fix(app): complete two-module consolidation

Add the ZeroGPU model module, remove every remaining legacy launcher and package file, and update the runtime documentation after the source move.

README.md

@@ -30,4 +30,4 @@ uv sync
 uv run app.py
 ```
 
-The Hugging Face Space starts from `app.py`, which exposes the Gradio app defined in `app/hf_space.py`.
+The Hugging Face Space starts from `app.py`. Model loading and the ZeroGPU pipelines are defined in `model.py`.

app.py

@@ -1,15 +1,8 @@
-# ruff: noqa: I001
-
-import spaces
-
-import json
 from pathlib import Path
 
 import gradio as gr
-import torch
-from transformers import AutoModelForImageTextToText, AutoProcessor
 
-model_id = "numind/NuExtract3"
+from model import extract, generate_template
 
 structured_json_templates = {
     "Basic receipt": """{
@@ -215,98 +208,6 @@ structured_json_templates = {
 default_structured_json_template = "Invoice with line items"
 
 
-def select_structured_json_template(name):
-    return structured_json_templates[name]
-
-
-processor = AutoProcessor.from_pretrained(
-    model_id,
-    trust_remote_code=True,
-)
-
-model = (
-    AutoModelForImageTextToText.from_pretrained(
-        model_id,
-        attn_implementation="sdpa",
-        dtype=torch.bfloat16,
-        trust_remote_code=True,
-    )
-    .to("cuda")
-    .eval()
-)
-
-
-def run_nuextract(messages, **chat_template_kwargs):
-    inputs = processor.apply_chat_template(
-        messages,
-        add_generation_prompt=True,
-        tokenize=True,
-        return_dict=True,
-        return_tensors="pt",
-        **chat_template_kwargs,
-    ).to(model.device)
-
-    with torch.inference_mode():
-        generated_ids = model.generate(  # pyright: ignore[reportAttributeAccessIssue]
-            **inputs,
-            max_new_tokens=4096,
-            do_sample=False,
-        )
-
-    generated_ids = generated_ids[:, inputs.input_ids.shape[1] :]
-    return processor.batch_decode(
-        generated_ids,
-        skip_special_tokens=True,
-        clean_up_tokenization_spaces=False,
-    )[0].strip()
-
-
-@spaces.GPU
-def extract(image, text, template, enable_thinking):
-    result = run_nuextract(
-        [
-            {
-                "role": "user",
-                "content": [
-                    {"type": "image", "image": image},
-                    {"type": "text", "text": text},
-                ],
-            }
-        ],
-        mode="structured",
-        template=template,
-        enable_thinking=enable_thinking,
-    )
-
-    return json.loads(result)
-
-
-@spaces.GPU
-def generate_template(image):
-    result = run_nuextract(
-        [
-            {
-                "role": "user",
-                "content": [
-                    {"type": "image", "image": image},
-                    {
-                        "type": "text",
-                        "text": (
-                            "Create a reusable structured extraction template grounded only "
-                            "in the visible document. Include fields supported by the document, "
-                            "represent repeated records as arrays, use NuExtract template leaf "
-                            "types, and return only the JSON template."
-                        ),
-                    },
-                ],
-            }
-        ],
-        mode="template-generation",
-    )
-
-    return json.dumps(json.loads(result), indent=2)
-
-
 with gr.Blocks(title="NuMarkApp") as demo:
     gr.Markdown(
         "# NuMarkApp\n"
@@ -411,7 +312,7 @@ with gr.Blocks(title="NuMarkApp") as demo:
     )
 
     template_preset.change(
-        select_structured_json_template,
+        structured_json_templates.__getitem__,
         inputs=template_preset,
         outputs=template,
     )

app/hf_space.py

@@ -1,7 +0,0 @@
-from app.gradio import demo
-
-__all__ = ["demo"]
-
-
-if __name__ == "__main__":
-    demo.launch()

app/nuextract.py

@@ -1,68 +0,0 @@
-from pathlib import Path
-
-import torch
-from PIL import Image
-from transformers import AutoModelForImageTextToText, AutoProcessor
-
-model_id = "numind/NuExtract3"
-
-processor = AutoProcessor.from_pretrained(
-    model_id,
-    trust_remote_code=True,
-)
-
-model = AutoModelForImageTextToText.from_pretrained(
-    model_id,
-    dtype=torch.bfloat16,
-    device_map="auto",
-    trust_remote_code=True,
-).eval()
-
-
-def run_nuextract(messages, **chat_template_kwargs):
-    inputs = processor.apply_chat_template(
-        messages,
-        add_generation_prompt=True,
-        tokenize=True,
-        return_dict=True,
-        return_tensors="pt",
-        **chat_template_kwargs,
-    ).to(model.device)
-
-    with torch.inference_mode():
-        generated_ids = model.generate(  # pyright: ignore[reportAttributeAccessIssue]
-            **inputs,
-            max_new_tokens=4096,
-            do_sample=False,
-        )
-
-    generated_ids = generated_ids[:, inputs.input_ids.shape[1] :]
-    return processor.batch_decode(
-        generated_ids,
-        skip_special_tokens=True,
-        clean_up_tokenization_spaces=False,
-    )[0].strip()
-
-
-sample_path = (
-    Path(__file__).parents[1] / "data" / "samples" / "Screenshot 2026-07-18 171416.png"
-)
-sample_image = Image.open(sample_path).convert("RGB")
-messages = [
-    {
-        "role": "user",
-        "content": [
-            {
-                "type": "image",
-                "image": sample_image,
-            }
-        ],
-    }
-]
-
-output = run_nuextract(
-    messages,
-    mode="content",
-    enable_thinking=False,
-)
-print(output)

gradio_app.py

@@ -1,10 +0,0 @@
-from runpy import run_path
-
-app = run_path("app/gradio.py")
-demo = app["demo"]
-
-__all__ = ["demo"]
-
-
-if __name__ == "__main__":
-    demo.launch()

launch.py

@@ -1,28 +0,0 @@
-import argparse
-import runpy
-
-
-MODULES = {
-    "gradio": "app.gradio",
-    "hf_space": "app.hf_space",
-    "nuextract": "app.nuextract",
-}
-
-
-def build_parser() -> argparse.ArgumentParser:
-    parser = argparse.ArgumentParser(description="Launch a NuMark application.")
-    parser.add_argument(
-        "app",
-        choices=MODULES,
-        help="Application to launch.",
-    )
-    return parser
-
-
-def main() -> None:
-    args = build_parser().parse_args()
-    runpy.run_module(MODULES[args.app], run_name="__main__")
-
-
-if __name__ == "__main__":
-    main()

model.py

@@ -0,0 +1,108 @@
+# ruff: noqa: I001
+
+import spaces
+
+import json
+
+import torch
+from transformers import AutoModelForImageTextToText, AutoProcessor
+
+
+processor = AutoProcessor.from_pretrained(
+    "numind/NuExtract3",
+    trust_remote_code=True,
+)
+
+model = (
+    AutoModelForImageTextToText.from_pretrained(
+        "numind/NuExtract3",
+        attn_implementation="sdpa",
+        dtype=torch.bfloat16,
+        trust_remote_code=True,
+    )
+    .to("cuda")
+    .eval()
+)
+
+
+@spaces.GPU
+def extract(image, text, template, enable_thinking):
+    inputs = processor.apply_chat_template(
+        [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image", "image": image},
+                    {"type": "text", "text": text},
+                ],
+            }
+        ],
+        add_generation_prompt=True,
+        tokenize=True,
+        return_dict=True,
+        return_tensors="pt",
+        mode="structured",
+        template=template,
+        enable_thinking=enable_thinking,
+    ).to(model.device)
+
+    with torch.inference_mode():
+        generated_ids = model.generate(
+            **inputs,
+            max_new_tokens=4096,
+            do_sample=False,
+        )
+
+    return json.loads(
+        processor.batch_decode(
+            generated_ids[:, inputs.input_ids.shape[1] :],
+            skip_special_tokens=True,
+            clean_up_tokenization_spaces=False,
+        )[0].strip()
+    )
+
+
+@spaces.GPU
+def generate_template(image):
+    inputs = processor.apply_chat_template(
+        [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image", "image": image},
+                    {
+                        "type": "text",
+                        "text": (
+                            "Create a reusable structured extraction template grounded only "
+                            "in the visible document. Include fields supported by the document, "
+                            "represent repeated records as arrays, use NuExtract template leaf "
+                            "types, and return only the JSON template."
+                        ),
+                    },
+                ],
+            }
+        ],
+        add_generation_prompt=True,
+        tokenize=True,
+        return_dict=True,
+        return_tensors="pt",
+        mode="template-generation",
+    ).to(model.device)
+
+    with torch.inference_mode():
+        generated_ids = model.generate(
+            **inputs,
+            max_new_tokens=4096,
+            do_sample=False,
+        )
+
+    return json.dumps(
+        json.loads(
+            processor.batch_decode(
+                generated_ids[:, inputs.input_ids.shape[1] :],
+                skip_special_tokens=True,
+                clean_up_tokenization_spaces=False,
+            )[0].strip()
+        ),
+        indent=2,
+    )