Upgrade GRACE-VLM to a live ZeroGPU demo

README.md

@@ -3,10 +3,16 @@ title: GRACE-VLM
 emoji: 🦢
 colorFrom: blue
 colorTo: purple
-sdk: static
-app_file: index.html
+sdk: gradio
+python_version: "3.10"
+sdk_version: 5.49.1
+app_file: app.py
 pinned: true
-short_description: Deployable INT4 Qwen3-VL through GRACE distillation.
+short_description: Try GRACE-VLM on an image, then deploy the INT4 build.
+suggested_hardware: zero-a10g
+startup_duration_timeout: 1h
+preload_from_hub:
+  - ForeverBlue/Qwen3-VL-2B-GRACE-BF16
 models:
   - ForeverBlue/Qwen3-VL-2B-GRACE-W4G128-AWQ
   - ForeverBlue/Qwen3-VL-2B-GRACE-BF16
@@ -21,10 +27,11 @@ tags:
 
 # GRACE-VLM
 
-Project showcase for **GRACE-VLM: INT4 Quantization-Aware Distillation for
+Live demo for **GRACE-VLM: INT4 Quantization-Aware Distillation for
 Vision-Language Models**, accepted at ICML 2026. Read the paper at
 [arXiv:2601.22709](https://arxiv.org/abs/2601.22709).
 
-This free static Space presents the benchmark results, a model output, and the
-copy-ready real INT4 deployment command. Interactive inference requires hosted
-GPU hardware; the model and loader remain fully available for local deployment.
+Upload an image and ask a question to run the GRACE 2B BF16 checkpoint on free
+ZeroGPU hardware. For genuine packed INT4 inference, use
+[`ForeverBlue/Qwen3-VL-2B-GRACE-W4G128-AWQ`](https://huggingface.co/ForeverBlue/Qwen3-VL-2B-GRACE-W4G128-AWQ)
+with the copy-ready [GRACE loader](https://github.com/ForeverBlue816/GRACE#quick-start-real-int4).

app.py

@@ -0,0 +1,102 @@
+import threading
+
+import gradio as gr
+import spaces
+import torch
+from transformers import AutoModelForMultimodalLM, AutoProcessor
+
+
+MODEL_ID = "ForeverBlue/Qwen3-VL-2B-GRACE-BF16"
+
+processor = AutoProcessor.from_pretrained(MODEL_ID)
+model = AutoModelForMultimodalLM.from_pretrained(
+    MODEL_ID,
+    dtype=torch.bfloat16,
+    low_cpu_mem_usage=True,
+).to("cuda").eval()
+generation_lock = threading.Lock()
+
+
+@spaces.GPU(duration=120)
+def answer(image, prompt, max_new_tokens):
+    if image is None:
+        raise gr.Error("Please upload an image first.")
+    if not prompt or not prompt.strip():
+        raise gr.Error("Please enter a question about the image.")
+
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image", "image": image},
+                {"type": "text", "text": prompt.strip()},
+            ],
+        }
+    ]
+    inputs = processor.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+        tokenize=True,
+        return_dict=True,
+        return_tensors="pt",
+    ).to(model.device)
+
+    with generation_lock, torch.inference_mode():
+        outputs = model.generate(
+            **inputs,
+            max_new_tokens=int(max_new_tokens),
+            do_sample=False,
+        )
+
+    prompt_length = inputs["input_ids"].shape[-1]
+    return processor.decode(
+        outputs[0][prompt_length:],
+        skip_special_tokens=True,
+        clean_up_tokenization_spaces=False,
+    ).strip()
+
+
+with gr.Blocks(title="GRACE-VLM Demo") as demo:
+    gr.Markdown(
+        """
+        # 🦢 GRACE-VLM
+        **An ICML 2026 Qwen3-VL-2B student distilled from an 8B teacher.**
+
+        Upload an image and ask a question. This online demo runs the GRACE
+        BF16 checkpoint on ZeroGPU. The primary deployment checkpoint uses
+        [real AWQ-packed INT4](https://huggingface.co/ForeverBlue/Qwen3-VL-2B-GRACE-W4G128-AWQ)
+        and retains 98% of the GRACE BF16 benchmark average.
+        """
+    )
+    with gr.Row():
+        image_input = gr.Image(type="pil", label="Upload an image")
+        with gr.Column():
+            prompt_input = gr.Textbox(
+                value="Describe this image in detail.",
+                label="Question",
+                lines=3,
+            )
+            token_input = gr.Slider(
+                minimum=32,
+                maximum=256,
+                value=128,
+                step=32,
+                label="Maximum new tokens",
+            )
+            run_button = gr.Button("Run GRACE", variant="primary")
+    output = gr.Textbox(label="GRACE response", lines=10)
+    gr.Markdown(
+        "[Paper](https://arxiv.org/abs/2601.22709) · "
+        "[Code](https://github.com/ForeverBlue816/GRACE) · "
+        "[Models](https://huggingface.co/collections/ForeverBlue/grace)"
+    )
+    run_button.click(
+        fn=answer,
+        inputs=[image_input, prompt_input, token_input],
+        outputs=output,
+    )
+
+
+if __name__ == "__main__":
+    demo.queue(default_concurrency_limit=1).launch()
+

requirements.txt

@@ -0,0 +1,5 @@
+accelerate==1.13.0
+pillow>=12,<13
+torch==2.8.0
+transformers==5.9.0
+