Update app.py

app.py

@@ -9,7 +9,8 @@ from transformers import (
     AutoModelForCausalLM,
     AutoTokenizer,
     GenerationConfig,
-    StoppingCriteriaList
+    StoppingCriteriaList,
+    TextIteratorStreamer # Importar TextIteratorStreamer
 )
 import uvicorn
 import asyncio
@@ -120,44 +121,27 @@ async def stream_text(model, tokenizer, input_text, generation_config, stop_sequ
 
     stopping_criteria = StoppingCriteriaList([stop_criteria])
 
-    output_text = ""
-    outputs = model.generate(
+    streamer = TextIteratorStreamer(tokenizer, chunk_delay=chunk_delay, skip_prompt=True) # Inicializar streamer
+
+    generation_kwargs = dict(
         **encoded_input,
-        do_sample=generation_config.do_sample,
-        max_new_tokens=generation_config.max_new_tokens,
-        temperature=generation_config.temperature,
-        top_p=generation_config.top_p,
-        top_k=generation_config.top_k,
-        repetition_penalty=generation_config.repetition_penalty,
-        num_return_sequences=generation_config.num_return_sequences,
+        generation_config=generation_config,
         stopping_criteria=stopping_criteria,
-        output_scores=True,
-        return_dict_in_generate=True
+        streamer=streamer, # Pasar streamer a generate
+        return_dict_in_generate=True,
+        output_scores=True
     )
 
-    for output in outputs.sequences:
-        for token_id in output:
-            token = tokenizer.decode(token_id, skip_special_tokens=True)
-            yield token
-            await asyncio.sleep(chunk_delay)
+    async def generate_task():
+        model.generate(**generation_kwargs) # Ejecutar generate en background
+
+    asyncio.create_task(generate_task()) # Iniciar la tarea de generación
 
-        if stop_sequences and any(stop in output_text for stop in stop_sequences):
-            yield output_text
+    for token in streamer: # Iterar sobre el streamer para obtener tokens uno por uno
+        yield token
+        if stop_sequences and any(stop in token for stop in stop_sequences): # Comprobar stop sequences en cada token
             return
 
-        outputs = model.generate(
-            **encoded_input,
-            do_sample=generation_config.do_sample,
-            max_new_tokens=generation_config.max_new_tokens,
-            temperature=generation_config.temperature,
-            top_p=generation_config.top_p,
-            top_k=generation_config.top_k,
-            repetition_penalty=generation_config.repetition_penalty,
-            num_return_sequences=generation_config.num_return_sequences,
-            stopping_criteria=stopping_criteria,
-            output_scores=True,
-            return_dict_in_generate=True
-        )
 
 @app.post("/generate-image")
 async def generate_image(request: GenerateRequest):