Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -2,43 +2,21 @@ import gradio as gr
|
|
2 |
import torch
|
3 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
4 |
from peft import PeftModel
|
5 |
-
# !python -c "import torch; assert torch.cuda.get_device_capability()[0] >= 8, 'Hardware not supported for Flash Attention'"
|
6 |
import json
|
7 |
-
import torch
|
8 |
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, GemmaTokenizer, StoppingCriteria, StoppingCriteriaList, GenerationConfig
|
9 |
import os
|
10 |
|
11 |
-
#sft_model = "somosnlp/gemma-FULL-RAC-Colombia_v2"
|
12 |
#sft_model = "somosnlp/RecetasDeLaAbuela_mistral-7b-instruct-v0.2-bnb-4bit"
|
13 |
#base_model_name = "unsloth/Mistral-7B-Instruct-v0.2"
|
14 |
-
|
15 |
-
sft_model2 = "somosnlp/RecetasDeLaAbuela_mistral-7b-instruct-v0.2-bnb-4bit"
|
16 |
base_model_name = "unsloth/gemma-2b-it-bnb-4bit"
|
17 |
|
18 |
-
bnb_config = BitsAndBytesConfig(
|
19 |
-
load_in_4bit=True,
|
20 |
-
bnb_4bit_quant_type="nf4",
|
21 |
-
bnb_4bit_compute_dtype=torch.bfloat16
|
22 |
-
)
|
23 |
max_seq_length=400
|
24 |
-
|
25 |
-
# if torch.cuda.get_device_capability()[0] >= 8:
|
26 |
-
# # print("Flash Attention")
|
27 |
-
# attn_implementation="flash_attention_2"
|
28 |
-
# else:
|
29 |
-
# attn_implementation=None
|
30 |
-
attn_implementation=None
|
31 |
-
|
32 |
-
#base_model = AutoModelForCausalLM.from_pretrained(model_name,return_dict=True,torch_dtype=torch.float16,)
|
33 |
base_model = AutoModelForCausalLM.from_pretrained(base_model_name,return_dict=True,device_map="auto", torch_dtype=torch.float16,)
|
34 |
-
#base_model = AutoModelForCausalLM.from_pretrained(base_model_name, return_dict=True, device_map = {"":0}, attn_implementation = attn_implementation,).eval()
|
35 |
-
|
36 |
tokenizer = AutoTokenizer.from_pretrained(base_model_name, max_length = max_seq_length)
|
37 |
-
sft_model = sft_model1
|
38 |
ft_model = PeftModel.from_pretrained(base_model, sft_model)
|
39 |
model = ft_model.merge_and_unload()
|
40 |
model.save_pretrained(".")
|
41 |
-
#model.to('cuda')
|
42 |
tokenizer.save_pretrained(".")
|
43 |
|
44 |
class ListOfTokensStoppingCriteria(StoppingCriteria):
|
@@ -68,15 +46,9 @@ stopping_criteria = ListOfTokensStoppingCriteria(tokenizer, stop_tokens)
|
|
68 |
# Añade tu criterio de parada a una StoppingCriteriaList
|
69 |
stopping_criteria_list = StoppingCriteriaList([stopping_criteria])
|
70 |
|
71 |
-
def generate_text(
|
72 |
-
print('Modelo es: '+modelin)
|
73 |
-
#sft_model = modelin
|
74 |
-
#ft_model = PeftModel.from_pretrained(base_model, sft_model)
|
75 |
-
#model = ft_model.merge_and_unload()
|
76 |
-
|
77 |
prompt=prompt.replace("\n", "").replace("¿","").replace("?","")
|
78 |
-
|
79 |
-
input_text = str(context)+str(prompt)
|
80 |
inputs = tokenizer.encode(input_text, return_tensors="pt", add_special_tokens=False).to("cuda:0")
|
81 |
max_new_tokens=max_length
|
82 |
generation_config = GenerationConfig(
|
@@ -90,25 +62,23 @@ def generate_text(modelin, prompt, context, max_length=2100):
|
|
90 |
outputs = model.generate(generation_config=generation_config, input_ids=inputs, stopping_criteria=stopping_criteria_list,)
|
91 |
return tokenizer.decode(outputs[0], skip_special_tokens=False) #True
|
92 |
|
93 |
-
def mostrar_respuesta(
|
94 |
try:
|
95 |
-
|
96 |
-
print('Pregunta: '+str(pregunta))
|
97 |
-
print('Contexto: '+str(contexto))
|
98 |
-
res= generate_text(modelo, pregunta, contexto, max_length=500)
|
99 |
-
print('Respuesta: '+str(contexto))
|
100 |
return str(res)
|
101 |
except Exception as e:
|
102 |
return str(e)
|
103 |
|
104 |
# Ejemplos de preguntas
|
105 |
-
mis_ejemplos = [
|
106 |
-
|
|
|
|
|
|
|
107 |
|
108 |
iface = gr.Interface(
|
109 |
fn=mostrar_respuesta,
|
110 |
-
inputs=[gr.
|
111 |
-
gr.Textbox(label="Contexto", value="You are a helpful AI assistant. Eres un experto cocinero hispanoamericano."),],
|
112 |
outputs=[gr.Textbox(label="Respuesta", lines=2),],
|
113 |
title="Recetas de la Abuel@",
|
114 |
description="Introduce tu pregunta sobre recetas de cocina.",
|
|
|
2 |
import torch
|
3 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
4 |
from peft import PeftModel
|
|
|
5 |
import json
|
|
|
6 |
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, GemmaTokenizer, StoppingCriteria, StoppingCriteriaList, GenerationConfig
|
7 |
import os
|
8 |
|
|
|
9 |
#sft_model = "somosnlp/RecetasDeLaAbuela_mistral-7b-instruct-v0.2-bnb-4bit"
|
10 |
#base_model_name = "unsloth/Mistral-7B-Instruct-v0.2"
|
11 |
+
sft_model = "somosnlp/RecetasDeLaAbuela_gemma-2b-it-bnb-4bit"
|
|
|
12 |
base_model_name = "unsloth/gemma-2b-it-bnb-4bit"
|
13 |
|
|
|
|
|
|
|
|
|
|
|
14 |
max_seq_length=400
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
base_model = AutoModelForCausalLM.from_pretrained(base_model_name,return_dict=True,device_map="auto", torch_dtype=torch.float16,)
|
|
|
|
|
16 |
tokenizer = AutoTokenizer.from_pretrained(base_model_name, max_length = max_seq_length)
|
|
|
17 |
ft_model = PeftModel.from_pretrained(base_model, sft_model)
|
18 |
model = ft_model.merge_and_unload()
|
19 |
model.save_pretrained(".")
|
|
|
20 |
tokenizer.save_pretrained(".")
|
21 |
|
22 |
class ListOfTokensStoppingCriteria(StoppingCriteria):
|
|
|
46 |
# Añade tu criterio de parada a una StoppingCriteriaList
|
47 |
stopping_criteria_list = StoppingCriteriaList([stopping_criteria])
|
48 |
|
49 |
+
def generate_text(prompt, context, max_length=2100):
|
|
|
|
|
|
|
|
|
|
|
50 |
prompt=prompt.replace("\n", "").replace("¿","").replace("?","")
|
51 |
+
input_text = f'''<bos><start_of_turn>system ¿{context}?<end_of_turn><start_of_turn>user ¿{prompt}?<end_of_turn><start_of_turn>model'''
|
|
|
52 |
inputs = tokenizer.encode(input_text, return_tensors="pt", add_special_tokens=False).to("cuda:0")
|
53 |
max_new_tokens=max_length
|
54 |
generation_config = GenerationConfig(
|
|
|
62 |
outputs = model.generate(generation_config=generation_config, input_ids=inputs, stopping_criteria=stopping_criteria_list,)
|
63 |
return tokenizer.decode(outputs[0], skip_special_tokens=False) #True
|
64 |
|
65 |
+
def mostrar_respuesta(pregunta, contexto):
|
66 |
try:
|
67 |
+
res= generate_text(pregunta, contexto, max_length=500)
|
|
|
|
|
|
|
|
|
68 |
return str(res)
|
69 |
except Exception as e:
|
70 |
return str(e)
|
71 |
|
72 |
# Ejemplos de preguntas
|
73 |
+
mis_ejemplos = [
|
74 |
+
["¿Dime la receta de la tortilla de patatatas?"],
|
75 |
+
["¿Dime la receta del ceviche?"],
|
76 |
+
["¿Como se cocinan unos autenticos frijoles?"],
|
77 |
+
]
|
78 |
|
79 |
iface = gr.Interface(
|
80 |
fn=mostrar_respuesta,
|
81 |
+
inputs=[gr.Textbox(label="Pregunta"), gr.Textbox(label="Contexto", value="You are a helpful AI assistant. Eres un experto cocinero hispanoamericano."),],
|
|
|
82 |
outputs=[gr.Textbox(label="Respuesta", lines=2),],
|
83 |
title="Recetas de la Abuel@",
|
84 |
description="Introduce tu pregunta sobre recetas de cocina.",
|