update

app.py

@@ -23,7 +23,15 @@ Besides, it can also used as user simulator.
 """
 
 survey = """\
+## knowledge distillation 知识蒸馏
 
+Essentially, it is a form of model compression.
+
+## distilling knowledge != knowledge distillation
+
+知识的形式可以是 QA纯文本，也可以是 QA+概率。
+
+## 有不用概率的知识蒸馏吗？
 """
 
 with gr.Blocks() as demo:

models/cpp_qwen2.py

@@ -58,7 +58,7 @@ class Qwen2Simulator(Simulator):
                 filename="*fp16.gguf",
                 n_ctx=config.MAX_SEQUENCE_LENGTH,
                 # use_mlock=True,
-                verbose=False,
+                verbose=True,
             )
         logger.info(f"llm has been initialized: {self.llm}, "
                     f"n_threads={self.llm.n_threads}, n_ctx={self.llm.n_ctx}, "
@@ -116,6 +116,9 @@ class Qwen2Simulator(Simulator):
             stream = copy.deepcopy(out)
             if stream["choices"][0]["finish_reason"] is None:
                 yield stream["choices"][0]["completion_text"], stream["choices"][0]["completion_tokens"]
+            else:
+                print(f'finish with text: {stream["choices"][0]["completion_text"]}, tokens: {stream["choices"][0]["completion_tokens"]}')
+
 
 
 bot = Qwen2Simulator()