Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
@@ -4,6 +4,7 @@ import torch
|
|
4 |
from huggingface_hub import login
|
5 |
import os
|
6 |
|
|
|
7 |
def setup_llama3_auth():
|
8 |
"""Configurar autenticación para Llama 3"""
|
9 |
if 'HUGGING_FACE_TOKEN_3' in st.secrets:
|
@@ -17,14 +18,17 @@ def setup_llama3_auth():
|
|
17 |
|
18 |
class Llama3Demo:
|
19 |
def __init__(self):
|
20 |
-
# Verificar autenticación antes de cargar el modelo
|
21 |
setup_llama3_auth()
|
22 |
-
|
23 |
-
# Usando el modelo de 3B con instrucciones
|
24 |
self.model_name = "meta-llama/Llama-3.2-3B-Instruct"
|
25 |
self._model = None
|
26 |
self._tokenizer = None
|
27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
@property
|
29 |
def model(self):
|
30 |
if self._model is None:
|
@@ -33,12 +37,11 @@ class Llama3Demo:
|
|
33 |
self.model_name,
|
34 |
torch_dtype=torch.float16,
|
35 |
device_map="auto",
|
36 |
-
|
37 |
-
|
38 |
)
|
39 |
except Exception as e:
|
40 |
st.error(f"Error cargando el modelo: {str(e)}")
|
41 |
-
st.error("Verifica tu acceso a Llama 3.2 en https://huggingface.co/meta-llama")
|
42 |
raise e
|
43 |
return self._model
|
44 |
|
@@ -48,21 +51,26 @@ class Llama3Demo:
|
|
48 |
try:
|
49 |
self._tokenizer = AutoTokenizer.from_pretrained(
|
50 |
self.model_name,
|
51 |
-
|
52 |
)
|
53 |
except Exception as e:
|
54 |
st.error(f"Error cargando el tokenizer: {str(e)}")
|
55 |
raise e
|
56 |
return self._tokenizer
|
57 |
|
|
|
|
|
58 |
def generate_response(self, prompt: str, max_new_tokens: int = 512) -> str:
|
59 |
-
# Formato específico para Llama 3.2
|
60 |
formatted_prompt = f"""<|system|>You are a helpful AI assistant.</s>
|
61 |
<|user|>{prompt}</s>
|
62 |
<|assistant|>"""
|
63 |
|
64 |
inputs = self.tokenizer(formatted_prompt, return_tensors="pt").to(self.model.device)
|
65 |
|
|
|
|
|
|
|
|
|
66 |
with torch.no_grad():
|
67 |
outputs = self.model.generate(
|
68 |
**inputs,
|
@@ -70,16 +78,16 @@ class Llama3Demo:
|
|
70 |
num_return_sequences=1,
|
71 |
temperature=0.7,
|
72 |
do_sample=True,
|
73 |
-
top_p=0.9
|
|
|
74 |
)
|
75 |
|
76 |
-
# Limpiar memoria GPU
|
77 |
torch.cuda.empty_cache()
|
78 |
|
79 |
response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
|
80 |
-
# Extraer solo la respuesta del asistente
|
81 |
return response.split("<|assistant|>")[-1].strip()
|
82 |
|
|
|
83 |
def main():
|
84 |
st.set_page_config(page_title="Llama 3.2 Chat", page_icon="🦙")
|
85 |
|
|
|
4 |
from huggingface_hub import login
|
5 |
import os
|
6 |
|
7 |
+
##################################################################
|
8 |
def setup_llama3_auth():
|
9 |
"""Configurar autenticación para Llama 3"""
|
10 |
if 'HUGGING_FACE_TOKEN_3' in st.secrets:
|
|
|
18 |
|
19 |
class Llama3Demo:
|
20 |
def __init__(self):
|
|
|
21 |
setup_llama3_auth()
|
|
|
|
|
22 |
self.model_name = "meta-llama/Llama-3.2-3B-Instruct"
|
23 |
self._model = None
|
24 |
self._tokenizer = None
|
25 |
|
26 |
+
# Configuración de cuantización
|
27 |
+
self.quantization_config = BitsAndBytesConfig(
|
28 |
+
load_in_8bit=True,
|
29 |
+
bnb_4bit_compute_dtype=torch.float16
|
30 |
+
)
|
31 |
+
|
32 |
@property
|
33 |
def model(self):
|
34 |
if self._model is None:
|
|
|
37 |
self.model_name,
|
38 |
torch_dtype=torch.float16,
|
39 |
device_map="auto",
|
40 |
+
quantization_config=self.quantization_config, # Nueva forma de configurar cuantización
|
41 |
+
token=st.secrets['HUGGING_FACE_TOKEN_3'] # Actualizado de use_auth_token a token
|
42 |
)
|
43 |
except Exception as e:
|
44 |
st.error(f"Error cargando el modelo: {str(e)}")
|
|
|
45 |
raise e
|
46 |
return self._model
|
47 |
|
|
|
51 |
try:
|
52 |
self._tokenizer = AutoTokenizer.from_pretrained(
|
53 |
self.model_name,
|
54 |
+
token=st.secrets['HUGGING_FACE_TOKEN_3'] # Actualizado de use_auth_token a token
|
55 |
)
|
56 |
except Exception as e:
|
57 |
st.error(f"Error cargando el tokenizer: {str(e)}")
|
58 |
raise e
|
59 |
return self._tokenizer
|
60 |
|
61 |
+
|
62 |
+
##################################################################
|
63 |
def generate_response(self, prompt: str, max_new_tokens: int = 512) -> str:
|
|
|
64 |
formatted_prompt = f"""<|system|>You are a helpful AI assistant.</s>
|
65 |
<|user|>{prompt}</s>
|
66 |
<|assistant|>"""
|
67 |
|
68 |
inputs = self.tokenizer(formatted_prompt, return_tensors="pt").to(self.model.device)
|
69 |
|
70 |
+
# Asegurar que tenemos un pad_token_id válido
|
71 |
+
if self.tokenizer.pad_token_id is None:
|
72 |
+
self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
|
73 |
+
|
74 |
with torch.no_grad():
|
75 |
outputs = self.model.generate(
|
76 |
**inputs,
|
|
|
78 |
num_return_sequences=1,
|
79 |
temperature=0.7,
|
80 |
do_sample=True,
|
81 |
+
top_p=0.9,
|
82 |
+
pad_token_id=self.tokenizer.pad_token_id # Explícitamente establecer pad_token_id
|
83 |
)
|
84 |
|
|
|
85 |
torch.cuda.empty_cache()
|
86 |
|
87 |
response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
|
|
|
88 |
return response.split("<|assistant|>")[-1].strip()
|
89 |
|
90 |
+
##################################################################
|
91 |
def main():
|
92 |
st.set_page_config(page_title="Llama 3.2 Chat", page_icon="🦙")
|
93 |
|