Deploy vision-ocr-studio on ZeroGPU

README.md

@@ -1,13 +1,26 @@
 ---
-title: Vision Ocr Studio
-emoji: 🌍
-colorFrom: pink
-colorTo: pink
+title: Smart Document, Blueprint & Invoice OCR Studio
+emoji: 📄
+colorFrom: blue
+colorTo: green
 sdk: gradio
-sdk_version: 6.31.0
-python_version: '3.12'
+sdk_version: 6.13.0
 app_file: app.py
-pinned: false
+python_version: "3.12"
+startup_duration_timeout: 1h
+short_description: Qwen2.5-VL Document, Blueprint, Table & Invoice OCR
+models:
+  - Qwen/Qwen2.5-VL-3B-Instruct
+tags:
+  - ocr
+  - vision-language
+  - qwen2.5-vl
+  - document-ai
+  - invoice-parser
+  - zerogpu
+  - mcp-server
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 📄 Smart Document, Blueprint & Invoice OCR Studio
+
+Upload **complex invoices, electrical panel schedules, blueprints, receipts, tables, or UI screenshots** and extract clean **Markdown or structured JSON** powered by **[Qwen/Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct)** on **ZeroGPU (`zero-a10g`)** at **$0/hour**.

app.py

@@ -0,0 +1,132 @@
+"""Qwen2.5-VL Smart Document, Blueprint & Invoice OCR Studio on ZeroGPU."""
+
+import os
+
+os.environ.setdefault("PYTORCH_ALLOC_CONF", "backend:cudaMallocAsync")
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "backend:cudaMallocAsync")
+
+import spaces  # Must be imported before torch
+import gradio as gr
+import PIL.Image
+import torch
+from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
+
+MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct"
+
+EXTRACTION_PRESETS = {
+    "Full Clean Markdown + Tables (Default)": (
+        "Extract all text, headings, and tables from this image into clean, well-formatted GitHub Flavored Markdown. "
+        "Preserve exact numbers, part numbers, prices, and table columns."
+    ),
+    "Structured Invoice / Receipt JSON": (
+        "Extract all fields from this invoice or receipt as valid JSON with keys: vendor, date, invoice_number, "
+        "line_items (description, quantity, unit_price, total), subtotal, tax, and total_amount."
+    ),
+    "Electrical Panel / Blueprint / Spec Analysis": (
+        "Analyze this electrical diagram, panel schedule, or technical blueprint. List all visible circuits, "
+        "breaker ratings, wire gauges, load calculations, labels, and technical notes in structured Markdown."
+    ),
+    "Custom Question / Prompt": "",
+}
+
+print(f"[vision-ocr-studio] Preloading {MODEL_ID} at module scope for ZeroGPU...", flush=True)
+model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+    MODEL_ID,
+    torch_dtype=torch.bfloat16,
+)
+model.to("cuda")
+model.eval()
+processor = AutoProcessor.from_pretrained(MODEL_ID)
+print("[vision-ocr-studio] Qwen2.5-VL OCR engine ready on ZeroGPU!", flush=True)
+
+
+@spaces.GPU(duration=60)
+def analyze_document(
+    image_path: str | None,
+    preset: str = "Full Clean Markdown + Tables (Default)",
+    custom_instruction: str = "",
+    max_tokens: int = 1024,
+) -> str:
+    """Run vision-language OCR or structured extraction on an uploaded document/image."""
+    if not image_path:
+        raise gr.Error("Please upload a document, blueprint, invoice, or screenshot image first.")
+
+    instruction = (custom_instruction or "").strip() or EXTRACTION_PRESETS.get(
+        preset, EXTRACTION_PRESETS["Full Clean Markdown + Tables (Default)"]
+    )
+    img = PIL.Image.open(image_path).convert("RGB")
+    if max(img.size) > 1600:
+        scale = 1600.0 / max(img.size)
+        img = img.resize((int(img.width * scale), int(img.height * scale)), PIL.Image.Resampling.LANCZOS)
+
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image", "image": img},
+                {"type": "text", "text": instruction},
+            ],
+        }
+    ]
+    text_prompt = processor.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+    inputs = processor(
+        text=[text_prompt],
+        images=[img],
+        padding=True,
+        return_tensors="pt",
+    ).to("cuda")
+
+    with torch.no_grad():
+        generated_ids = model.generate(**inputs, max_new_tokens=int(max_tokens))
+
+    trimmed_ids = [
+        out_ids[len(in_ids):]
+        for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
+    ]
+    output_text = processor.batch_decode(
+        trimmed_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
+    )[0]
+    return output_text.strip()
+
+
+with gr.Blocks(title="Smart Document, Blueprint & Invoice OCR Studio", fill_width=True) as demo:
+    gr.Markdown(
+        "# 📄 Smart Document, Blueprint & Invoice OCR Studio\n"
+        "Powered by **[Qwen/Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct)** on **ZeroGPU (`zero-a10g`)** at **$0/hour**—extract clean **Markdown, JSON, tables, invoices, or electrical specs** from any image."
+    )
+    with gr.Row():
+        with gr.Column(scale=1):
+            img_in = gr.Image(
+                label="Upload Document, Invoice, Electrical Schedule, Blueprint, or Screenshot",
+                type="filepath",
+            )
+            preset_dd = gr.Dropdown(
+                choices=list(EXTRACTION_PRESETS),
+                value="Full Clean Markdown + Tables (Default)",
+                label="Extraction Mode",
+            )
+            custom_box = gr.Textbox(
+                label="Custom Instruction / Question (Optional — overrides preset when filled)",
+                lines=3,
+                placeholder="e.g., Extract all circuit breaker amps and wire sizes into a JSON array...",
+            )
+            max_tok = gr.Slider(256, 2048, value=1024, step=128, label="Max Output Tokens")
+            extract_btn = gr.Button("📄 Extract & Analyze Document (ZeroGPU)", variant="primary")
+        with gr.Column(scale=1):
+            md_out = gr.Textbox(
+                label="Extracted Output (Markdown / JSON)",
+                lines=18,
+                buttons=["copy"],
+            )
+
+    extract_btn.click(
+        fn=analyze_document,
+        inputs=[img_in, preset_dd, custom_box, max_tok],
+        outputs=md_out,
+        api_name="extract_document",
+    )
+
+if __name__ == "__main__":
+    demo.launch(theme=gr.themes.Ocean(), mcp_server=True)

requirements.txt

@@ -0,0 +1,6 @@
+transformers>=4.49.0
+accelerate>=1.6.0
+safetensors
+qwen-vl-utils
+Pillow
+torchvision