app.py

import os
import time
import torch
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

HF_TOKEN = os.getenv("hf_token")
MODEL_ID = "ruSpamModels/ruSpam-Qwen-0.5B-50k"

device = "cuda" if torch.cuda.is_available() else "cpu"

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B-Instruct",
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
    device_map=device,
    trust_remote_code=True,
    token=HF_TOKEN,
)

model = PeftModel.from_pretrained(
    base_model,
    MODEL_ID,
    token=HF_TOKEN,
)
model.eval()

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_ID,
    token=HF_TOKEN,
)

def classify(message):
    prompt = (
        "You are a spam classifier.\n"
        "Answer with one word: spam or ham.\n\n"
        f"Message:\n{message}\n\n"
        "Answer:"
    )

    inputs = tokenizer(prompt, return_tensors="pt").to(device)

    start = time.time()
    with torch.no_grad():
        out = model.generate(
            **inputs,
            max_new_tokens=1,
            do_sample=False,
            temperature=0.01,
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )

    elapsed = (time.time() - start) * 1000
    new_token_id = out[0, inputs["input_ids"].shape[1]]
    answer = tokenizer.decode(new_token_id).strip().lower()

    if answer.startswith("spam"):
        label = "SPAM"
    elif answer.startswith("ham"):
        label = "HAM"
    else:
        label = "UNKNOWN"

    return f"{label} ({elapsed:.1f} ms)"

iface = gr.Interface(
    fn=classify,
    inputs=gr.Textbox(lines=4),
    outputs=gr.Textbox(),
    title="ruSpam Qwen 0.5B",
)

iface.launch()