Update model reference to Qwen3.5-ocr-jp-2b

Dockerfile

@@ -0,0 +1,65 @@
+# Local-test Dockerfile for spaces/open_chandra_ocr.
+#
+# Mirrors the HF Space (ZeroGPU) stack so we can validate the inference path
+# locally without the host's CUDA-13/Python-3.12 incompatibilities:
+#   - Ubuntu 22.04 ships Python 3.10 (= Space's Python)
+#   - cu12.4 runtime libs (pre-built torch 2.6.0 cu12 wheels target this)
+#   - causal_conv1d wheel for cu12torch2.6 / cp310 / cxx11abi=FALSE
+#
+# Build:
+#   docker build -t open-chandra-ocr-local spaces/open_chandra_ocr/
+# Run:
+#   docker run --gpus all --rm -p 7860:7860 \
+#     -v $(pwd)/checkpoint/open-chandra-stage2-2b:/model:ro \
+#     -v $(pwd)/.cache/hf:/root/.cache/huggingface \
+#     open-chandra-ocr-local
+# Open http://localhost:7860
+
+FROM nvidia/cuda:12.4.1-devel-ubuntu22.04
+
+# `devel` (not `runtime`) is required: triton JIT-compiles its CUDA driver
+# bindings (`backends/nvidia/driver.c` → `cuda_utils.so`) at first invocation,
+# which needs gcc + cuda.h. Without these the kernels silently fall back to CPU.
+ENV DEBIAN_FRONTEND=noninteractive \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        python3.10 \
+        python3-pip \
+        python3.10-dev \
+        gcc \
+        g++ \
+        ca-certificates \
+        libgl1 \
+        libglib2.0-0 \
+    && rm -rf /var/lib/apt/lists/* \
+    && ln -sf /usr/bin/python3.10 /usr/bin/python
+
+WORKDIR /app
+
+# Pin torch+vision from the cu12.4 index so the wheel ABI matches our cuda image.
+RUN pip install --no-cache-dir \
+        --extra-index-url https://download.pytorch.org/whl/cu124 \
+        torch==2.6.0 torchvision==0.21.0
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Pre-install causal_conv1d for fla fast path (Python 3.10 / cu12 / torch 2.6 / cxx11abi=FALSE).
+# app.py also tries to install at runtime; this pre-install just makes startup faster.
+RUN pip install --no-cache-dir --no-deps \
+        https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.5.0.post8/causal_conv1d-1.5.0.post8+cu12torch2.6cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
+
+COPY app.py .
+
+# Defaults: load model from the bind-mounted /model, expose Gradio on all interfaces.
+ENV MODEL_ID=/model \
+    GRADIO_SERVER_NAME=0.0.0.0 \
+    GRADIO_SERVER_PORT=7860 \
+    HF_HOME=/root/.cache/huggingface
+
+EXPOSE 7860
+
+CMD ["python", "app.py"]

README.md

@@ -1,5 +1,5 @@
 ---
-title: Open Chandra OCR
+title: Qwen3.5-OCR-JP-2B
 emoji: 📄
 colorFrom: blue
 colorTo: gray
@@ -10,18 +10,18 @@ pinned: false
 license: apache-2.0
 hardware: zero-a10g
 models:
-  - ebinan92/open-chandra-stage2-2b
+  - ebinan92/Qwen3.5-ocr-jp-2b
 ---
 
-# Open Chandra OCR
+# Qwen3.5-OCR-JP-2B
 
-Gradio app for hosting a custom Chandra-like OCR model on Hugging Face Spaces.
+Gradio demo for [`ebinan92/Qwen3.5-ocr-jp-2b`](https://huggingface.co/ebinan92/Qwen3.5-ocr-jp-2b) on Hugging Face Spaces.
 
 ## Configuration
 
 Set these Space variables as needed:
 
-- `MODEL_ID`: Hugging Face model repo or local model path. Defaults to `ebinan92/open-chandra-stage2-2b`.
+- `MODEL_ID`: Hugging Face model repo or local model path. Defaults to `ebinan92/Qwen3.5-ocr-jp-2b`.
 - `HF_TOKEN`: required Space **Secret** for private models (read scope).
 - `MAX_IMAGE_PIXELS`: area-based image cap (px²). Defaults to `2359296` (1536²) ≈ 2300 visual tokens — matches the [chandra-ocr-2 reference Space](https://huggingface.co/spaces/victor/chandra-ocr-2) for ZeroGPU-friendly latency.
 - `PDF_DPI`: PDF render DPI. Defaults to `180`.

app.py

@@ -12,26 +12,33 @@ import time
 from pathlib import Path
 
 # Pre-built causal_conv1d wheel (linear-attention dependency, can't compile in Space build).
-# Pulled from Dao-AILab official GitHub release; matches torch==2.6.0 / cu12 / py3.10 / cxx11abi=FALSE.
+# Pulled from Dao-AILab official GitHub release. Auto-selects the cpXY wheel matching the
+# running Python version (HF Space = cp310, local dev py3.12 = cp312, etc.).
+# Soft-fails on unsupported combos — flash-linear-attention falls back to its torch impl.
+_CAUSAL_CONV1D_PY = f"cp{sys.version_info.major}{sys.version_info.minor}"
 _CAUSAL_CONV1D_WHEEL = (
     "https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.5.0.post8/"
-    "causal_conv1d-1.5.0.post8+cu12torch2.6cxx11abiFALSE-cp310-cp310-linux_x86_64.whl"
+    f"causal_conv1d-1.5.0.post8+cu12torch2.6cxx11abiFALSE-"
+    f"{_CAUSAL_CONV1D_PY}-{_CAUSAL_CONV1D_PY}-linux_x86_64.whl"
 )
 try:
     import causal_conv1d  # noqa: F401
 except ImportError:
-    subprocess.run(
-        [sys.executable, "-m", "pip", "install", _CAUSAL_CONV1D_WHEEL, "--no-deps", "-q"],
-        check=True,
-    )
-    import causal_conv1d  # noqa: F401
+    try:
+        subprocess.run(
+            [sys.executable, "-m", "pip", "install", _CAUSAL_CONV1D_WHEEL, "--no-deps", "-q"],
+            check=True,
+        )
+        import causal_conv1d  # noqa: F401
+    except Exception as e:
+        print(f"[warn] causal_conv1d unavailable ({e}); slower linear-attention conv1d path will be used")
 
 import gradio as gr
 import spaces
 from PIL import Image
 
 
-MODEL_ID = os.getenv("MODEL_ID", "ebinan92/open-chandra-stage2-2b")
+MODEL_ID = os.getenv("MODEL_ID", "ebinan92/Qwen3.5-ocr-jp-2b")
 PROMPT = os.getenv(
     "OCR_PROMPT", "OCR this image as HTML layout blocks with bbox and label."
 )
@@ -667,8 +674,8 @@ CSS = """
 """
 
 
-with gr.Blocks(title="Open Chandra OCR", css=CSS) as demo:
-    gr.Markdown("# Open Chandra OCR", elem_classes=["app-title"])
+with gr.Blocks(title="Qwen3.5-OCR-JP-2B", css=CSS) as demo:
+    gr.Markdown("# Qwen3.5-OCR-JP-2B", elem_classes=["app-title"])
     gr.Markdown(
         "Upload an image or PDF and run OCR. Predicted blocks are shown in the right pane.",
         elem_classes=["app-sub"],