app.py

"""
PetBull-7B-VL demo Space
------------------------

• Base model  : Qwen/Qwen2.5-VL-7B-Instruct
• LoRA adapter: ColdSlim/PetBull-7B   (light-weight repo you just pushed)

Put this file in your Space, add a `requirements.txt` with:
    transformers>=4.41.0
    peft>=0.11.0
    accelerate
    gradio>=4.33

Then (optionally) switch the Space hardware to **GPU (shared)** in
Settings → Hardware for much faster vision-language inference.
"""

import torch, gradio as gr
from PIL import Image
from transformers import AutoProcessor, AutoModelForVision2Seq
from peft import PeftModel

# ---------------------------------------------------------------------
# 1  Load base + LoRA  (≈ 12 GB VRAM in bf16; falls back to CPU if needed)
# ---------------------------------------------------------------------
BASE_MODEL   = "Qwen/Qwen2.5-VL-7B-Instruct"
ADAPTER_REPO = "ColdSlim/PetBull-7B"     # 👉 replace with your HF path if different
ADAPTER_REV  = "master" 

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype  = torch.float16 if device == "cuda" else torch.float32

processor = AutoProcessor.from_pretrained(BASE_MODEL, trust_remote_code=True)
base      = AutoModelForVision2Seq.from_pretrained(
                BASE_MODEL, torch_dtype=dtype, trust_remote_code=True)
model     = PeftModel.from_pretrained(base, ADAPTER_REPO, revision=ADAPTER_REV)
model.to(device).eval()

# ---------------------------------------------------------------------
# 2  Inference helper
# ---------------------------------------------------------------------
def generate_answer(image: Image.Image | None,
                    question: str,
                    temperature: float = 0.7,
                    top_p: float = 0.95,
                    max_tokens: int = 512) -> str:
    """
    Runs one‐shot VQA chat. Image is optional; if None we still obey the
    prompt format required by Qwen-VL by inserting a blank white image.
    """
    if image is None:
        image = Image.new("RGB", (224, 224), color="white")

    inputs = processor(text=[question],
                       images=[image],
                       return_tensors="pt").to(device)

    with torch.no_grad():
        output_ids = model.generate(**inputs,
                                    max_new_tokens=max_tokens,
                                    temperature=temperature,
                                    top_p=top_p)
    return processor.batch_decode(output_ids,
                                  skip_special_tokens=True)[0]

# ---------------------------------------------------------------------
# 3  Gradio UI
# ---------------------------------------------------------------------
with gr.Blocks(title="PetBull-7B-VL – Ask a Vet Bot") as demo:
    gr.Markdown(
        """
        ## 🐾 PetBull-7B-VL  
        Upload a photo of your pet **and/or** ask a question.  
        The model will analyse the image (if provided) and give tailored advice.
        """
    )

    with gr.Row():
        with gr.Column(scale=1):
            img_in   = gr.Image(type="pil", label="Pet photo (optional)")
            txt_in   = gr.Textbox(lines=3, placeholder="Describe the issue or ask a question…")
            run_btn  = gr.Button("Ask PetBull")
            temp_sl  = gr.Slider(0.1, 1.5, 0.7, label="Temperature")
            topp_sl  = gr.Slider(0.1, 1.0, 0.95, label="Top-p")
            max_sl   = gr.Slider(32, 1024, 512, step=8, label="Max new tokens")
        with gr.Column(scale=1):
            answer   = gr.Textbox(lines=12, label="Assistant", interactive=False)

    run_btn.click(fn=generate_answer,
                  inputs=[img_in, txt_in, temp_sl, topp_sl, max_sl],
                  outputs=answer)

demo.queue().launch()