Update api.py

api.py

@@ -135,7 +135,6 @@ def generate_text_mirostat_top_p(model, prompt, max_len=100, max_gen=98,
     model_input = text_to_ids(f"<start> {prompt} <sep>")
     model_input = model_input[:max_len]
     generated = list(model_input)
-    text_so_far = []
 
     tau = 5.0  # 초기 목표 surprise
 
@@ -193,16 +192,14 @@ def generate_text_mirostat_top_p(model, prompt, max_len=100, max_gen=98,
         generated.append(int(final_token))
 
         next_word = sp.id_to_piece(int(final_token))
-        text_so_far.append(next_word)
-        decoded_text = decode_sp_tokens(text_so_far)
+        decoded_piece = decode_sp_tokens([next_word])  # << 요것만 바뀐 부분!
+        yield decoded_piece  # 누적 텍스트가 아니라, 새로 생성된 토큰만 출력!
 
         if len(generated) >= min_len and final_token == end_id:
             break
-        if len(generated) >= min_len and decoded_text.endswith(('.', '!', '?', '<end>')):
+        if len(generated) >= min_len and decoded_piece.endswith(('.', '!', '?', '<end>')):
             break
 
-        yield decoded_text
-
 async def async_generator_wrapper(prompt: str):
     # 동기 제너레이터를 비동기로 감싸기
     loop = asyncio.get_event_loop()