app.py

import os

import spaces
import torch
import gradio as gr
from transformers import AutoModelForImageTextToText, AutoProcessor

MODEL_ID = "CohereLabs/North-Micro-Vision-Instruct"

# Load once at startup. On ZeroGPU the weights stay resident and
# @spaces.GPU allocates a worker per call.
print(f"Loading {MODEL_ID} ...")
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID,
    dtype=torch.bfloat16,
    device_map="cuda",
)
print("Model loaded!")


def _estimate_duration(image, prompt, max_new_tokens, temperature, top_p, top_k) -> int:
    """Rough wall-clock estimate (seconds) for one VLM call. Requesting less
    than the 60s default raises queue priority and frees the GPU slot sooner
    for the next visitor. Scaled by max_new_tokens; clamped to a safe range."""
    seconds = 10 + int(max_new_tokens) * 0.15
    return max(20, min(int(seconds), 120))


def _friendly_gpu_error(err: Exception) -> str:
    msg = (str(err) or "").lower()
    capacity_hints = (
        "gpu limit", "reached its gpu limit", "gpu quota", "out of quota",
        "quota", "no gpu", "could not allocate", "gpu is busy", "too many",
        "concurrent",
    )
    if any(h in msg for h in capacity_hints):
        return (
            "⛔ This demo's shared GPU is at capacity right now — it's not a "
            "problem with your input or your account. Please wait a minute and retry."
        )
    if "out of memory" in msg or "oom" in msg:
        return (
            "💥 Ran out of GPU memory. Try a smaller image or fewer max new "
            "tokens, then retry."
        )
    return "⚠️ Generation failed. Please try again in a moment."


@spaces.GPU(duration=_estimate_duration)
def _run_vlm_gpu(image, prompt, max_new_tokens, temperature, top_p, top_k):
    """GPU worker: runs only under a ZeroGPU allocation."""
    if not prompt or not prompt.strip():
        raise gr.Error("Please enter a prompt.")
    if image is None:
        raise gr.Error("Please provide an image.")

    if isinstance(image, dict):
        image_ref = image.get("path") or image.get("url")
    else:
        image_ref = image

    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "url": image_ref},
                {"type": "text", "text": prompt},
            ],
        }
    ]

    inputs = processor.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt",
        return_dict=True,
    ).to(model.device)
    if "pixel_values" in inputs:
        inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

    do_sample = float(temperature) > 0
    gen_kwargs = dict(max_new_tokens=int(max_new_tokens), do_sample=do_sample)
    if do_sample:
        gen_kwargs.update(
            temperature=float(temperature),
            top_p=float(top_p),
            top_k=int(top_k),
        )

    outputs = model.generate(**inputs, **gen_kwargs)
    generated_ids = outputs[0][inputs["input_ids"].shape[1]:]
    return processor.decode(
        generated_ids,
        skip_special_tokens=True,
        clean_up_tokenization_spaces=False,
    )


def run_vlm(image, prompt, max_new_tokens, temperature, top_p, top_k):
    """Workflow-facing wrapper bound to the canvas as a `fn` operator node.
    Catches ZeroGPU allocator rejections and rewords them for users."""
    try:
        return _run_vlm_gpu(image, prompt, max_new_tokens, temperature, top_p, top_k)
    except gr.Error:
        raise
    except Exception as e:
        raise gr.Error(_friendly_gpu_error(e)) from e


# The workflow (workflow.json) wires `run_vlm` as a `fn` operator:
#   Image, Prompt, Max New Tokens, Temperature, Top P, Top K ─▶
#   run_vlm (fn operator, kind="fn") ─▶ Response
demo = gr.Workflow(
    graph="workflow.json",
    bind={"run_vlm": run_vlm},
)

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