import os import shlex import subprocess from threading import Thread from typing import Iterator from huggingface_hub import hf_hub_download whl_path = hf_hub_download("google/gemma-2-9b-it", "transformers/transformers-4.42.0.dev0-py3-none-any.whl") subprocess.run(shlex.split(f"pip install {whl_path}")) import gradio as gr import spaces import torch from transformers import ( AutoModelForCausalLM, BitsAndBytesConfig, GemmaTokenizerFast, TextIteratorStreamer, ) DESCRIPTION = """\ # Gemma 2 9B IT Gemma 2 is Google's latest iteration of open LLMs. This is a demo of [`google/gemma-2-9b-it`](https://huggingface.co/google/gemma-2-9b-it), fine-tuned for instruction following. For more details, please check [our post](https://huggingface.co/blog/gemma2). """ MAX_MAX_NEW_TOKENS = 2048 DEFAULT_MAX_NEW_TOKENS = 1024 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096")) device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") model_id = "google/gemma-2-9b-it" tokenizer = GemmaTokenizerFast.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, device_map="auto", quantization_config=BitsAndBytesConfig(load_in_8bit=True), ) model.config.sliding_window = 4096 model.eval() @spaces.GPU(duration=90) def generate( message: str, chat_history: list[tuple[str, str]], max_new_tokens: int = 1024, temperature: float = 0.6, top_p: float = 0.9, top_k: int = 50, repetition_penalty: float = 1.2, ) -> Iterator[str]: conversation = [] for user, assistant in chat_history: conversation.extend( [ {"role": "user", "content": user}, {"role": "assistant", "content": assistant}, ] ) conversation.append({"role": "user", "content": message}) input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt") if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH: input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:] gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.") input_ids = input_ids.to(model.device) streamer = TextIteratorStreamer(tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True) generate_kwargs = dict( {"input_ids": input_ids}, streamer=streamer, max_new_tokens=max_new_tokens, do_sample=True, top_p=top_p, top_k=top_k, temperature=temperature, num_beams=1, repetition_penalty=repetition_penalty, ) t = Thread(target=model.generate, kwargs=generate_kwargs) t.start() outputs = [] for text in streamer: outputs.append(text) yield "".join(outputs) chat_interface = gr.ChatInterface( fn=generate, additional_inputs=[ gr.Slider( label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS, ), gr.Slider( label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6, ), gr.Slider( label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9, ), gr.Slider( label="Top-k", minimum=1, maximum=1000, step=1, value=50, ), gr.Slider( label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2, ), ], stop_btn=None, examples=[ ["Hello there! How are you doing?"], ["Can you explain briefly to me what is the Python programming language?"], ["Explain the plot of Cinderella in a sentence."], ["How many hours does it take a man to eat a Helicopter?"], ["Write a 100-word article on 'Benefits of Open-Source in AI research'"], ], ) with gr.Blocks(css="style.css", fill_height=True) as demo: gr.Markdown(DESCRIPTION) gr.DuplicateButton(value="Duplicate Space for private use", elem_id="duplicate-button") chat_interface.render() if __name__ == "__main__": demo.queue(max_size=20).launch()