app.py

# app.py
import os
from typing import Dict
import gradio as gr
import spaces
import torch
from PIL import Image
from transformers import AutoProcessor, LlavaOnevisionForConditionalGeneration

MODEL_ID = os.environ.get("MODEL_ID", "BSC-LT/salamandra-7b-vision")
DTYPE = torch.float16
DEVICE = "cuda"

_model = None
_processor = None

def _lazy_load():
    global _model, _processor
    if _model is None or _processor is None:
        _processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
        _model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            MODEL_ID,
            torch_dtype=DTYPE,
            low_cpu_mem_usage=True,
            trust_remote_code=True,
            use_safetensors=True,
            device_map=None,
        )
    return _model, _processor

def _compose_prompt(user_text: str):
    convo = [{"role": "user", "content": [{"type": "image"},
                                          {"type": "text", "text": user_text or "Describe la imagen con detalle."}]}]
    return convo

@spaces.GPU
def infer_core(image: Image.Image, text: str, max_new_tokens: int = 256, temperature: float = 0.7) -> str:
    model, processor = _lazy_load()
    prompt = processor.apply_chat_template(_compose_prompt(text), add_generation_prompt=True)
    model = model.to(DEVICE)
    inputs = processor(images=image, text=prompt, return_tensors="pt").to(DEVICE, DTYPE)
    with torch.inference_mode():
        out = model.generate(**inputs, max_new_tokens=int(max_new_tokens), temperature=float(temperature))
    return processor.decode(out[0], skip_special_tokens=True).strip()

# ---------- UI ----------
with gr.Blocks(title="Salamandra Vision 7B · ZeroGPU") as demo:
    gr.Markdown("## Salamandra-Vision 7B · ZeroGPU\nImagen + texto → descripción.")
    with gr.Row():
        with gr.Column():
            in_img = gr.Image(label="Imagen", type="pil")
            in_txt = gr.Textbox(label="Texto/prompt", value="Describe la imagen con detalle (ES/CA).")
            max_new = gr.Slider(16, 1024, value=256, step=16, label="max_new_tokens")
            temp = gr.Slider(0.0, 1.5, value=0.7, step=0.05, label="temperature")
            btn = gr.Button("Generar", variant="primary")
        with gr.Column():
            out = gr.Textbox(label="Descripción", lines=18)
    btn.click(infer_core, [in_img, in_txt, max_new, temp], out, api_name="describe")

# ---------- API pura (sin UI) ----------
# Exponemos un endpoint REST nítido (multipart/form-data o JSON base64) sin depender de componentes UI.
# /api/describe_raw -> recibe {image,file} y campos simples.
@gr.api()
@spaces.GPU
def describe_raw(image: gr.File, text: str = "Describe la imagen con detalle.", 
                 max_new_tokens: int = 256, temperature: float = 0.7) -> Dict[str, str]:
    img = Image.open(image)
    result = infer_core(img, text, max_new_tokens, temperature)
    return {"text": result}

demo.queue(concurrency_count=1, max_size=16).launch()