Hey ! I am trying to build a Space with Streamlit where you can run inferences, but my code doesn’t seem to be able to complete the generate()
function. The code goes to st.write(f'inst: {instruction}')
but not up to t.write('Gen done')
. Here is the code:
import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
import warnings
import os
import streamlit as st
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
warnings.filterwarnings("ignore", category=UserWarning, module='transformers.generation.utils')
@st.cache_resource
def load_model():
base_model = "microsoft/phi-2"
peft_model_id = "STEM-AI-mtl/phi-2-electrical-engineering"
config = PeftConfig.from_pretrained(peft_model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="cpu", return_dict=True, trust_remote_code=True)
model = PeftModel.from_pretrained(model, peft_model_id, trust_remote_code=True)
#model = model.to('cpu')
return model
@st.cache_resource
def load_tokenizer():
base_model = "microsoft/phi-2"
return AutoTokenizer.from_pretrained(base_model)
model = load_model()
tokenizer = load_tokenizer()
def generate(instruction, model, tokenizer):
inputs = tokenizer(instruction, return_tensors="pt", return_attention_mask=False)
#inputs = inputs.to('cpu')
st.write(f'inst: {instruction}')
outputs = model.generate(
**inputs,
max_length=350,
do_sample=True,
temperature=0.7,
top_k=50,
top_p=0.9,
repetition_penalty=1,
)
st.write('Gen done')
text = tokenizer.batch_decode(outputs)[0]
st.write('Decode done')
return text
# Streamlit interface
st.title('Electrical_engineer.AI')
instruction = st.text_input("Enter your instruction:")
if st.button('Submit'):
if instruction:
answer = generate(instruction, model, tokenizer)
st.write(f'Answer: {answer}')
else:
st.write("Please enter an instruction.")