Hug0endob's picture
Update app.py
d125cdc verified
raw
history blame
3.01 kB
import os
import io
import time
import sys
import requests
from PIL import Image, ImageSequence
import gradio as gr
# llama-cpp-python import
try:
from llama_cpp import Llama
except Exception as e:
raise RuntimeError("llama-cpp-python import failed: " + str(e))
MODEL_PATH = os.path.join("model", "llama-joycaption-q4_k_m.gguf")
if not os.path.exists(MODEL_PATH):
raise FileNotFoundError(f"Model not found at {MODEL_PATH}. Ensure start.sh downloaded the GGUF.")
def download_bytes(url: str, timeout: int = 30) -> bytes:
with requests.get(url, stream=True, timeout=timeout) as r:
r.raise_for_status()
return r.content
def load_first_frame_from_bytes(raw: bytes):
img = Image.open(io.BytesIO(raw))
if getattr(img, "is_animated", False):
img = next(ImageSequence.Iterator(img))
if img.mode != "RGB":
img = img.convert("RGB")
return img
def make_prompt_for_image(image_path: str, user_prompt: str = "Describe the image."):
# JoyCaption-style multimodal GGUFs accept <img>{path}</img>
return f"<img>{image_path}</img>\nUser: {user_prompt}\nAssistant:"
# Initialize model (low-resource options)
print("Loading GGUF model (this can take 30–120s)...", file=sys.stderr)
llm = Llama(model_path=MODEL_PATH, n_ctx=2048, n_threads=2)
def generate_caption_from_url(url: str, prompt: str = "Describe the image."):
if not url:
return "No URL provided."
try:
raw = download_bytes(url)
except Exception as e:
return f"Download error: {e}"
try:
img = load_first_frame_from_bytes(raw)
except Exception as e:
return f"Image processing error: {e}"
tmp_dir = "/tmp/joycap"
os.makedirs(tmp_dir, exist_ok=True)
ts = int(time.time() * 1000)
tmp_path = os.path.join(tmp_dir, f"{ts}.jpg")
try:
img.save(tmp_path, format="JPEG", quality=85)
except Exception as e:
return f"Failed to save temp image: {e}"
prompt_full = make_prompt_for_image(tmp_path, prompt)
try:
resp = llm.create(
prompt=prompt_full,
max_tokens=256,
temperature=0.2,
top_p=0.95,
stop=["User:", "Assistant:"],
)
text = resp.get("choices", [{}])[0].get("text", "").strip()
return text or "No caption generated."
except Exception as e:
return f"Inference error: {e}"
finally:
try:
os.remove(tmp_path)
except Exception:
pass
iface = gr.Interface(
fn=generate_caption_from_url,
inputs=[
gr.Textbox(label="Image URL", placeholder="https://example.com/photo.jpg"),
gr.Textbox(label="Prompt (optional)", value="Describe the image."),
],
outputs=gr.Textbox(label="Generated caption"),
title="JoyCaption GGUF (Q4_K_M)",
description="Runs a quantized JoyCaption GGUF locally via llama.cpp (no external API).",
)
if __name__ == "__main__":
iface.launch(server_name="0.0.0.0", server_port=7860)