window = 100;
         std::vector<int> cu_seqlens;
         int curseq = 0;
@@ -569,6 +569,15 @@ void Omni::response(const std::vector<int>& input_ids, std::ostream* os, const c
         mTalker->generate_init();
     }
     generate(input_ids, max_new_tokens);
+}
+
+void Omni::setWavformCallback(std::function<bool(const float*, size_t, bool)> callback) {
+    if (mTalker) {
+        mTalker->setWavformCallback(callback);
+    }
+}
+
+void Omni::generateWavform() {
     if (mTalker) {
         mTalker->generate();
 #ifdef DUMP_TALKER_PERFORMANCE
@@ -599,12 +608,6 @@ void Omni::response(const std::vector<int>& input_ids, std::ostream* os, const c
     }
 }
 
-void Omni::setWavformCallback(std::function<bool(const float*, size_t, bool)> callback) {
-    if (mTalker) {
-        mTalker->setWavformCallback(callback);
-    }
-}
-
 void Talker::load() {
     MNN::BackendConfig backendConfig;
     auto executor = MNN::Express::Executor::newExecutor(MNN_FORWARD_CPU, backendConfig, 1);

transformers/llm/engine/src/omni.hpp

@@ -110,6 +110,7 @@ public:
     virtual Express::VARP gen_position_ids(int seq_len) override;
     virtual void response(const std::vector<int>& input_ids, std::ostream* os = &std::cout, const char* end_with = nullptr, int max_new_tokens = -1) override;
     virtual void setWavformCallback(std::function<bool(const float*, size_t, bool)> callback) override;
+    virtual void generateWavform() override;
 private:
     int mVisionHeight = 448, mVisionWidth = 448, mVisionStart = 151857,
         mVisionEnd = 151858, mVisionPad = 151859, mAudioPad = 151646;