app_peft.py

import spaces

import torch
import torch._dynamo

import gradio as gr

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer, BitsAndBytesConfig

torch._dynamo.config.suppress_errors = True
torch._dynamo.disable()

max_seq_length = 2048
dtype = (
    None
)
load_in_4bit = True


quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    # bnb_4bit_compute_dtype=torch.float16,
)

tokenizer = AutoTokenizer.from_pretrained("ua-l/gemma-2-9b-legal-uk")
model = AutoPeftModelForCausalLM.from_pretrained(
    "ua-l/gemma-2-9b-legal-uk",
    quantization_config=quantization_config,
    device_map='auto'
)


@spaces.GPU
def predict(question):
    inputs = tokenizer(
    [f'''### Question:
    {question}
    
    ### Answer:
'''], return_tensors = "pt").to("cuda")

    outputs = model.generate(**inputs, max_new_tokens = 128)
    
    results = tokenizer.batch_decode(outputs, skip_special_tokens=True)

    return results[0]

inputs = gr.Textbox(lines=2, label="Enter a question", value="Як отримати виплати ВПО?")

outputs = gr.Textbox(label="Answer")

demo = gr.Interface(fn=predict, inputs=inputs, outputs=outputs)
demo.launch()