Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
|
| 3 |
+
|
| 4 |
+
model_id = "clibrain/Llama-2-7b-ft-instruct-es"
|
| 5 |
+
|
| 6 |
+
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True).to("cuda")
|
| 7 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 8 |
+
|
| 9 |
+
def create_instruction(instruction, input_data=None, context=None):
|
| 10 |
+
sections = {
|
| 11 |
+
"Instrucci贸n": instruction,
|
| 12 |
+
"Entrada": input_data,
|
| 13 |
+
"Contexto": context,
|
| 14 |
+
}
|
| 15 |
+
|
| 16 |
+
system_prompt = "A continuaci贸n hay una instrucci贸n que describe una tarea, junto con una entrada que proporciona m谩s contexto. Escriba una respuesta que complete adecuadamente la solicitud.\n\n"
|
| 17 |
+
prompt = system_prompt
|
| 18 |
+
|
| 19 |
+
for title, content in sections.items():
|
| 20 |
+
if content is not None:
|
| 21 |
+
prompt += f"### {title}:\n{content}\n\n"
|
| 22 |
+
|
| 23 |
+
prompt += "### Respuesta:\n"
|
| 24 |
+
|
| 25 |
+
return prompt
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def generate(
|
| 29 |
+
instruction,
|
| 30 |
+
input=None,
|
| 31 |
+
context=None,
|
| 32 |
+
max_new_tokens=128,
|
| 33 |
+
temperature=0.1,
|
| 34 |
+
top_p=0.75,
|
| 35 |
+
top_k=40,
|
| 36 |
+
num_beams=4,
|
| 37 |
+
**kwargs
|
| 38 |
+
):
|
| 39 |
+
|
| 40 |
+
prompt = create_instruction(instruction, input, context)
|
| 41 |
+
print(prompt.replace("### Respuesta:\n", ""))
|
| 42 |
+
inputs = tokenizer(prompt, return_tensors="pt")
|
| 43 |
+
input_ids = inputs["input_ids"].to("cuda")
|
| 44 |
+
attention_mask = inputs["attention_mask"].to("cuda")
|
| 45 |
+
generation_config = GenerationConfig(
|
| 46 |
+
temperature=temperature,
|
| 47 |
+
top_p=top_p,
|
| 48 |
+
top_k=top_k,
|
| 49 |
+
num_beams=num_beams,
|
| 50 |
+
**kwargs,
|
| 51 |
+
)
|
| 52 |
+
with torch.no_grad():
|
| 53 |
+
generation_output = model.generate(
|
| 54 |
+
input_ids=input_ids,
|
| 55 |
+
attention_mask=attention_mask,
|
| 56 |
+
generation_config=generation_config,
|
| 57 |
+
return_dict_in_generate=True,
|
| 58 |
+
output_scores=True,
|
| 59 |
+
max_new_tokens=max_new_tokens,
|
| 60 |
+
early_stopping=True
|
| 61 |
+
)
|
| 62 |
+
s = generation_output.sequences[0]
|
| 63 |
+
output = tokenizer.decode(s)
|
| 64 |
+
return output.split("### Respuesta:")[1].lstrip("\n")
|
| 65 |
+
|
| 66 |
+
instruction = "Dame una lista de lugares a visitar en Espa帽a."
|
| 67 |
+
print(generate(instruction))
|