Restore native Gradio UI and use GroundingPI account model

Use the native light Gradio theme with Chinese UI and disabled Video. Preserve 9 task routes, 22 image cases, canvas controls and CJK annotation fonts. Pin GroundingPI/GroundingPI at 127b0e1c54c68f24bf769a80067521b815b695dd. Update native architecture contract and Transformers 5.7 compatibility. Local: 51 tests, release fingerprint and CPU model/processor import passed; GPU verification follows deployment.

.env.example

@@ -1,7 +1,7 @@
 # Hugging Face Space -> Settings -> Variables
 GROUNDING_BACKEND=zerogpu
-GROUNDING_MODEL_REPO=XPENG-AI/GroundingPI-4B
-GROUNDING_MODEL_REVISION=REPLACE_WITH_40_CHARACTER_MODEL_COMMIT_SHA
+GROUNDING_MODEL_REPO=GroundingPI/GroundingPI
+GROUNDING_MODEL_REVISION=127b0e1c54c68f24bf769a80067521b815b695dd
 GROUNDING_MAX_TOKENS=1024
 GROUNDING_GPU_SECONDS=120
 

README.md

@@ -1,128 +1,49 @@
 ---
-title: GroundingPI
+title: GroundingPI 视觉感知工作台
 emoji: 🎯
-colorFrom: purple
+colorFrom: blue
 colorTo: green
 sdk: gradio
 sdk_version: 5.49.1
 python_version: 3.12.12
 app_file: space_app.py
 pinned: false
-license: cc-by-nc-sa-4.0
 models:
-  - XPENG-AI/GroundingPI-4B
+  - GroundingPI/GroundingPI
 ---
 
-# GroundingPI · Precise Visual Grounding
+# GroundingPI · 视觉感知工作台
 
-**From language to spatial evidence.**
+中文视觉感知 demo。支持目标检测、密集检测、描述定位、三类点定位、视觉提示、OCR 与 GUI 定位，共 9 类任务、22 张原始案例。
 
-An interactive research workspace for locating objects, reading text and exploring
-interaction targets. Choose a task, provide an image and a prompt, then inspect
-the model's predictions directly on the image.
+模型固定为公开的 `GroundingPI/GroundingPI`，修订版本为 `127b0e1c54c68f24bf769a80067521b815b695dd`。保留模型原生处理器、聊天模板和 GAM 输出协议，中文 OCR 标签使用随项目打包的 Noto Sans CJK。
 
-Model: [XPENG-AI/GroundingPI-4B](https://huggingface.co/XPENG-AI/GroundingPI-4B).
-This private preview is hosted at `FanLR/GroundingPI-4B`. Two real HF ZeroGPU grounding requests passed on 2026-09-16; full task coverage
-remains pending. Licensing and asset permissions are pending:
-see [license status](LICENSE_PENDING.md).
+页面支持图像上传、案例切换、画布缩放、结果聚焦、逐实例查看、原图局部放大、图片下载、结构化坐标与原始输出。视频入口保留为不可选状态。案例提示词和坐标协议与原 demo 一致。
 
-## Explore visual evidence
-
-| Task | Interaction |
-| --- | --- |
-| Grounding | Enter object categories separated by commas |
-| Dense | Localize instances in crowded or small-object scenes |
-| Referring | Describe a particular object or person |
-| Dense Point | Locate multiple targets as image-relative points |
-| Referring Point in Mask | Predict a point inside a target described by a referring expression |
-| Dense Point in Mask | Predict an interior point for each matching instance |
-| Visual Prompt | Supply an input region as a visual condition |
-| OCR | Read text with its associated location |
-| GUI | Point to an interface target; no UI action is executed |
-
-1. Select a task and a curated example, or upload your own image.
-2. Write the request in the prompt field and select **Run Grounding**.
-3. Select a category to focus individual instances. Use next/previous, zoom and
-   fit-to-canvas; press Escape to leave focus.
-4. Inspect the compiled prompt, normalized coordinates and raw GAM tokens.
-
-Predictions are drawn directly on the canvas. There is no original/model/reference
-view switch. Large regions use unfilled outlines; labels avoid obscuring small
-targets. The renderer does not delete, merge or relocate predictions for appearance.
-Visual Prompt's marked region is an **input condition**, not a model prediction.
-
-The gallery contains 22 original images from the selected candidate pool, with
-their corresponding task prompts. Original image formats and dimensions are preserved.
-Examples are qualitative scenarios selected from training pools, not a held-out
-benchmark. An operational demo is not evidence of a reported accuracy or speedup.
-
-## Run locally
-
-Use a separate Python 3.12 environment:
+## 本地预览
 
 ```bash
 python3 -m venv .venv
 .venv/bin/python -m pip install -r requirements.txt
-.venv/bin/python -m pip check
-.venv/bin/python -m unittest discover -s tests -v
-.venv/bin/python scripts/verify_ui.py
-bash scripts/start.sh
+GROUNDING_BACKEND=http DEMO_PORT=7862 .venv/bin/python space_app.py
 ```
 
-Without a model source, the page is an explicitly disconnected UI preview.
-For a compatible NVIDIA GPU, set `GROUNDING_MODEL_PATH=/absolute/model/directory`
-before startup. Do not also set `GROUNDING_MODEL_REPO`. Local execution does not
-automatically allocate Hugging Face GPU resources.
+预览默认 HTTP 地址为 `http://127.0.0.1:8000/v1`。未启动模型服务时，页面明确显示“推理未连接”，可以浏览图像与案例，不会展示伪造预测。已有推理服务可通过 `GROUNDING_API_BASE` 和 `GROUNDING_MODEL` 配置。
 
-The model is custom Oxygen-VL-v3 with a Kimi-K3 vision component, not a standard
-Qwen3-VL package. Keep the checkpoint's tokenizer, processor, template and custom
-code. Native generation uses BF16, SDPA and greedy decoding; vLLM/SGLang are not
-required for this Space runtime.
+## ZeroGPU 部署
 
-## Deploy a private Hugging Face Space
+上传本目录内容到已有的 Gradio / ZeroGPU Space。本次目标为 `FanLR/GroundingPI-4B`，使用该账号已有的算力。模型仓库公开，不需要新建访问令牌。
 
-Follow [the deployment guide](docs/HF_SPACE_DEPLOYMENT.md). Upload the contents of
-this directory to a **Gradio Space**, not the parent release workspace and not a ZIP.
+`space_app.py` 固定加载上述模型版本，并替换原 Space 的旧模型变量。不要���时设置本地模型目录与远程模型仓库。
 
-Configure `GROUNDING_BACKEND=zerogpu`, `GROUNDING_MODEL_REPO` and
-`GROUNDING_MODEL_REVISION` with the reviewed full 40-character model commit SHA.
-Store a scoped read-only `HF_TOKEN` as a Space **Secret** for private weights.
-Do not put credentials in code or Variables.
+Static Space 无法在服务器加载模型。没有 ZeroGPU 权限的新账号可以发布前端，但完整 Gradio 应用需要计算后端。本次直接复用原 FanLR Space，避免跨账号访问私有服务的问题。
 
-The model is initialized once at process startup. ZeroGPU leases GPU compute for
-inference; it is not an instruction to reload model weights on every request.
-See [runtime design](docs/ZEROGPU_DEPLOYMENT.md).
+## 验证
 
-## Verification boundary
-
-The existing PPU implementation has completed real generation and UI checks.
-This package includes automated regressions for prompts, parsing, stale-result
-clearing, error recovery and overlay geometry. **Two private HF ZeroGPU grounding smoke checks passed on 2026-09-16.**
-See [the smoke-check receipt](docs/HF_PRIVATE_SMOKE.md). Full route coverage and
-benchmark accuracy remain unverified. See [verification](docs/VERIFICATION.md).
-
-The default queue is single-concurrency with at most eight waiting requests.
-Uploads are limited to 30 MiB / 50 megapixels, and input plus output to 8,192 tokens.
-The default output budget is 1,024 tokens. Truncation is reported rather than
-presented as a complete prediction. Failed calls never fall back to annotations.
-
-## Project map
-
-```text
-space_app.py        Startup and queue
-app.py + assets/    Canvas, controls and presentation
-grounding_demo/     Native inference, prompt compiler, GAM parser and overlays
-cases/ + data/     Qualitative examples and provenance metadata
-scripts/           Launch, package checks and regression helpers
-tests/             Automated behavior and integrity checks
-docs/              Deployment, runtime and verification notes
+```bash
+.venv/bin/python -m unittest discover -s tests -v
+.venv/bin/python scripts/verify_ui.py
+.venv/bin/python scripts/build_bundle.py /absolute/path/studio.zip
 ```
 
-## Responsible release
-
-Do not upload confidential or personal images without permission. The demo is not
-a safety-critical controller, and its coordinates are not calibrated robot/world
-coordinates. Source data licenses, example redistribution, XPENG branding and final
-model/code licenses require owner review. Font and asset notices are in
-[assets/SOURCES.md](assets/SOURCES.md). Paper citation and approved benchmark results
-will be added after confirmation; they are not invented in this release candidate.
+本地验证与线上实际 GPU 推理分开记录。尚未完成的线上检查不会标注为通过。图片、模型及字体分别遵循其来源许可；见 `assets/SOURCES.md` 与模型仓库许可。

app.py

@@ -17,15 +17,15 @@ from grounding_demo.protocol import parse_gam, parse_reference_boxes
 from grounding_demo.render import draw_primitives
 
 ROUTE_META = {
-    "grounding": ("Grounding", "Open-vocabulary object localization"),
-    "dense": ("Dense", "Dense instance localization"),
-    "referring": ("Referring", "Language-guided referring"),
-    "dense_point": ("Dense Point", "Dense point prediction"),
-    "referring_point_in_mask": ("Referring Point in Mask", "Point inside a referred target"),
-    "dense_point_in_mask": ("Dense Point in Mask", "Point inside multiple targets"),
-    "visual_prompt": ("Visual Prompt", "Exemplar-guided localization"),
-    "ocr": ("OCR", "Text recognition and localization"),
-    "gui": ("GUI", "Interface element grounding"),
+    "grounding": ("目标检测", "开放词汇目标定位"),
+    "dense": ("密集检测", "密集场景实例定位"),
+    "referring": ("描述定位", "根据语言描述定位目标"),
+    "dense_point": ("密集点定位", "定位多个目标的中心点"),
+    "referring_point_in_mask": ("描述内点", "在描述目标内部预测位置"),
+    "dense_point_in_mask": ("密集内点", "在多个目标内部预测位置"),
+    "visual_prompt": ("视觉提示", "以参考区域寻找相似目标"),
+    "ocr": ("文字识别 OCR", "识别文字并定位文本区域"),
+    "gui": ("界面定位 GUI", "将操作意图映射到界面元素"),
 }
 CASES = load_cases()
 BY_ID = {row["id"]: row for row in CASES}
@@ -62,10 +62,10 @@ def query_update(canonical, route):
 
 def case_meta(row):
     return (
-        "<div class='source-meta'><span>TRAINING SOURCE</span>"
-        f"<b>{html.escape(row['source'])}</b><span>IMAGE SHA256</span>"
+        "<div class='source-meta'><span>样例来源</span>"
+        f"<b>{html.escape(row['source'])}</b><span>图像校验值</span>"
         f"<code>{row.get('image_sha256', '')[:16]}</code>"
-        "<small>Curated research examples for qualitative exploration, not a held-out benchmark. Inference receives only the image and task prompt.</small></div>"
+        "<small>精选案例用于体验不同任务。模型仅接收当前图像与任务提示词。</small></div>"
     )
 
 
@@ -90,7 +90,7 @@ def reset_image(path, case_id, route, canonical, meta):
         targets_html([]),
         "[]",
         "",
-        status_html("Image ready · Run grounding to extract visual evidence"),
+        status_html("图像已就绪 · 点击开始解析"),
         task_help_html(route),
     )
 
@@ -124,7 +124,7 @@ def choose_gallery(route, evt: gr.SelectData):
 
 def upload_image(path, route, canonical):
     if not path:
-        raise gr.Error("Select an image to continue.")
+        raise gr.Error("请先选择或上传一张图像。")
     with Image.open(path) as image:
         image.verify()
     return reset_image(
@@ -132,8 +132,8 @@ def upload_image(path, route, canonical):
         None,
         route,
         canonical,
-        "<div class='source-meta'><span>YOUR IMAGE</span><b>Uploaded image</b>"
-        "<small>Select a task and describe the visual evidence to extract.</small></div>",
+        "<div class='source-meta'><span>你的图像</span><b>已上传图像</b>"
+        "<small>选择任务，描述你希望识别或定位的内容。</small></div>",
     )
 
 
@@ -153,7 +153,7 @@ def edit_query(text, route, path):
         targets_html([]),
         "[]",
         "",
-        status_html("Prompt updated · Run grounding to refresh the overlay"),
+        status_html("提示词已更新 · 再次解析以刷新结果"),
     )
 
 
@@ -178,9 +178,9 @@ def result_bundle(path, canonical, route, raw, elapsed=None, usage=None):
     if not parsed.valid:
         state = "Protocol error"
     message = (
-        f"Complete · {len(valid_items)} visual targets"
+        f"解析完成 · {len(valid_items)} 个视觉目标"
         if parsed.valid
-        else "Output validation failed · " + "; ".join(parsed.errors)
+        else "输出校验未通过 · " + "; ".join(parsed.errors)
     )
     return dict(
         image=rendered,
@@ -203,15 +203,15 @@ def display_bundle(bundle):
 def inference_error(exc):
     timed_out = "timeout" in type(exc).__name__.lower()
     return (
-        "Request timed out. Check the connection and retry; your image and prompt are preserved."
+        "请求超时，请检查连接后重试。图像与提示词已保留。"
         if timed_out else
-        "Inference unavailable. Check the connection and retry; your image and prompt are preserved."
+        "推理暂不可用，请检查连接后重试。图像与提示词已保留。"
     )
 
 
 def run_inference(path, query_text, route):
     if not path:
-        raise gr.Error("Choose an example or upload an image first.")
+        raise gr.Error("请先选择案例或上传图像。")
     canonical = compile_query(query_text, route)
     try:
         raw, elapsed, usage = GroundingClient().infer(path, canonical)
@@ -229,21 +229,21 @@ def run_inference(path, query_text, route):
     return display_bundle(bundle)
 
 
-RUN_LABEL = "Run Grounding"
+RUN_LABEL = "开始解析"
 
 
 def run_inference_stream(path, query_text, route):
     """Expose real request state without an opaque progress mask or synthetic steps."""
     if not path:
-        raise gr.Error("Choose an example or upload an image first.")
+        raise gr.Error("请先选择案例或上传图像。")
     yield (
         path,
-        metrics_html(ROUTE_META[route][0], "Awaiting inference"),
+        metrics_html(ROUTE_META[route][0], "等待推理"),
         targets_html([], pending=True),
         "[]",
         "",
-        status_html("Awaiting inference", "busy"),
-        gr.update(value="Grounding in progress…", interactive=False),
+        status_html("等待推理", "busy"),
+        gr.update(value="正在解析…", interactive=False),
     )
     try:
         result = run_inference(path, query_text, route)
@@ -272,28 +272,16 @@ def service_status():
     ok, detail = GroundingClient().health()
     return (
         f"<div class='api-status {'online' if ok else 'offline'}' role='status' aria-live='polite'>"
-        f"<i aria-hidden='true'></i><b>{'Inference ready' if ok else 'Inference offline'}</b>"
-        f"<small>{'Model loaded. Ready to process your image.' if ok else 'Explore examples while the inference service is unavailable.'}</small></div>"
-        f"<details class='connection-detail'><summary>Runtime details</summary><code>{html.escape(detail)}</code></details>"
+        f"<i aria-hidden='true'></i><b>{'模型已就绪' if ok else '推理未连接'}</b>"
+        f"<small>{'随时开始，探索图像中的目标。' if ok else '可以浏览案例；连接推理服务后即可解析。'}</small></div>"
+        f"<details class='connection-detail'><summary>运行详情</summary><code>{html.escape(detail)}</code></details>"
     )
 
 
-CSS = (ROOT / "assets" / "paper.css").read_text(encoding="utf-8")
-CSS += "\n" + (ROOT / "assets" / "original.css").read_text(encoding="utf-8")
-CSS += "\n" + (ROOT / "assets" / "studio.css").read_text(encoding="utf-8")
-CSS = CSS.replace(
-    "__MANROPE_DATA__",
-    base64.b64encode((ROOT / "assets" / "Manrope-variable.ttf").read_bytes()).decode(),
-)
-# Gradio 5 scopes css= rules inside .contain, including media-query selectors.
-# Host sizing must live in head; workbench styling remains scoped.
-HOST_HEAD = """<style id="grounding-host-layout">
-gradio-app:has(#console) .gradio-container { width:100%!important; max-width:1440px!important; min-width:0!important; box-sizing:border-box; overflow:clip!important; padding:28px 32px 32px!important; margin:0 auto!important; }
-gradio-app:has(#console) .gradio-container>main { width:100%!important; max-width:none!important; min-width:0!important; padding:0!important; }
-gradio-app:has(#console) .gradio-container .contain { max-width:none!important; }
-gradio-app:has(#console) footer:not(.studio-footer) { display:none!important; }
-@media(max-width:760px) { gradio-app:has(#console) .gradio-container { padding:16px 14px!important; } }
-@media(max-width:430px) { gradio-app:has(#console) .gradio-container { padding:12px 10px!important; } }
+CSS = (ROOT / "assets" / "native.css").read_text(encoding="utf-8")
+HOST_HEAD = """<style>
+gradio-app:has(#console) .gradio-container {max-width:1280px!important;margin:0 auto!important;padding:24px!important;}
+@media(max-width:760px){gradio-app:has(#console) .gradio-container{padding:14px!important;}}
 </style>"""
 JS = (ROOT / "assets" / "console.js").read_text(encoding="utf-8")
 first_route = "grounding"
@@ -303,89 +291,85 @@ with gr.Blocks(
     css=CSS,
     js=JS,
     head=HOST_HEAD,
-    title="GroundingPI · Perception Studio",
-    theme=gr.themes.Base(font=["Manrope", "sans-serif"], font_mono=["ui-monospace", "monospace"]),
+    title="GroundingPI · 视觉感知工作台",
+    theme=gr.themes.Default(font=["Arial", "PingFang SC", "Microsoft YaHei", "sans-serif"], font_mono=["ui-monospace", "monospace"]),
     analytics_enabled=False,
 ) as demo:
     image_state = gr.State(str(ROOT / first_case["image"]))
     case_state = gr.State(first_case["id"])
+    gr.Markdown("# GroundingPI\n视觉定位、文字识别与交互目标探索 · 9 类任务 / 22 张案例")
     with gr.Column(elem_id="console"):
         with gr.Row(elem_id="workspace-top", equal_height=True):
             with gr.Column(scale=3, min_width=280, elem_id="controls"):
-                gr.HTML(
-                    """<div class="control-brand">
-                      <h2>Grounding<span>PI</span></h2>
-                      <p>Visual primitives for physical intelligence.</p>
-                    </div>"""
-                )
+                gr.Markdown("### 输入")
                 with gr.Column(elem_id="control-settings"):
                     with gr.Row(elem_id="control-fields"):
                         gr.HTML(
-                            """<div class="media-field"><span>Media Type</span><div>
-                              <span class="media-option is-active">Image</span>
+                            """<div class="media-field"><span>输入类型</span><div>
+                              <span class="media-option is-active" >图像</span>
                               <button class="media-option" type="button" disabled
-                                title="Video is not available">Video</button></div></div>""",
+                                title="暂未开放视频解析">视频</button></div></div>""",
                             elem_id="media-type",
                         )
                         route = gr.Dropdown(
                             [(name, key) for key, (name, _) in ROUTE_META.items()],
                             value=first_route,
-                            label="Task Type",
+                            label="选择任务",
                             filterable=False,
                             elem_id="route-select",
                         )
                     case = gr.Dropdown(
                         route_choices(first_route),
                         value=first_case["id"],
-                        label="Selected example",
+                        label="已选案例",
                         filterable=False,
                         visible=False,
                         elem_id="case-select",
                     )
                     upload = gr.UploadButton(
-                        "Upload Image", file_types=["image"], type="filepath", elem_id="upload-button"
+                        "上传图像", file_types=["image"], type="filepath", elem_id="upload-button"
                     )
-                    with gr.Accordion("Advanced Parameters", open=False, elem_id="advanced-options"):
+                    with gr.Accordion("运行配置", open=False, elem_id="advanced-options"):
                         gr.HTML(
-                            "<div class='runtime-options'><span>Model</span><b>"
-                            + html.escape(os.getenv("GROUNDING_MODEL", "grounding-pi-vlm"))
-                            + "</b><span>Max output tokens</span><b>"
+                            "<div class='runtime-options'><span>模型</span><b>"
+                            + html.escape(os.getenv("GROUNDING_MODEL", "GroundingPI"))
+                            + "</b><span>最大输出长度</span><b>"
                             + html.escape(os.getenv("GROUNDING_MAX_TOKENS", "1024"))
-                            + "</b><small>Read-only runtime configuration. Only the current image and task prompt enter inference.</small></div>"
+                            + "</b><small>当前模型的运行配置。</small></div>"
                         )
                 run_btn = gr.Button(RUN_LABEL, variant="primary", elem_id="run-button")
                 with gr.Column(elem_id="service-panel"):
-                    health = gr.HTML("<div class='api-status'>Checking inference service…</div>")
-                    health_btn = gr.Button("Refresh connection", size="sm", elem_id="health-button")
+                    health = gr.HTML("<div class='api-status'>正在检查推理连接…</div>")
+                    health_btn = gr.Button("刷新连接", size="sm", elem_id="health-button")
             with gr.Column(scale=7, min_width=400, elem_id="visual-stage"):
                 task_help = gr.HTML(task_help_html(first_route), elem_id="task-guide")
                 with gr.Row(elem_id="stage-toolbar"):
-                    status = gr.HTML(status_html("Image ready"), elem_id="stage-status")
+                    status = gr.HTML(status_html("图像已就绪"), elem_id="stage-status")
                 with gr.Column(elem_id="canvas-viewport"):
                     stage = gr.Image(
                         value=input_view(str(ROOT / first_case["image"]), first_case["prompt"], first_route),
                         type="filepath",
                         interactive=False,
                         show_label=False,
-                        height=400,
+                        height=440,
                         elem_id="stage-image",
                         show_download_button=False,
                         show_share_button=False,
                         show_fullscreen_button=False,
                     )
                     gr.HTML(
-                        """<div class="canvas-tools" aria-label="Canvas tools">
-                      <span class="image-dimensions">Image canvas</span>
+                        """<div class="canvas-tools" aria-label="画布工具">
+                      <span class="image-dimensions">图像画布</span>
                       <div class="zoom-tools">
-                        <button type="button" data-canvas="out" aria-label="Zoom out" title="Zoom out">
+                        <button type="button" data-canvas="out" aria-label="缩小" title="缩小">
                           <svg viewBox="0 0 20 20" aria-hidden="true"><path d="M5 10h10"/></svg></button>
-                        <output class="zoom-value" aria-label="Canvas zoom">100%</output>
-                        <button type="button" data-canvas="in" aria-label="Zoom in" title="Zoom in">
+                        <output class="zoom-value" aria-label="画布缩放">100%</output>
+                        <button type="button" data-canvas="in" aria-label="放大" title="放大">
                           <svg viewBox="0 0 20 20" aria-hidden="true"><path d="M5 10h10M10 5v10"/></svg></button>
                         <span class="tool-divider"></span>
-                        <button type="button" data-canvas="fit" aria-label="Fit to canvas" title="Fit to canvas">
+                        <button type="button" data-canvas="fit" aria-label="适应画布" title="适应画布">
                           <svg viewBox="0 0 20 20" aria-hidden="true"><path d="M7 3H3v4m10-4h4v4M3 13v4h4m10-4v4h-4"/></svg></button>
-                        <a class="canvas-download" download="grounding.png" aria-label="Download image" title="Download image">
+                        <a class="canvas-download" download="grounding.png" aria-label="下载图像" title="下载图像">
                           <svg viewBox="0 0 20 20" aria-hidden="true"><path d="M10 3v10m-4-4 4 4 4-4M4 14v3h12v-3"/></svg></a>
                       </div>
                     </div>""",
@@ -394,27 +378,27 @@ with gr.Blocks(
                 with gr.Row(elem_id="query-row"):
                     query = gr.Textbox(
                         value=query_from_prompt(first_case["prompt"], first_route),
-                        label="What would you like to locate?",
-                        show_label=False,
+                        label="描述你的目标",
+                        show_label=True,
                         lines=2,
                         max_lines=4,
                         placeholder=TASK_HELP[first_route][3],
                         elem_id="query-dock",
                     )
-                    gr.HTML("<p class='query-hint'>Enter to run · Shift + Enter for a new line</p>")
+                    gr.HTML("<p class='query-hint'>Enter 开始解析 · Shift + Enter 换行</p>")
                     mobile_run = gr.Button(RUN_LABEL, variant="primary", elem_id="mobile-run-button")
                 with gr.Column(elem_id="targets-panel"):
                     targets = gr.HTML(targets_html([]), elem_id="targets-content")
         with gr.Row(elem_id="review-row", equal_height=True):
             with gr.Column(scale=4, min_width=320, elem_id="sandbox-panel"):
                 gr.HTML(
-                    "<div class='film-heading'><h2>Explore Examples</h2><small>Curated visual scenarios</small></div>"
+                    "<div class='film-heading'><h2>示例图像</h2></div>"
                 )
                 gallery = gr.Gallery(
                     value=gallery_items(first_route),
                     columns=3,
                     rows=1,
-                    height=170,
+                    height=190,
                     show_label=False,
                     allow_preview=False,
                     object_fit="cover",
@@ -422,36 +406,36 @@ with gr.Blocks(
                     interactive=False,
                     elem_id="case-gallery",
                 )
-                with gr.Accordion("Data provenance", open=False, elem_id="source-details"):
+                with gr.Accordion("案例信息", open=False, elem_id="source-details"):
                     meta = gr.HTML(case_meta(first_case))
             with gr.Column(scale=6, min_width=420, elem_id="review-summary"):
                 with gr.Row(elem_id="result-columns", equal_height=True):
                     with gr.Column(scale=4, min_width=200, elem_id="metrics-panel"):
-                        gr.HTML("<h2 class='inspector-subheading'>Inference Metrics</h2>")
+                        gr.Markdown("### 解析概览")
                         metrics = gr.HTML(metrics_html(ROUTE_META[first_route][0]))
                     with gr.Column(scale=6, min_width=240, elem_id="sidebar"):
                         gr.HTML(
-                            "<div class='inspector-heading'><h2>Target Detail</h2><span class='live-mark'>Select a category below the canvas to magnify its original pixels.</span></div>"
+                            "<div class='inspector-heading'><h2>目标细节</h2><span class='live-mark'>选择结果即可放大原始图像。</span></div>"
                         )
                         gr.HTML(
-                            "<div class='evidence-detail' hidden><div class='detail-heading'><span>Detail View</span><small>Original pixels · magnified</small></div><div class='detail-image'></div><div class='detail-coordinates'></div></div>",
+                            "<div class='evidence-detail' hidden><div class='detail-heading'><span>局部细节</span><small>原始像素 · 局部放大</small></div><div class='detail-image'></div><div class='detail-coordinates'></div></div>",
                             elem_id="detail-panel",
                         )
-    with gr.Accordion("Prompt Console", open=False, elem_id="prompt-console"):
+    with gr.Accordion("模型提示词", open=False, elem_id="prompt-console"):
         prompt = gr.Textbox(
-            value=first_case["prompt"], label="COMPILED PROMPT", lines=3, interactive=False
+            value=first_case["prompt"], label="实际发送的提示词", lines=3, interactive=False
         )
         gr.HTML(
-            "<div class='console-note'>Transparent inputs: only this image and compiled prompt are sent to the model. No annotations, training answers, or expected counts are included.</div>"
+            "<div class='console-note'>只发送当前图像与提示词，模型独立生成预测结果。</div>"
         )
-    with gr.Accordion("Structured Output", open=False, elem_id="output-console"):
+    with gr.Accordion("结构化结果", open=False, elem_id="output-console"):
         normalized = gr.Code(
-            value="[]", language="json", label="PARSED COORDINATES · 0–999", interactive=False
+            value="[]", language="json", label="目标与坐标 · 0–999", interactive=False
         )
-        with gr.Accordion("Native Token Sequence", open=False):
-            raw = gr.Textbox(label="RAW GAM SPECIAL TOKENS", lines=4, interactive=False)
+        with gr.Accordion("模型原始输出", open=False):
+            raw = gr.Textbox(label="原始 GAM 序列", lines=4, interactive=False)
     gr.HTML(
-        "<footer class='studio-footer'><span>GroundingPI · Towards Physical Intelligence</span><span>Research demonstration <i>Qualitative examples, not benchmark results.</i></span></footer>"
+        "<footer class='studio-footer'><span>GroundingPI · 感知走向行动</span><span>探索目标、文字与交互位置 <i>图像 → 视觉基元</i></span></footer>"
     )
 
     reset_outputs = [

assets/console.js

@@ -43,7 +43,7 @@
     preview.style.backgroundImage = `url(${JSON.stringify(source?.src || image.currentSrc || image.src)})`;
     preview.style.backgroundSize = `${iw*scale}px ${ih*scale}px`;
     preview.style.backgroundPosition = `${preview.clientWidth/2-cx*scale}px ${preview.clientHeight/2-cy*scale}px`;
-    $('.detail-coordinates').textContent = `Instance ${instanceIndex + 1} / ${coords.length} · [${p.join(', ')}]`;
+    $('.detail-coordinates').textContent = `实例 ${instanceIndex + 1} / ${coords.length} · [${p.join(', ')}]`;
   }
   function canvasTransform() {
     const stage = $('#stage-image'); if (!stage) return;
@@ -104,21 +104,21 @@
     const caption = document.createElement('div'); caption.className = 'focus-caption';
     const label = document.createElement('span');
     label.setAttribute('role', 'status'); label.setAttribute('aria-live', 'polite');
-    const evidence = 'Prediction';
+    const evidence = '预测目标';
     label.textContent = `${evidence} · ${selected.dataset.label} · ${instanceIndex + 1} / ${coords.length}`;
     label.title = label.textContent;
     caption.append(label);
-    for (const [action, name, path] of [['prev', 'Previous instance', 'm12 5-5 5 5 5'], ['next', 'Next instance', 'm8 5 5 5-5 5']]) {
+    for (const [action, name, path] of [['prev', '上一个实例', 'm12 5-5 5 5 5'], ['next', '下一个实例', 'm8 5 5 5-5 5']]) {
       const button = document.createElement('button'); button.type = 'button'; button.dataset.instance = action;
       button.setAttribute('aria-label', name); button.title = name; button.disabled = coords.length < 2;
       const icon = svg('svg', {viewBox: '0 0 20 20', width: 16, height: 16, fill: 'none', 'aria-hidden': 'true'});
       icon.append(svg('path', {d: path, stroke: 'currentColor', 'stroke-width': 1.7}));
       button.append(icon); caption.append(button);
     }
-    const close = document.createElement('button'); close.type = 'button'; close.className = 'focus-close'; close.setAttribute('aria-label', 'Close target focus');
+    const close = document.createElement('button'); close.type = 'button'; close.className = 'focus-close'; close.setAttribute('aria-label', '关闭目标聚焦');
     const icon = svg('svg', { viewBox: '0 0 20 20', width: 16, height: 16, fill: 'none', 'aria-hidden': 'true' });
     icon.append(svg('path', { d: 'm6 6 8 8m0-8-8 8', stroke: 'currentColor', 'stroke-width': 1.7, 'stroke-linecap': 'round' }));
-    close.title = 'Close focus and return to targets';
+    close.title = '关闭聚焦，显示全部目标';
     close.append(icon); caption.append(close); stage.append(layer); $('#canvas-viewport').append(caption);
     if (activeAction) caption.querySelector(`[data-instance=${activeAction}]`)?.focus({preventScroll: true});
     showDetail(coords);
@@ -135,7 +135,7 @@
     for (const button of document.querySelectorAll('#run-button,#mobile-run-button')) {
       button.disabled = value;
       button.classList.toggle('is-waiting', value);
-      if (value) button.setAttribute('aria-label', 'Grounding in progress');
+      if (value) button.setAttribute('aria-label', '正在解析');
       else button.removeAttribute('aria-label');
     }
     for (const id of ['controls', 'sandbox-panel', 'query-dock', 'upload-button', 'canvas-chrome']) {
@@ -169,10 +169,10 @@
     if (!ready || waiting) return;
     previousRunId = $('#stage-status .stage-status')?.dataset.runId || '';
     clientPending = true;
-    clientStatus('Request submitted · Awaiting inference', 'busy'); setWaiting(true);
+    clientStatus('请求已提交 · 等待推理', 'busy'); setWaiting(true);
     watchdog = setTimeout(() => {
       if (!waiting) return;
-      clientStatus('No result received. Check the connection and refresh. The server request has not been cancelled.', 'error');
+      clientStatus('尚未收到结果，请检查连接并刷新。服务器请求仍可能在处理。', 'error');
       setWaiting(false);
     }, 330000);
   };
@@ -201,7 +201,7 @@
       clearSelection(); resetCanvas(); lastSrc = src;
     }
     const dimensions = $('.image-dimensions');
-    if (dimensions && image?.naturalWidth) dimensions.textContent = `${image.naturalWidth} × ${image.naturalHeight} px${zoom>1 ? ' / drag to pan' : ''}`;
+    if (dimensions && image?.naturalWidth) dimensions.textContent = `${image.naturalWidth} × ${image.naturalHeight} px${zoom>1 ? ' / 拖动平移' : ''}`;
     drawSelection();
   }
   function onClick(event) {

assets/native.css

@@ -0,0 +1,41 @@
+/* Native Gradio layout. Extra styles only support canvas interactions. */
+#console{gap:20px!important}#workspace-top{gap:24px!important;align-items:flex-start!important}
+#controls{flex:0 0 300px!important;min-width:0!important;gap:16px!important}
+#control-fields{flex-direction:column!important}#control-fields>.form{width:100%!important}
+.media-field{display:flex;align-items:center;justify-content:space-between;gap:14px;font-size:13px}
+.media-field>div{display:flex;gap:6px}.media-option{display:inline-flex;align-items:center;justify-content:center;min-width:54px;padding:7px 11px;border-radius:6px;font-size:13px;font-family:inherit;border:1px solid #d1d5db}
+.media-option.is-active{border-color:#2563eb;background:#eff6ff;color:#1d4ed8}.media-option:disabled{background:#f3f4f6;color:#9ca3af;cursor:not-allowed}
+#visual-stage{min-width:0!important;gap:12px!important}.task-guide strong{font-size:19px}.task-guide p{font-size:14px;line-height:1.6;margin:8px 0}.task-guide small{font-size:12px;color:#64748b}
+#stage-toolbar{min-height:24px;align-items:center;gap:0!important}.stage-status{display:flex;align-items:center;gap:8px;font-size:13px;line-height:1.6;overflow-wrap:anywhere}
+.stage-status i,.api-status i{width:7px;height:7px;border-radius:50%;background:#16a34a;flex-shrink:0}.stage-status.error i,.api-status.offline i{background:#dc2626}.stage-status.busy i{background:#2563eb;animation:pulse 1s ease-in-out infinite}.wait-clock{margin-left:auto;font-size:12px;font-variant-numeric:tabular-nums}
+#stage-toolbar.client-pending #stage-status{display:none!important}.client-notice{flex:1}
+#canvas-viewport{position:relative;gap:8px!important;overflow:hidden!important;padding:0!important;border:0!important}
+#stage-image{position:relative!important;transform-origin:center;transition:transform 160ms ease;overflow:visible!important;touch-action:pan-y}
+#stage-image img{object-fit:contain!important}#stage-image .icon-button-wrapper{display:none!important}
+#canvas-viewport.is-zoomed #stage-image{touch-action:none;cursor:grab}#canvas-viewport.is-panning #stage-image{cursor:grabbing;transition:none!important}
+.canvas-tools{display:flex;align-items:center;justify-content:space-between;gap:10px;font-size:12px;color:#64748b;padding:3px 0}
+.zoom-tools{display:flex;align-items:center;gap:3px;border:1px solid #d1d5db;border-radius:6px;padding:3px;background:white}
+.zoom-tools button,.canvas-download{width:31px;height:31px;border:0;border-radius:4px;background:white;display:grid;place-items:center;color:#475569;cursor:pointer}
+.zoom-tools button:hover,.canvas-download:hover{background:#f1f5f9}.zoom-tools button:disabled{opacity:.35;cursor:default}.zoom-tools svg{width:18px;height:18px;fill:none;stroke:currentColor;stroke-width:1.6;stroke-linecap:round;stroke-linejoin:round}
+.zoom-value{min-width:42px;text-align:center;font-size:12px}.tool-divider{width:1px;height:18px;background:#e2e8f0;margin:0 4px}
+#query-row{gap:6px!important;flex-direction:column!important}#query-dock{width:100%;flex:0 0 auto!important}.query-hint{color:#64748b;font-size:11px;margin:0}
+#targets-panel{gap:8px!important;min-width:0!important}.empty-targets{display:flex;flex-direction:column;gap:5px;color:#64748b;font-size:12px}.empty-targets svg{display:none}.empty-targets b{font-size:13px;color:#475569}
+.targets-summary{display:flex;gap:8px;align-items:baseline;font-size:13px;color:#475569}.targets-summary>span{margin-left:auto;font-size:12px;color:#64748b}.targets-summary b{font-size:18px}
+.target-chips{display:flex;flex-wrap:wrap;gap:7px;max-height:220px;overflow:auto;padding:3px}
+.target-chip{display:flex;align-items:center;gap:7px;padding:8px 10px;border:1px solid #d1d5db;border-radius:6px;background:white;color:#334155;font:13px inherit;cursor:pointer}
+.target-chip i{width:7px;height:7px;border-radius:2px;background:var(--category-color);flex-shrink:0}.target-chip span{overflow-wrap:anywhere}.target-chip b{font-size:11px;color:#64748b}.target-chip:hover{background:#f8fafc}.target-chip[aria-pressed=true]{border-color:#2563eb;background:#eff6ff}.target-chip:focus-visible{outline:2px solid #2563eb;outline-offset:2px}
+.focus-help{display:flex;align-items:center;justify-content:space-between;gap:10px;color:#64748b;font-size:12px}.clear-focus{background:none;border:0;color:#2563eb;font:inherit;cursor:pointer;text-decoration:underline}.clear-focus[hidden],.evidence-detail[hidden]{display:none!important}.targets-caption{display:none}
+.focus-layer,.focus-pixels{position:absolute;pointer-events:none}.focus-pixels{z-index:4}.focus-layer{z-index:5;overflow:visible}.focus-dimmer{fill:#0f172a;opacity:.22}.focus-outline{fill:none;stroke-width:2.5;vector-effect:non-scaling-stroke}.focus-halo{fill:none;stroke:white;stroke-width:5;vector-effect:non-scaling-stroke}
+.focus-caption{display:flex;align-items:center;gap:7px;position:relative;z-index:6;padding:7px 9px;border:1px solid #dbeafe;border-radius:6px;background:#eff6ff;color:#1e40af;font-size:12px}
+.focus-caption span{flex:1;overflow-wrap:anywhere}.focus-caption button{width:30px;height:30px;border:0;border-radius:4px;background:#dbeafe;color:#1e40af;display:grid;place-items:center;cursor:pointer}
+#review-row{gap:24px!important}#review-summary{min-width:0!important}#result-columns{gap:20px!important}#metrics-panel,#sidebar{min-width:0!important;gap:10px!important}
+.film-heading h2,.inspector-heading h2{font-size:17px;margin:0 0 8px}.film-heading small,.live-mark{font-size:12px;color:#64748b}
+.metrics-list{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:12px 20px;margin:0;font-size:13px}.metrics-list>div{display:flex;align-items:center;justify-content:space-between;gap:10px}.metrics-list dt{color:#64748b}.metrics-list dd{margin:0;color:#334155;overflow-wrap:anywhere;text-align:right}
+#sidebar:not(:has(.evidence-detail:not([hidden]))){display:none!important}.detail-heading{display:flex;align-items:center;justify-content:space-between;gap:12px;font-size:13px;margin-bottom:8px}.detail-heading small{color:#64748b;font-size:11px}.detail-image{height:190px;background-color:#f1f5f9;background-repeat:no-repeat;border-radius:6px}.detail-coordinates{font-size:11px;color:#64748b;margin-top:6px;overflow-wrap:anywhere}
+.api-status{display:flex;align-items:center;gap:7px;flex-wrap:wrap;font-size:13px}.api-status small{flex-basis:100%;font-size:12px;color:#64748b;line-height:1.6}.connection-detail{font-size:12px;color:#64748b}.connection-detail code{display:block;font-size:11px;overflow-wrap:anywhere}
+.runtime-options,.source-meta{display:grid;gap:8px;font-size:12px;color:#64748b}.source-meta b,.runtime-options b{color:#334155}.console-note{font-size:12px;color:#64748b}
+.studio-footer{display:flex;justify-content:space-between;gap:15px;padding-top:12px;font-size:11px;color:#64748b}.studio-footer i{font-style:normal;margin-left:12px;color:#94a3b8}
+#mobile-run-button{display:none!important}#console[aria-busy=true] #canvas-viewport{cursor:progress}
+@keyframes pulse{50%{opacity:.3}}
+@media(max-width:760px){#workspace-top,#review-row,#result-columns{flex-direction:column!important}#controls{width:100%;flex:0 0 auto!important}#visual-stage{width:100%}#controls>h3{margin:0}#stage-image{height:330px!important}#run-button{display:none!important}#mobile-run-button{display:flex!important;margin-top:8px!important;width:100%!important}#review-summary,#sandbox-panel{width:100%!important;flex:0 0 auto!important}.studio-footer{flex-direction:column;gap:5px}.metrics-list{font-size:12px}.canvas-tools{font-size:10px}.image-dimensions{white-space:nowrap}#service-panel{gap:6px!important}.target-chip{font-size:12px}.task-guide strong{font-size:18px}}
+@media(prefers-reduced-motion:reduce){*,*::before,*::after{animation:none!important;transition:none!important;scroll-behavior:auto!important}}

constraints-app.txt

@@ -20,7 +20,7 @@ h11==0.16.0
 hf-xet==1.6.0
 httpcore==1.0.9
 httpx==0.28.1
-huggingface_hub==1.4.1
+huggingface_hub==1.5.0
 idna==3.19
 Jinja2==3.1.6
 markdown-it-py==4.2.0
@@ -57,7 +57,7 @@ sympy==1.14.0
 tokenizers==0.22.2
 tomlkit==0.13.3
 tqdm==4.70.1
-transformers==5.3.0
+transformers==5.7.0
 typer==0.27.2
 typer-slim==0.24.0
 typing-inspection==0.4.4

docs/HF_SPACE_DEPLOYMENT.md

@@ -1,67 +1,13 @@
-# Hugging Face deployment
+# 部署到已有 FanLR ZeroGPU Space
 
-This guide describes a candidate deployment, not an already-created HF service.
-The model repository belongs to XPENG-AI; this private Space is hosted under
-FanLR. Both repository names and private deployment have been confirmed.
+目标：`FanLR/GroundingPI-4B`。SDK 为 Gradio 5.49.1，硬件保持已有 ZeroGPU。Space 可见性保持私有。
 
-## Two repositories
+1. 先运行全部单元测试、案例完整性校验及打包校验。
+2. 上传更新文件，保持项目相对路径。
+3. 模型变量：`GROUNDING_MODEL_REPO=GroundingPI/GroundingPI`，`GROUNDING_MODEL_REVISION=127b0e1c54c68f24bf769a80067521b815b695dd`，`GROUNDING_MODEL=GroundingPI`。
+4. `GROUNDING_BACKEND=zerogpu`，输出预算 1024 tokens，GPU 调用上限 120 秒。
+5. 等待构建和模型初始化完成，确认“模型已就绪”。
+6. 使用无个人信息的图片测试 9 条任务路径。中文 OCR 使用合成图片；验证中文标签、结构化文本与边框。
+7. 验证上传、切换任务/案例后清空旧预测、缩放、聚焦、实例切换、下载与错误恢复。
 
-- Model repository: contents of the sibling `model/` directory; proposed ID
-  `XPENG-AI/GroundingPI-4B`.
-- Gradio Space: contents of this `demo/` directory; proposed ID
-  `FanLR/GroundingPI-4B`.
-
-The parent release workspace includes internal review material and must not be
-uploaded. Uploading files privately is still a transfer to an external platform.
-
-## Private preview and hardware
-
-Create both repositories as **Private from the beginning**. Protected Spaces hide
-source code but expose the running application, so Protected is not a confidential
-preview mode. Organization access depends on its roles and grants; verify the
-actual reader set rather than assuming Private means only the uploader.
-[Visibility](https://huggingface.co/docs/hub/spaces-overview#space-visibility),
-[organization access](https://huggingface.co/docs/hub/organizations-security).
-
-Choose Gradio and confirm available hardware before upload. According to current
-official documentation, organization ZeroGPU hosting requires a Team/Enterprise
-plan. Personal free-tier eligibility does not establish organization eligibility.
-Do not enable paid hardware or change subscriptions without approval.
-[ZeroGPU eligibility](https://huggingface.co/docs/hub/spaces-zerogpu).
-
-## Configuration
-
-| Location | Name | Value |
-| --- | --- | --- |
-| Variable | GROUNDING_BACKEND | zerogpu |
-| Variable | GROUNDING_MODEL_REPO | Reviewed model repository ID |
-| Variable | GROUNDING_MODEL_REVISION | Full 40-character commit SHA |
-| Variable | GROUNDING_MAX_TOKENS | 1024 |
-| Variable | GROUNDING_GPU_SECONDS | 120; tune only after measurement |
-| Secret | HF_TOKEN | Read access limited to the private model |
-
-Do not set internal deployment variables or a local model path in HF.
-The repository snapshot is downloaded using its pinned revision and then loaded
-locally, because the custom processor opens local configuration files.
-Never place a token in README, source, Variables, a URL or a screenshot.
-[Secrets](https://huggingface.co/docs/hub/spaces-overview#managing-secrets-and-environment-variables).
-
-Keep the README SDK configuration: `app_file: space_app.py`,
-Gradio 5.49.1 and Python 3.12.12. Use the supplied standard NVIDIA Torch dependency
-pins, not a vendor PPU environment. The runtime imports spaces before loading
-the model; GPU work alone runs through `@spaces.GPU`.
-
-## Acceptance before wider access
-
-1. Validate build dependencies, complete model loading and memory use.
-2. Run every task family at least once; inspect actual pixels, raw tokens and EOS.
-3. Record unsuccessful calls as well as successes. Curated training examples are
-   not an independent accuracy benchmark.
-4. Verify first request, repeated requests, GPU lease release/reallocation, quota
-   exhaustion, timeout recovery and session isolation.
-5. Confirm owner access and unauthorized-user denial for both repositories and
-   the running application. A logged-out access check is part of acceptance.
-6. Pin and record model and Space revisions. Do not automatically make either public.
-
-No HF login, upload, repository creation, hardware purchase or publication is
-performed by this document. Recheck platform policies when deploying.
+新账号 GroundingPI 目前的创建页禁用了 Gradio/Docker，因此不在该账号新建计算型 Space。此次使用它的公开模型，复用 FanLR 已有 GPU 服务。

docs/VERIFICATION.md

@@ -1,44 +1,7 @@
-# Verification and evidence boundaries
+# 验证范围
 
-## Existing service evidence — 2026-09-15
+自动测试覆盖案例提示词、GAM 解析、预测状态更新、HTML 安全转义、图像上传、错误恢复、中文与长标签绘制、画布几何、打包文件和模型请求边界。
 
-The native implementation was installed in a new isolated application environment
-on a single PPU. The full model loaded and generated real outputs. A 27-example
-Gradio regression completed uploads, generation with EOS and GAM parsing without
-transport/format failures. This is not 100% semantic accuracy.
+模型包预检与 CPU 处理器测试验证 `GroundingPI/GroundingPI` 的固定版本、模型代码导入、原生聊天模板、图像张量与网格，不等同于实际 GPU 推理。
 
-The original serial PPU request window had median 5.069 s and P95 14.532 s
-(nearest rank, 27 requests). These are historical observations for that hardware
-and workload, not HF latency guarantees or paper benchmark measurements.
-
-The paper-r7 UI regression included 39 passing tests and browser verification of
-direct prediction overlays, category focus, instance navigation, zoom, stale-state
-clearing and responsive layout. Example live calls returned 27 boxes and 79 points;
-counts are output observations, not verified recall.
-
-## This release candidate
-
-The presentation and inference implementation are copied from the paper-r7 release.
-README and deployment materials have been rewritten for external publication.
-The UI fingerprint and example contracts remain locked. Model weights, tokenizer
-and image processing remain byte-identical to the source. The cleaned custom
-code has canonical Oxygen class names and no unused Mage/StreamMind branches.
-Strict loading has no missing, unexpected or mismatched keys. Before/after
-Transformers regression matched all generated tokens for all 27 examples, with
-identical preprocessing tensors. This is equivalence, not semantic accuracy.
-
-The parent release workspace contains the technical inspection report. Reported
-tests must not be read as license approval, independent benchmark evaluation, or
-approval to disclose model/data assets.
-
-## Still pending
-
-- Standard NVIDIA GPU and final HF ZeroGPU execution.
-- GPU lease capture, allocation/release, quota and long-idle cold starts.
-- Private organization access verification and external-upload authorization.
-- Independent accuracy evaluation and any paper speedup claims.
-- Final legal, model-card, example-image and branding review.
-
-The code does not run diffusion/parallel decoding just because a paper illustration
-discusses it. Publish only results verified for the exact released checkpoint,
-decoding method, hardware and evaluation protocol.
+线上 GPU 检查必须记录实际输出、完成状态、耗时和截图。9 条任务路径使用无个人信息图片；中文 OCR 使用合成文本。未执行或失败的检查应记录其具体状态，不标注为通过。

docs/ZEROGPU_DEPLOYMENT.md

@@ -1,47 +1,9 @@
-# Runtime and model contract
+# ZeroGPU 运行设计
 
-## Initialization versus requests
+`spaces` 在 Torch/Transformers 之前导入。模型在进程启动时加载并放到 CUDA，由 ZeroGPU 捕获。只在 `@spaces.GPU` 函数内执行生成与设备转换；模型不随每个请求重新加载。
 
-`space_app.py` installs the spaces CUDA hooks and initializes the model once at
-process startup. In a ZeroGPU Space, the decorated generation function leases GPU
-compute and returns CPU token IDs. It does not call from_pretrained per request.
-The same code can run on a normal compatible GPU, where the decorator is effect-free.
-[Official model-loading guidance](https://huggingface.co/docs/hub/spaces-zerogpu#model-loading).
+固定仓库 `GroundingPI/GroundingPI` 与 40 位修订 `127b0e1c54c68f24bf769a80067521b815b695dd`，使用自定义 GroundingPI 配置、MoonViT 视觉编码器和原生处理器。模型合同检查验证完整 GAM 词汇、图像 token、EOS 与权重分片。
 
-The default implementation is native Transformers, not a persistent vLLM or SGLang
-server. An explicitly configured HTTP backend exists for approved deployments, but
-no automatic backend fallback is used.
+Transformers 5.7.0，Torch 2.10.0，BF16，SDPA，贪心生成。保留特殊坐标 tokens；没有 EOS 的截断结果不展示为成功。初始化或推理失败会明确显示离线/错误状态，不回退到样例答案。
 
-## Frozen behavior
-
-- `oxygen_vl_v3` / `OxygenVLv3ForConditionalGeneration`; vision model type
-  `oxygen_k3_vision`.
-- Download a reviewed, pinned snapshot; load the checkpoint's processor from a local
-  directory with `trust_remote_code=True`.
-- Apply the original chat template to the current image and request.
-- Preserve native image processing, image token/grid alignment and GAM vocabulary.
-- BF16, SDPA, greedy generation and `use_cache=True`; native EOS from generation config.
-- Preserve `skip_special_tokens=False` and `spaces_between_special_tokens=False`.
-- Reject missing pixels/grid, context overflow, empty answers and missing EOS.
-- Do not send annotations, reference counts or previous results to the model.
-- Do not claim token-identical output or equal latency across hardware/backends.
-
-The model package contains optional StreamMind source for provenance; this demo does
-not enable that branch or require mamba-ssm. Do not copy a training environment into
-the Space or install vendor-specific Torch wheels.
-
-## Limits and failures
-
-Default 1,024 new tokens; supported UI/runtime range 1–2,048. Input plus requested
-output must not exceed 8,192 tokens. GPU duration is a maximum lease, not a latency
-promise. Queue waits and initialization must be distinguished from generation.
-
-Incomplete output is a failure, not a correct empty result. On exceptions, the UI
-clears stale predictions, restores its controls and exposes a recoverable error.
-The example annotations are never used as fallback predictions.
-
-## Deployment boundary
-
-PPU service verification does not validate NVIDIA CUDA, ZeroGPU capture/leases,
-platform quotas or an organization's permissions. These are final-platform checks.
-See [HF deployment](HF_SPACE_DEPLOYMENT.md) and [verification](VERIFICATION.md).
+公开模型不需要访问秘密。GPU 配额与排队由既有 FanLR Space 管理；页面改版不改变 Space 可见性和硬件。

docs/ui_case_pool_22_lock.json

@@ -1,13 +1,14 @@
 {
   "files": {
-    "app.py": "79282810a5f15138950ffb9d2ff108067f7f73cf52c797a1931920e1fbff40b8",
+    "app.py": "11ba24d8264b7a5b5bb9c3aafbb45df3d7ef36adaf556938a74e731e59885dc5",
     "assets/Manrope-OFL.txt": "e01b637272e0cbdfb240184dd98ea5cc671556d9894dae2668d92ab2c906787c",
     "assets/Manrope-variable.ttf": "d0639be45d0af36e798172419d7bd173c4bd4f29e2b76cbb69db1d11bf8b0a40",
     "assets/NotoSansCJK-OFL.txt": "6a73f9541c2de74158c0e7cf6b0a58ef774f5a780bf191f2d7ec9cc53efe2bf2",
     "assets/NotoSansCJKsc-Regular.otf": "2c76254f6fc379fddfce0a7e84fb5385bb135d3e399294f6eeb6680d0365b74b",
     "assets/SOURCES.md": "0299a8a04951a07c896b28168496209ff900e60f6ae0c98bfdc86a7cc47c1712",
-    "assets/console.js": "3af36a3aeafb72e85018b1c663f624b92877bad0152ee9e264a1c440d689f60e",
+    "assets/console.js": "822e670f29a584ba201dc69e76cdbc6a2b836bee99b8dd1bc806aabdf3c7f5ec",
     "assets/frosted.css": "393bf5f1c04a49a6e9b17e855e573646489b4add85bc7973eb6bb277686b5c80",
+    "assets/native.css": "24b80e61a37ca92f62a871f30114bf2839290abf9c899242033b1aeef3998c38",
     "assets/original.css": "7a244c7f65e40797850dcde7d6b6505b1b5de60aa24b129f44b33ae52291fd2e",
     "assets/paper.css": "61fc60bf8c379a0894fbfe06574964ec2ed256791a23860fa4dbb872eb2869f9",
     "assets/studio.css": "17a4d588ae8e1099d54043c10e3bc83c9182c33bcf1dda1914053040d4b63fc0",
@@ -35,7 +36,7 @@
     "cases/visual_prompt/21_visual_prompt_round5_candidates_gam_visual_dense200_142.jpg": "34101e5a696ca089b3d6026fbafda21ba8a144bb2b4d4f36fd309273ed48e286",
     "cases/visual_prompt/22_visual_prompt_round5_candidates_gam_visual_dense200_187.jpg": "3daa8e2776a2dd0ab17abebd669d95728f64bfbd063986f14becfb3f1966463a",
     "grounding_demo/catalog.py": "75e543fed7f0ca018fb41c4b528efc2901aa6a0f05232959297d6d40798f676f",
-    "grounding_demo/console.py": "5754564abe58099c9406de639663d9d1416b96bc65d7a89b24f402ec41978fb8",
+    "grounding_demo/console.py": "3097928c7c6a77d5d4d379cbc298222e23d2751e5269e5efad39a71e8050c311",
     "grounding_demo/protocol.py": "cf043c9f026a1dcd63d55c5393f0d0921cd15eae4a0cc9a42e2834fab23486ec",
     "grounding_demo/render.py": "3b3fe5901edc41586bf5f1a1c263265d47c0fbe4d2ee4c30697c1bd0f78d484b"
   },

grounding_demo/console.py

@@ -23,60 +23,15 @@ CATEGORY_ROUTES = {"grounding", "dense", "layout", "dense_point_in_mask"}
 
 # Presentation only: never alter the released model prompt or expected answer.
 TASK_HELP = {
-    "grounding": [
-        "Open-Vocabulary Grounding",
-        "Turn category names into visual primitives: what is present, and where it is.",
-        "Separate categories with commas. Select a result to inspect its original pixels.",
-        "car, person"
-    ],
-    "dense": [
-        "Grounding in Dense Scenes",
-        "Explore object-level perception in crowded scenes, including small and overlapping instances.",
-        "Enter object categories. Inspect overlapping predictions one instance at a time.",
-        "person, chair"
-    ],
-    "referring": [
-        "Language, Grounded in Context",
-        "Link descriptions of appearance, position, and relationships to a specific visual target.",
-        "Describe appearance, position, or relationships in a complete phrase.",
-        "The person standing beside the table"
-    ],
-    "dense_point": [
-        "Dense Object Pointing",
-        "Represent multiple objects as localized points: a compact form of visual grounding.",
-        "Describe the targets to locate. Points indicate positions, not object boundaries.",
-        "All visible people"
-    ],
-    "referring_point_in_mask": [
-        "Point Inside a Referred Target",
-        "Use a description to predict a point inside the target object.",
-        "Describe the target. The output is an image-relative point inside the object.",
-        "The person wearing a flotation device"
-    ],
-    "visual_prompt": [
-        "Grounding by Visual Example",
-        "Use spatial exemplars to identify related instances through a shared grounding interface.",
-        "Green INPUT boxes come from your prompt, not model predictions. Coordinates use the 0–999 range.",
-        "Given reference boxes <|box_start|><100><100><200><200><|box_end|> ..."
-    ],
-    "ocr": [
-        "Reading with Spatial Context",
-        "Bind recognized text to its image region, connecting what is read with where it appears.",
-        "Use an OCR instruction. Select a transcription to inspect the original pixels.",
-        "Read and locate the text in the image"
-    ],
-    "dense_point_in_mask": [
-        "Point Inside Multiple Targets",
-        "Predict an interior point for each matching instance in a crowded scene.",
-        "Separate categories with commas. Each prediction marks a point inside an instance.",
-        "van, bus, truck, car, pedestrian"
-    ],
-    "gui": [
-        "Grounding Intent in Interfaces",
-        "Map an interaction instruction to a visual target within an interface.",
-        "Describe the element to select. The demo predicts a point; it does not execute clicks.",
-        "Click the search button"
-    ]
+    "grounding": ["语言所指，清晰可见", "从一句描述中发现目标，以精确边框呈现它的位置。", "用逗号分隔类别；选择结果可查看原始像素。", "汽车，行人"],
+    "dense": ["密集场景，逐一看清", "探索拥挤场景里的小目标、遮挡物和多个实例。", "输入目标类别，再逐个查看检测结果。", "行人，椅子"],
+    "referring": ["把描述，落在图像上", "用外观、位置或关系，锁定你想找的那个目标。", "描述目标的特征与位置，例如颜色、衣着或相邻物体。", "站在桌子旁边的人"],
+    "dense_point": ["多个目标，一点即明", "以轻量的坐标点表达多个对象的位置。", "描述需要定位的目标，点表示位置而非物体边界。", "所有可见的行人"],
+    "referring_point_in_mask": ["在目标内部，找到落点", "根据文字描述，在目标对象内部预测一个位置。", "描述目标，结果将以图像内的坐标点呈现。", "穿着救生衣的人"],
+    "dense_point_in_mask": ["复杂场景里的每个落点", "为同类目标逐一预测位于对象内部的位置。", "用逗号分隔类别，每个点对应一个目标实例。", "货车，公交车，汽车，行人"],
+    "visual_prompt": ["给出参照，发现相似", "以参考框为线索，在图像中寻找相似对象。", "绿色 INPUT 框来自你的输入；坐标范围为 0–999。", "给出参考框，再描述需要寻找的相似目标"],
+    "ocr": ["读懂文字，也找到它", "将识别的文字与所在区域连接，完整保留中文与标点。", "输入文字识别指令；选择一条结果即可检查原始图像。", "OCR task detect all the text in box format."],
+    "gui": ["让交互意图，找到位置", "把你的操作描述映射为界面上的目标位置。", "描述要选择的元素；此功能仅定位，不执行点击。", "点击搜索按钮"]
 }
 
 
@@ -111,12 +66,12 @@ def compile_query(query: str, route: str) -> str:
 
 def metrics_html(route, state="Ready", count=0, elapsed=None, usage=None):
     rows = [
-        ("Status", state),
-        ("Task", route),
-        ("Model", "grounding-pi-vlm"),
-        ("Detections", str(count)),
-        ("Latency", f"{elapsed:.2f} s" if elapsed is not None else "—"),
-        ("Output tokens", str((usage or {}).get("completion_tokens", "—"))),
+        ("状态", {"Ready": "就绪", "Completed": "已完成", "Protocol error": "协议错误", "API error": "连接异常", "等待推理": "等待推理"}.get(state, state)),
+        ("任务", route),
+        ("模型", "GroundingPI"),
+        ("目标数", str(count)),
+        ("耗时", f"{elapsed:.2f} s" if elapsed is not None else "—"),
+        ("输出 tokens", str((usage or {}).get("completion_tokens", "—"))),
     ]
     return (
         "<dl class='metrics-list'>"
@@ -127,11 +82,11 @@ def metrics_html(route, state="Ready", count=0, elapsed=None, usage=None):
 
 def targets_html(items, pending=False, image_source=""):
     if not items:
-        title = "Awaiting visual evidence" if pending else "Explore a visual target"
+        title = "正在寻找视觉线索" if pending else "让图像给出答案"
         hint = (
-            "Your image stays visible while the model processes the request."
+            "解析进行中，当前图像会保持显示。"
             if pending
-            else "Choose an example, run grounding, then select a category to inspect."
+            else "选择案例或上传图像，开始解析后在这里探索结果。"
         )
         return (
             "<div class='empty-targets'><svg viewBox='0 0 40 40' width='40' height='40' "
@@ -143,18 +98,18 @@ def targets_html(items, pending=False, image_source=""):
     counts = Counter(item["label"] for item in items)
     chips = "".join(
         "<button type='button' class='target-chip' aria-pressed='false' "
-        f"aria-label='{html.escape(f'Inspect {label}, {count} target' + ('s' if count != 1 else ''), quote=True)}' "
+        f"aria-label='{html.escape(f'查看 {label}，共 {count} 个目标', quote=True)}' "
         f"data-label='{html.escape(label, quote=True)}' "
         f"data-coordinates='{html.escape(json.dumps([item['coordinates'] for item in items if item['label'] == label]), quote=True)}'>"
         f"<i aria-hidden='true' style='--category-color:rgb{PALETTE[idx % len(PALETTE)]}'></i><span>{html.escape(label)}</span><b>{count}</b></button>"
         for idx, (label, count) in enumerate(counts.items())
     )
-    caption = "Live model predictions · select a target or category to inspect"
+    caption = "模型实时预测 · 选择目标查看细节"
     return (
         (f"<img class='focus-source' hidden alt='' src='{html.escape(image_source, quote=True)}'>" if image_source else "")
-        + f"<div class='targets-summary'><b>{len(items)}</b> {'target' if len(items) == 1 else 'targets'} <span>{len(counts)} {'category' if len(counts) == 1 else 'categories'}</span></div>"
+        + f"<div class='targets-summary'><b>{len(items)}</b> 个目标 <span>{len(counts)} 个类别</span></div>"
         f"<div class='target-chips'>{chips}</div>"
-        "<div class='focus-help'><span>Inspect individual instances</span>"
-        "<button type='button' class='clear-focus' hidden>Show all</button></div>"
+        "<div class='focus-help'><span>选择目标，查看局部细节</span>"
+        "<button type='button' class='clear-focus' hidden>显示全部</button></div>"
         f"<div class='targets-caption'>{caption}</div>"
     )

grounding_demo/model_spec.py

@@ -0,0 +1,4 @@
+"""Public model release selected for this independent demo."""
+MODEL_REPO = "GroundingPI/GroundingPI"
+MODEL_REVISION = "127b0e1c54c68f24bf769a80067521b815b695dd"
+MODEL_NAME = "GroundingPI"

grounding_demo/native_runtime.py

@@ -43,21 +43,22 @@ def resolve_snapshot() -> Path:
 
 
 def inspect_contract(root: Path, *, require_weights: bool = True) -> dict:
-    """Check the custom Kimi-K3-ViT package without importing its executable code."""
+    """Validate the public GroundingPI release without executing model code."""
     required = [
         "config.json", "preprocessor_config.json", "generation_config.json",
         "tokenizer.json", "tokenizer_config.json", "chat_template.jinja",
-        "configuration_oxygen_vl_v3.py",
-        "modeling_oxygen_vl_v3.py", "modeling_kimi_k3.py",
-        "processing_oxygen_vl_v3.py", "kimi_k3_vision_processing.py", "media_utils.py",
+        "configuration_groundingpi.py", "configuration_groundingpi_vision.py",
+        "modeling_groundingpi.py", "modeling_groundingpi_vision.py",
+        "processing_groundingpi.py", "image_processing_groundingpi.py", "media_utils.py",
+        "streammind_gate.py",
     ]
     missing = [name for name in required if not (root / name).is_file()]
     if missing:
         raise ConfigurationError(f"Model package is missing files: {', '.join(missing)}")
     config = json.loads((root / "config.json").read_text())
-    if config.get("model_type") != "oxygen_vl_v3":
-        raise ConfigurationError("This runtime requires the custom Oxygen-VL-v3 model, not standard Qwen3-VL weights.")
-    if config.get("vision_config", {}).get("model_type") != "oxygen_k3_vision":
+    if config.get("model_type") != "groundingpi":
+        raise ConfigurationError("This demo requires the GroundingPI/GroundingPI model package.")
+    if config.get("vision_config", {}).get("model_type") != "groundingpi_vision":
         raise ConfigurationError("The model does not use the expected Kimi-K3 ViT.")
     if not config.get("auto_map", {}).get("AutoModelForImageTextToText"):
         raise ConfigurationError("The model is missing the AutoModelForImageTextToText mapping.")
@@ -93,7 +94,7 @@ def inspect_contract(root: Path, *, require_weights: bool = True) -> dict:
             raise ConfigurationError("Model weight shards are incomplete. Upload the weights as well as the tokenizer and configuration.")
     prep = json.loads((root / "preprocessor_config.json").read_text())
     return {
-        "architecture": config["architectures"], "vision": "oxygen_k3_vision",
+        "architecture": config["architectures"], "vision": "groundingpi_vision",
         "config_sha256": hashlib.sha256((root / "config.json").read_bytes()).hexdigest(),
         "processor_sha256": hashlib.sha256((root / "preprocessor_config.json").read_bytes()).hexdigest(),
         "media_proc_cfg": prep.get("media_proc_cfg"), "gam_coordinate_tokens": 1000,

requirements.txt

@@ -20,7 +20,7 @@ h11==0.16.0
 hf-xet==1.6.0
 httpcore==1.0.9
 httpx==0.28.1
-huggingface-hub==1.4.1
+huggingface-hub==1.5.0
 idna==3.19
 Jinja2==3.1.6
 markdown-it-py==4.2.0
@@ -59,7 +59,7 @@ tomlkit==0.13.3
 torch==2.10.0
 torchvision==0.25.0
 tqdm==4.70.1
-transformers==5.3.0
+transformers==5.7.0
 typer==0.27.2
 typer-slim==0.24.0
 typing_extensions==4.16.0

space_app.py

@@ -1,9 +1,15 @@
-"""HF Gradio entrypoint. The original app.py and visual assets stay unchanged."""
+"""Chinese GroundingPI Studio with a pinned public checkpoint."""
 import os
 
 from grounding_demo.client import zerogpu_configured
+from grounding_demo.model_spec import MODEL_REPO, MODEL_REVISION, MODEL_NAME
 
 os.environ.setdefault("GROUNDING_BACKEND", "zerogpu")
+if not os.getenv("GROUNDING_MODEL_PATH"):
+    # This release deliberately replaces the older Space's model variables.
+    os.environ["GROUNDING_MODEL_REPO"] = MODEL_REPO
+    os.environ["GROUNDING_MODEL_REVISION"] = MODEL_REVISION
+os.environ["GROUNDING_MODEL"] = MODEL_NAME
 os.environ["GROUNDING_BACKEND"] = os.environ["GROUNDING_BACKEND"].strip().lower()
 os.environ.setdefault("GRADIO_ANALYTICS_ENABLED", "False")
 

tests/test_console.py

@@ -44,12 +44,12 @@ class ConsoleTest(unittest.TestCase):
                 self.assertEqual(len(choices["choices"]), len(app.BY_ROUTE[route]))
                 self.assertEqual(len(gallery["value"]), len(app.BY_ROUTE[route]))
                 self.assertEqual(loaded[8], "[]")
-                self.assertIn("Image ready", loaded[10])
+                self.assertIn("图像已就绪", loaded[10])
 
     def test_case_switch_resets_prediction(self):
         output = app.load_case(app.CASES[1]["id"])
         self.assertEqual(output[8], "[]")
-        self.assertIn("Image ready", output[10])
+        self.assertIn("图像已就绪", output[10])
 
     def test_upload_cannot_reveal_unrelated_gt(self):
         row = app.CASES[0]
@@ -70,7 +70,7 @@ class ConsoleTest(unittest.TestCase):
             str(app.ROOT / row["image"]), row["prompt"], row["route"], valid + invalid
         )
         self.assertEqual(result["normalized"], "[]")
-        self.assertIn("Protocol error", result["metrics"])
+        self.assertIn("协议错误", result["metrics"])
 
     def test_actual_request_uses_latest_query_without_reference(self):
         row = app.CASES[0]
@@ -81,7 +81,7 @@ class ConsoleTest(unittest.TestCase):
         infer.assert_called_once_with(path, compile_query("car", row["route"]))
         self.assertIsInstance(output[0], app.Image.Image)
         self.assertEqual(len(output), 6)
-        self.assertIn("1 visual targets", output[-1])
+        self.assertIn("1 个视觉目标", output[-1])
 
     def test_labels_are_html_escaped(self):
         self.assertIn(
@@ -97,7 +97,7 @@ class ConsoleTest(unittest.TestCase):
         ])
         self.assertIn(label, result["normalized"])
         self.assertIn(label, result["targets"])
-        self.assertIn("Complete", result["status"])
+        self.assertIn("解析完成", result["status"])
 
     def test_focus_buttons_carry_exact_safe_geometry(self):
         class Buttons(HTMLParser):
@@ -138,7 +138,7 @@ class ConsoleTest(unittest.TestCase):
             self.assertIn("data-phase='busy'", pending[-2])
             done = next(stream)
             self.assertTrue(done[-1]["interactive"])
-            self.assertIn("1 visual targets", done[-2])
+            self.assertIn("1 个视觉目标", done[-2])
             infer.assert_called_once_with(path, compile_query("car", row["route"]))
 
     def test_stream_error_releases_action_without_stale_prediction(self):
@@ -178,14 +178,14 @@ class ConsoleTest(unittest.TestCase):
         row = app.CASES[0]
         with patch.object(app.GroundingClient, "infer", side_effect=ConnectionError("test detail")):
             result = app.run_inference(str(app.ROOT / row["image"]), "car", "grounding")
-        self.assertIn("Check the connection and retry", result[-1])
+        self.assertIn("请检查连接后重试", result[-1])
         self.assertNotIn("test detail", result[-1])
         self.assertIn("test detail", result[-2])
 
     def test_connection_status_explicit_and_escaped(self):
         with patch.object(app.GroundingClient, "health", return_value=(False, "<unsafe>")):
             output = app.service_status()
-        self.assertIn("Inference offline", output)
+        self.assertIn("推理未连接", output)
         self.assertIn("&lt;unsafe&gt;", output)
 
     def test_both_run_buttons_share_real_stream(self):

tests/test_model_release.py

@@ -0,0 +1,57 @@
+import json
+import tempfile
+import unittest
+from pathlib import Path
+
+from grounding_demo.model_spec import MODEL_REPO, MODEL_REVISION
+from grounding_demo.native_runtime import inspect_contract, ConfigurationError
+
+
+class ModelReleaseTests(unittest.TestCase):
+    def package(self, root):
+        config = {
+            "model_type": "groundingpi", "architectures": ["GroundingPIForConditionalGeneration"],
+            "vision_config": {"model_type": "groundingpi_vision"}, "image_token_id": 151655,
+            "auto_map": {"AutoModelForImageTextToText": "modeling_groundingpi.GroundingPIForConditionalGeneration",
+                         "AutoProcessor": "processing_groundingpi.GroundingPIProcessor"},
+        }
+        for name in ["configuration_groundingpi.py", "configuration_groundingpi_vision.py",
+                     "modeling_groundingpi.py", "modeling_groundingpi_vision.py",
+                     "processing_groundingpi.py", "image_processing_groundingpi.py",
+                     "media_utils.py", "streammind_gate.py", "chat_template.jinja"]:
+            (root/name).write_text("test fixture")
+        vocab = {f"<{n}>": n for n in range(1000)}
+        vocab.update({token: 150000+i for i,token in enumerate(
+            ["<|box_start|>", "<|box_end|>", "<|object_ref_start|>", "<|object_ref_end|>"])})
+        vocab["<|image_pad|>"] = 151655
+        for name, data in {"config.json": config, "preprocessor_config.json": {"media_proc_cfg": {}},
+                           "generation_config.json": {"eos_token_id": [151645]},
+                           "tokenizer_config.json": {}, "tokenizer.json": {"model": {"vocab": vocab}}}.items():
+            (root/name).write_text(json.dumps(data))
+        return config
+
+    def test_release_targets_new_accounts_pinned_public_model(self):
+        self.assertEqual(MODEL_REPO, "GroundingPI/GroundingPI")
+        self.assertRegex(MODEL_REVISION, r"^[0-9a-f]{40}$")
+
+    def test_new_model_contract_is_accepted_without_executing_model_code(self):
+        with tempfile.TemporaryDirectory() as d:
+            root=Path(d);self.package(root)
+            report=inspect_contract(root,require_weights=False)
+            self.assertEqual(report["vision"],"groundingpi_vision")
+            self.assertEqual(report["gam_coordinate_tokens"],1000)
+            self.assertEqual(report["eos_token_ids"],[151645])
+
+    def test_wrong_old_architecture_is_rejected(self):
+        with tempfile.TemporaryDirectory() as d:
+            root=Path(d);config=self.package(root)
+            config["model_type"]="oxygen_vl_v3"
+            (root/"config.json").write_text(json.dumps(config))
+            with self.assertRaisesRegex(ConfigurationError,"GroundingPI/GroundingPI"):
+                inspect_contract(root,require_weights=False)
+
+    def test_production_contract_requires_complete_weights(self):
+        with tempfile.TemporaryDirectory() as d:
+            root=Path(d);self.package(root)
+            with self.assertRaisesRegex(ConfigurationError,"weight shards"):
+                inspect_contract(root)