Commit ·
81ff557
1
Parent(s): f2527c6
Se volvio a el approach de usar transformers
Browse files- handler.py +46 -118
- requirements.txt +1 -1
handler.py
CHANGED
|
@@ -1,125 +1,53 @@
|
|
| 1 |
-
from transformers import pipeline
|
| 2 |
-
import torch
|
| 3 |
-
import base64
|
| 4 |
from typing import Dict, List, Any
|
| 5 |
-
import
|
| 6 |
-
import
|
| 7 |
-
import
|
| 8 |
-
import tempfile
|
| 9 |
-
import numpy as np
|
| 10 |
-
|
| 11 |
-
# Nombre del modelo (usado como fallback si 'path' no se proporciona)
|
| 12 |
-
MODEL_NAME = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
|
| 13 |
|
| 14 |
class EndpointHandler():
|
| 15 |
-
def __init__(self, path=""):
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
"torch_dtype": torch.bfloat16 if torch.cuda.is_available() else None,
|
| 21 |
-
"enable_audio_output": True # Clave esencial para cargar el componente Talker (generador de voz) [4]
|
| 22 |
-
}
|
| 23 |
-
|
| 24 |
-
# 2. Carga del pipeline genérico de generación de texto (el wrapper para LLM multimodales) [3]
|
| 25 |
-
self.pipeline = pipeline(
|
| 26 |
-
task="text-generation",
|
| 27 |
-
model=path or MODEL_NAME,
|
| 28 |
-
**model_kwargs # Inyección de los parámetros específicos de Qwen3
|
| 29 |
-
)
|
| 30 |
-
|
| 31 |
-
# 3. System prompt obligatorio para Qwen3-Omni para generar audio natural [4]
|
| 32 |
-
self.system_prompt = (
|
| 33 |
-
"You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, "
|
| 34 |
-
"capable of perceiving auditory and visual inputs, as well as generating text and speech."
|
| 35 |
)
|
| 36 |
-
|
| 37 |
-
# 4. Tasa de muestreo del modelo (necesaria para la serialización de audio en __call__)
|
| 38 |
-
self.sampling_rate = getattr(self.pipeline.model.config, 'sampling_rate', 24000)
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
def _handle_audio_input(self, data: Dict[str, Any]) -> str:
|
| 42 |
-
""" Decodifica la entrada de audio Base64 y la guarda temporalmente como un archivo WAV. """
|
| 43 |
-
audio_data_base64 = data.get("audio_data")
|
| 44 |
-
if not audio_data_base64:
|
| 45 |
-
return None
|
| 46 |
-
|
| 47 |
-
temp_file_path = None
|
| 48 |
-
try:
|
| 49 |
-
audio_bytes = base64.b64decode(audio_data_base64)
|
| 50 |
-
# Guardar en un archivo temporal para que el pipeline lo pueda procesar [5]
|
| 51 |
-
temp_file = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
|
| 52 |
-
temp_file.write(audio_bytes)
|
| 53 |
-
temp_file.close()
|
| 54 |
-
temp_file_path = temp_file.name
|
| 55 |
-
return temp_file_path
|
| 56 |
-
except Exception as e:
|
| 57 |
-
if temp_file_path and os.path.exists(temp_file_path):
|
| 58 |
-
os.remove(temp_file_path)
|
| 59 |
-
raise ValueError(f"Error al decodificar y guardar el audio Base64: {e}")
|
| 60 |
-
|
| 61 |
-
def _handle_audio_output(self, generated_audio: torch.Tensor, sampling_rate: int) -> str:
|
| 62 |
-
""" Convierte el tensor de audio de salida a un buffer WAV y lo codifica en Base64. """
|
| 63 |
-
audio_array = generated_audio.cpu().numpy().squeeze()
|
| 64 |
-
if audio_array.dtype!= np.float32:
|
| 65 |
-
audio_array = audio_array.astype(np.float32)
|
| 66 |
-
|
| 67 |
-
with io.BytesIO() as buffer:
|
| 68 |
-
# Escribir el array como WAV [2]
|
| 69 |
-
wavfile.write(buffer, rate=sampling_rate, data=audio_array)
|
| 70 |
-
buffer.seek(0)
|
| 71 |
-
|
| 72 |
-
# Codificar a Base64 para la respuesta JSON
|
| 73 |
-
encoded_audio = base64.b64encode(buffer.read()).decode('utf-8')
|
| 74 |
-
return encoded_audio
|
| 75 |
|
| 76 |
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
"max_new_tokens": generation_kwargs.get("max_new_tokens", 512),
|
| 98 |
-
})
|
| 99 |
-
|
| 100 |
-
# 4. Ejecutar el pipeline
|
| 101 |
-
raw_output = self.pipeline(inputs_list, **generation_kwargs)
|
| 102 |
-
|
| 103 |
-
# El pipeline devuelve una lista de diccionarios, extraemos el primer resultado
|
| 104 |
-
response = raw_output
|
| 105 |
-
|
| 106 |
-
final_response = {
|
| 107 |
-
"generated_text": response.get("generated_text"),
|
| 108 |
-
"audio_output": None
|
| 109 |
-
}
|
| 110 |
-
|
| 111 |
-
# 5. Post-procesamiento (Tensor -> Base64-WAV)
|
| 112 |
-
if "audio_array" in response:
|
| 113 |
-
encoded_audio = self._handle_audio_output(response["audio_array"], self.sampling_rate)
|
| 114 |
-
final_response["audio_output"] = encoded_audio
|
| 115 |
-
|
| 116 |
-
return [final_response]
|
| 117 |
-
|
| 118 |
-
except Exception as e:
|
| 119 |
-
# Manejo de errores
|
| 120 |
-
return [{"error": str(e)}]
|
| 121 |
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
from typing import Dict, List, Any
|
| 2 |
+
import soundfile as sf
|
| 3 |
+
from transformers import Qwen3OmniMoeForConditionalGeneration, Qwen3OmniMoeProcessor
|
| 4 |
+
from qwen_omni_utils import process_mm_info
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
class EndpointHandler():
|
| 7 |
+
def __init__(self, path="./"):
|
| 8 |
+
self.model = Qwen3OmniMoeForConditionalGeneration.from_pretrained(
|
| 9 |
+
path,
|
| 10 |
+
dtype="auto",
|
| 11 |
+
device_map="auto",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
)
|
| 13 |
+
self.processor = Qwen3OmniMoeProcessor.from_pretrained(path)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
| 16 |
+
messages = data.get("messages", [])
|
| 17 |
+
use_audio_in_video = data.get("use_audio_in_video", True)
|
| 18 |
+
speaker = data.get("speaker", "Ethan")
|
| 19 |
+
|
| 20 |
+
text = self.processor.apply_chat_template(
|
| 21 |
+
messages,
|
| 22 |
+
tokenize=False,
|
| 23 |
+
add_generation_prompt=True,
|
| 24 |
+
)
|
| 25 |
+
audios, images, videos = process_mm_info(messages, use_audio_in_video=use_audio_in_video)
|
| 26 |
+
inputs = self.processor(
|
| 27 |
+
text=text,
|
| 28 |
+
audio=audios,
|
| 29 |
+
images=images,
|
| 30 |
+
videos=videos,
|
| 31 |
+
return_tensors="pt",
|
| 32 |
+
padding=True,
|
| 33 |
+
use_audio_in_video=use_audio_in_video
|
| 34 |
+
)
|
| 35 |
+
inputs = inputs.to(self.model.device).to(self.model.dtype)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
+
text_ids, audio = self.model.generate(
|
| 38 |
+
**inputs,
|
| 39 |
+
speaker=speaker,
|
| 40 |
+
thinker_return_dict_in_generate=True,
|
| 41 |
+
use_audio_in_video=use_audio_in_video
|
| 42 |
+
)
|
| 43 |
+
text_output = self.processor.batch_decode(
|
| 44 |
+
text_ids.sequences[:, inputs["input_ids"].shape[1]:],
|
| 45 |
+
skip_special_tokens=True,
|
| 46 |
+
clean_up_tokenization_spaces=False
|
| 47 |
+
)
|
| 48 |
+
result = {"generated_text": text_output}
|
| 49 |
+
if audio is not None:
|
| 50 |
+
# Guarda el audio en un archivo temporal y retorna la ruta
|
| 51 |
+
sf.write("output.wav", audio.reshape(-1).detach().cpu().numpy(), samplerate=24000)
|
| 52 |
+
result["audio_path"] = "output.wav"
|
| 53 |
+
return [result]
|
requirements.txt
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
soundfile
|
| 2 |
-
transformers
|
| 3 |
torch
|
| 4 |
qwen-omni-utils
|
| 5 |
torchvision
|
|
|
|
| 1 |
soundfile
|
| 2 |
+
transformers
|
| 3 |
torch
|
| 4 |
qwen-omni-utils
|
| 5 |
torchvision
|