import os from threading import Thread import gradio as gr import torch from transformers import (AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, set_seed) from huggingface_hub import Repository import json theme = gr.themes.Monochrome( primary_hue="indigo", secondary_hue="blue", neutral_hue="slate", radius_size=gr.themes.sizes.radius_sm, font=[gr.themes.GoogleFont("Open Sans"), "ui-sans-serif", "system-ui", "sans-serif"], ) HF_TOKEN = os.environ.get("HF_TOKEN", None) os.environ["TOKENIZERS_PARALLELISM"] = "false" if HF_TOKEN: repo = Repository( local_dir="data", clone_from="trl-lib/stack-llama-prompts", use_auth_token=HF_TOKEN, repo_type="dataset" ) device = "cuda" if torch.cuda.is_available() else "cpu" model_id = "trl-lib/llama-se-rl-merged" print(f"Loading model: {model_id}") if device == "cpu": model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, use_auth_token=HF_TOKEN) else: model = AutoModelForCausalLM.from_pretrained( model_id, device_map="auto", load_in_8bit=True, use_auth_token=HF_TOKEN ) tokenizer = AutoTokenizer.from_pretrained(model_id, use_auth_token=HF_TOKEN) PROMPT_TEMPLATE = """Question: {prompt}\n\nAnswer:""" def save_inputs_and_outputs(inputs, outputs, generate_kwargs): with open(os.path.join("data", "prompts.jsonl"), "a") as f: json.dump({"inputs": inputs, "outputs": outputs, "generate_kwargs": generate_kwargs}, f, ensure_ascii=False) f.write("\n") commit_url = repo.push_to_hub() def generate(instruction, temperature=0.8, max_new_tokens=128, top_p=0.95, top_k=40): set_seed(42) formatted_instruction = PROMPT_TEMPLATE.format(prompt=instruction) temperature = float(temperature) top_p = float(top_p) streamer = TextIteratorStreamer(tokenizer) model_inputs = tokenizer(formatted_instruction, return_tensors="pt", truncation=True, max_length=2048).to(device) generate_kwargs = dict( top_p=top_p, temperature=temperature, max_new_tokens=max_new_tokens, do_sample=True, top_k=top_k, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.eos_token_id, ) t = Thread(target=model.generate, kwargs={**dict(model_inputs, streamer=streamer), **generate_kwargs}) t.start() output = "" hidden_output = "" for new_text in streamer: # skip streaming until new text is available if len(hidden_output) <= len(formatted_instruction): hidden_output += new_text continue # replace eos token # if tokenizer.eos_token in new_text: # new_text = new_text.replace(tokenizer.eos_token, "") output += new_text yield output if HF_TOKEN: print("Pushing prompt and completion to the Hub") save_inputs_and_outputs(formatted_instruction, output, generate_kwargs) return output examples = [ "A llama is in my lawn. How do I get rid of him?", "How do I create an array in C++ which contains all even numbers between 1 and 10?", "How can I sort a list in Python?", "How can I write a Java function to generate the nth Fibonacci number?", "How many helicopters can a llama eat in one sitting?", ] def process_example(args): for x in generate(args): pass return x with gr.Blocks(theme=theme, analytics_enabled=False, css=".generating {visibility: hidden}") as demo: with gr.Column(): gr.Markdown( """

🦙🦙🦙 StackLLaMa 🦙🦙🦙

StackLLaMa is a 7 billion parameter language model that has been trained on pairs of questions and answers from [Stack Exchange](https://stackexchange.com) using Reinforcement Learning from Human Feedback with the [TRL library](https://github.com/lvwerra/trl). For more details, check out our [blog post](https://huggingface.co/blog/stackllama). Type in the box below and click the button to generate answers to your most pressing questions 🔥! **Note:** we are collecting your prompts and model completions for research purposes. """ ) with gr.Row(): with gr.Column(scale=3): instruction = gr.Textbox(placeholder="Enter your question here", label="Question") with gr.Box(): gr.Markdown("**Answer**") output = gr.Markdown() submit = gr.Button("Generate", variant="primary") gr.Examples( examples=examples, inputs=[instruction], cache_examples=True, fn=process_example, outputs=[output], ) with gr.Column(scale=1): temperature = gr.Slider( label="Temperature", value=0.8, minimum=0.01, maximum=2.0, step=0.1, interactive=True, info="Higher values produce more diverse outputs", ) max_new_tokens = gr.Slider( label="Max new tokens", value=128, minimum=0, maximum=2048, step=4, interactive=True, info="The maximum numbers of new tokens", ) top_p = gr.Slider( label="Top-p (nucleus sampling)", value=0.95, minimum=0.0, maximum=1, step=0.05, interactive=True, info="Higher values sample more low-probability tokens", ) top_k = gr.Slider( label="Top-k", value=40, minimum=0, maximum=100, step=2, interactive=True, info="Sample from top-k tokens", ) submit.click(generate, inputs=[instruction, temperature, max_new_tokens, top_p, top_k], outputs=[output]) instruction.submit(generate, inputs=[instruction, temperature, max_new_tokens, top_p, top_k], outputs=[output]) demo.queue(concurrency_count=1) demo.launch(enable_queue=True, share=True)