app_vllm.py

# app_vllm.py - Faster inference using vLLM
import os
import spaces
import gradio as gr
from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest
from transformers import AutoTokenizer

HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACEHUB_API_TOKEN")

BASE_MODEL_ID = "meta-llama/Meta-Llama-3.1-8B-Instruct"
PEFT_MODEL_ID = "befm/Be.FM-8B"

# Use /data for persistent storage to avoid re-downloading models
CACHE_DIR = "/data" if os.path.exists("/data") else None

def load_model():
    if HF_TOKEN is None:
        raise RuntimeError(
            "HF_TOKEN is not set. Add it in Space → Settings → Secrets. "
            "Also ensure your account has access to the gated base model."
        )

    # Initialize vLLM with PEFT support
    llm = LLM(
        model=BASE_MODEL_ID,
        tokenizer=BASE_MODEL_ID,
        enable_lora=True,
        max_lora_rank=64,
        dtype="float16",
        gpu_memory_utilization=0.7,  # Reduced from 0.9 to avoid OOM on T4 GPU
        trust_remote_code=True,
        download_dir=CACHE_DIR,  # Use persistent storage
    )

    print(f"[INFO] vLLM loaded base model: {BASE_MODEL_ID}")
    print(f"[INFO] Using cache directory: {CACHE_DIR}")

    # Load PEFT adapter
    lora_request = LoRARequest(
        lora_name="befm",
        lora_int_id=1,
        lora_path=PEFT_MODEL_ID,
    )
    print(f"[INFO] PEFT adapter prepared: {PEFT_MODEL_ID}")

    return llm, lora_request

# Lazy load model and tokenizer
_llm = None
_lora_request = None
_tokenizer = None

def get_model_and_tokenizer():
    global _llm, _lora_request, _tokenizer
    if _llm is None:
        _llm, _lora_request = load_model()
        _tokenizer = AutoTokenizer.from_pretrained(
            BASE_MODEL_ID,
            token=HF_TOKEN,
            cache_dir=CACHE_DIR  # Use persistent storage
        )
    return _llm, _lora_request, _tokenizer

@spaces.GPU
def generate_response(messages, max_new_tokens=512, temperature=0.7) -> str:
    llm, lora_request, tokenizer = get_model_and_tokenizer()

    # Apply Llama 3.1 chat template
    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

    sampling_params = SamplingParams(
        temperature=temperature,
        top_p=0.9,
        max_tokens=max_new_tokens,
    )

    # Generate with vLLM
    outputs = llm.generate(
        prompts=[prompt],
        sampling_params=sampling_params,
        lora_request=lora_request,
    )

    return outputs[0].outputs[0].text

def chat_fn(message, history, system_prompt, _prompt_reference, max_new_tokens, temperature):
    # Build conversation in Llama 3.1 chat format
    messages = []

    # Add system prompt (use default if not provided)
    if not system_prompt:
        system_prompt = (
            "Your are a Be.FM assistant. Be.FM is a family of open foundation models "
            "designed for human behavior modeling. Built on Llama 3.1 and fine-tuned on "
            "diverse behavioral datasets, Be.FM models are designed to enhance the "
            "understanding and prediction of human decision-making."
        )
    messages.append({"role": "system", "content": system_prompt})

    # History is already in dict format: [{"role": "user", "content": "..."}, ...]
    for msg in (history or []):
        messages.append(msg)

    if message:
        messages.append({"role": "user", "content": message})

    reply = generate_response(
        messages,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
    )
    return reply

demo = gr.ChatInterface(
    fn=lambda message, history, system_prompt, prompt_reference, max_new_tokens, temperature:
        chat_fn(message, history, system_prompt, prompt_reference, max_new_tokens, temperature),
    additional_inputs=[
        gr.Textbox(
            label="System prompt (optional)",
            placeholder=(
                "Your are a Be.FM assistant. Be.FM is a family of open foundation models "
                "designed for human behavior modeling. Built on Llama 3.1 and fine-"
                "tuned on diverse behavioral datasets, Be.FM models are designed to "
                "enhance the understanding and prediction of human decision-making."
            ),
            lines=2,
        ),
        gr.Markdown(
            "For system and user prompts in a variety of behavioral tasks, please refer "
            "to the appendix in our [paper](https://arxiv.org/abs/2505.23058)."
        ),
        gr.Slider(16, 2048, value=512, step=16, label="max_new_tokens"),
        gr.Slider(0.1, 1.5, value=0.6, step=0.05, label="temperature"),
    ],
    title="Be.FM: Open Foundation Models for Human Behavior (8B)",
)

if __name__ == "__main__":
    demo.launch()