Spaces:
Runtime error
Runtime error
import gradio as gr | |
import torch | |
from transformers import ( | |
AutoModelForCausalLM, | |
AutoTokenizer, | |
TextIteratorStreamer, | |
LlamaTokenizer, | |
) | |
import os | |
from threading import Thread | |
import spaces | |
import subprocess | |
# flash-attn ๋ผ์ด๋ธ๋ฌ๋ฆฌ ์ค์น. CUDA ๋น๋๋ ๊ฑด๋๋. | |
subprocess.run( | |
"pip install flash-attn --no-build-isolation", | |
env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"}, | |
shell=True, | |
) | |
# Hugging Face ํ ํฐ ๊ฐ์ ธ์ค๊ธฐ | |
token = os.environ["HF_TOKEN"] | |
# apple/OpenELM-270M ๋ชจ๋ธ๊ณผ ํ ํฌ๋์ด์ ๋ก๋ | |
# ํ ํฌ๋์ด์ ๊ฐ ์ค๋ฅ๋๋ ๋ฌธ์ ๊ฐ ์์ด์ NousResearch/Llama-2-7b-hf๋ฅผ ์ | |
# ํ๊ตญ์ด ๋ชจ๋ธ ํ ํฌ๋์ด์ ๋ก ๋ฐ๊ฟ๋ด beomi/llama-2-ko-7b | |
# apple/OpenELM-1.1B ํ ํฌ๋์ด์ ๋ง ํฌ๊ฒ ํด๋ด <- ์๋จ | |
# apple/OpenELM-3B-Instruct๋ก ๋๋ค ๋ณ๊ฒฝ ํด๋ด <- ์๋จ | |
model = AutoModelForCausalLM.from_pretrained( | |
"apple/OpenELM-270M-Instruct", | |
token=token, | |
trust_remote_code=True, | |
) | |
tok = AutoTokenizer.from_pretrained( | |
"NousResearch/Llama-2-7b-hf", | |
token=token, | |
trust_remote_code=True, | |
tokenizer_class=LlamaTokenizer, | |
) | |
# ์ข ๋ฃ ํ ํฐ ID ์ค์ | |
terminators = [ | |
tok.eos_token_id, | |
] | |
# GPU๊ฐ ์ฌ์ฉ ๊ฐ๋ฅํ ๊ฒฝ์ฐ GPU๋ก, ์๋๋ฉด CPU๋ก ๋ชจ๋ธ ๋ก๋ | |
if torch.cuda.is_available(): | |
device = torch.device("cuda") | |
print(f"Using GPU: {torch.cuda.get_device_name(device)}") | |
else: | |
device = torch.device("cpu") | |
print("Using CPU") | |
model = model.to(device) | |
# Spaces์ GPU ์์์ ์ฌ์ฉํ์ฌ chat ํจ์ ์คํ. ์ต๋ 60์ด ๋์ GPU ์์ ์ฌ์ฉ ๊ฐ๋ฅ. | |
def chat(message, history, temperature, do_sample, max_tokens): | |
# ์ฑํ ๊ธฐ๋ก์ ์ ์ ํ ํ์์ผ๋ก ๋ณํ | |
chat = [] | |
for item in history: | |
chat.append({"role": "user", "content": item[0]}) | |
if item[1] is not None: | |
chat.append({"role": "assistant", "content": item[1]}) | |
chat.append({"role": "user", "content": message}) | |
# ํ ํฌ๋์ด์ ๋ฅผ ์ฌ์ฉํ์ฌ ์ ๋ ฅ ์ฒ๋ฆฌ | |
messages = tok.apply_chat_template(chat, tokenize=False, add_generation_prompt=True) | |
model_inputs = tok([messages], return_tensors="pt").to(device) | |
# TextIteratorStreamer๋ฅผ ์ฌ์ฉํ์ฌ ๋ชจ๋ธ ์ถ๋ ฅ ์คํธ๋ฆฌ๋ฐ | |
streamer = TextIteratorStreamer( | |
tok, timeout=20.0, skip_prompt=True, skip_special_tokens=True | |
) | |
# ์์ฑ ๊ด๋ จ ๋งค๊ฐ๋ณ์ ์ค์ | |
generate_kwargs = dict( | |
model_inputs, | |
streamer=streamer, | |
max_new_tokens=max_tokens, # ์์ฑํ ์ต๋ ์ ํ ํฐ ์ | |
do_sample=True, # ์ํ๋ง ์ฌ๋ถ | |
temperature=temperature, # ์จ๋ ๋งค๊ฐ๋ณ์. ๋์์๋ก ๋ค์์ฑ ์ฆ๊ฐ | |
eos_token_id=terminators, # ์ข ๋ฃ ํ ํฐ ID | |
) | |
# ์จ๋๊ฐ 0์ด๋ฉด ์ํ๋งํ์ง ์์ | |
if temperature == 0: | |
generate_kwargs["do_sample"] = False | |
# ๋ณ๋ ์ค๋ ๋์์ ๋ชจ๋ธ ์์ฑ ์์ | |
t = Thread(target=model.generate, kwargs=generate_kwargs) | |
t.start() | |
# ์์ฑ๋ ํ ์คํธ๋ฅผ ๋ฐ๋ณต์ ์ผ๋ก yield | |
partial_text = "" | |
for new_text in streamer: | |
partial_text += new_text | |
yield partial_text | |
yield partial_text | |
# Gradio์ ChatInterface๋ฅผ ์ฌ์ฉํ์ฌ ๋ํํ ์ธํฐํ์ด์ค ์์ฑ | |
demo = gr.ChatInterface( | |
fn=chat, | |
examples=[["let's talk about korea"]], | |
additional_inputs_accordion=gr.Accordion( | |
label="โ๏ธ Parameters", open=False, render=False | |
), | |
additional_inputs=[ | |
gr.Slider( | |
minimum=0, maximum=1, step=0.1, value=0.7, label="Temperature", render=False | |
), | |
gr.Checkbox(label="Sampling", value=True), | |
gr.Slider( | |
minimum=128, | |
maximum=4096, | |
step=1, | |
value=512, | |
label="Max new tokens", | |
render=False, | |
), | |
], | |
stop_btn="Stop Generation", | |
title="Chat With LLMs", | |
description="Now Running [apple/OpenELM-270M](https://huggingface.co/apple/OpenELM-270M)", | |
) | |
# Gradio ์ธํฐํ์ด์ค ์คํ | |
demo.launch() |