Create app.py

app.py

@@ -0,0 +1,64 @@
+"""
+Qari-OCR ZeroGPU Space — Arabic page OCR for the history-textbook project.
+
+Deploy this as a Gradio Space with ZeroGPU hardware (free on HF PRO).
+It exposes an API endpoint "/ocr" that takes a page image and returns the
+transcribed Arabic text. The local batch script (../batch_ocr.py) drives it
+over all 120 pages.
+"""
+import spaces
+import torch
+import gradio as gr
+from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
+from qwen_vl_utils import process_vision_info
+
+MODEL_ID = "NAMAA-Space/Qari-OCR-v0.3-VL-2B-Instruct"
+
+# High max_pixels keeps small Arabic glyphs + tashkeel legible (OCR needs detail).
+# 28 is Qwen2-VL's patch factor; 4096*28*28 ≈ 3.2M px ≈ a 300-DPI textbook page.
+processor = AutoProcessor.from_pretrained(
+    MODEL_ID, min_pixels=256 * 28 * 28, max_pixels=4096 * 28 * 28
+)
+# Loaded on CPU at startup — ZeroGPU only attaches a GPU inside @spaces.GPU functions.
+model = Qwen2VLForConditionalGeneration.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16)
+
+PROMPT = (
+    "Below is an image of one page of an Arabic school textbook. "
+    "Transcribe ALL the Arabic text exactly as printed, preserving line breaks, "
+    "headings, and right-to-left reading order. Keep diacritics (tashkeel) if present. "
+    "Ignore any faint diagonal draft watermark. Output only the transcribed text."
+)
+
+
+@spaces.GPU(duration=120)
+def ocr(image):
+    if image is None:
+        return ""
+    model.to("cuda")
+    messages = [{"role": "user", "content": [
+        {"type": "image", "image": image},
+        {"type": "text", "text": PROMPT},
+    ]}]
+    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    image_inputs, video_inputs = process_vision_info(messages)
+    inputs = processor(
+        text=[text], images=image_inputs, videos=video_inputs,
+        padding=True, return_tensors="pt",
+    ).to("cuda")
+    with torch.no_grad():
+        generated = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
+    trimmed = [g[len(i):] for i, g in zip(inputs.input_ids, generated)]
+    return processor.batch_decode(
+        trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
+    )[0]
+
+
+with gr.Blocks(title="Qari-OCR — Arabic page OCR") as demo:
+    gr.Markdown("## Qari-OCR — Arabic textbook page OCR\nUpload a page image, or call the `/ocr` API.")
+    with gr.Row():
+        inp = gr.Image(type="pil", label="Page image")
+        out = gr.Textbox(label="Transcribed Arabic", lines=25, rtl=True)
+    gr.Button("Run OCR", variant="primary").click(ocr, inp, out, api_name="ocr")
+
+if __name__ == "__main__":
+    demo.queue(max_size=16).launch()