Dockerfile

# Local-test Dockerfile for spaces/Qwen3.5-ocr-jp-2b.
#
# 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)
#   - cu12.8 runtime libs (pre-built torch 2.8.0 cu12 wheels target this)
#   - causal_conv1d wheel for cu12torch2.8 / cp310 / cxx11abi=TRUE
#     (torch >=2.8 ships with cxx11abi=TRUE by default)
#
# Build:
#   docker build -t qwen3.5-ocr-jp-2b-local spaces/Qwen3.5-ocr-jp-2b/
# Run:
#   docker run --gpus all --rm -p 7860:7860 \
#     -v $(pwd)/checkpoint/Qwen3.5-ocr-jp-2b:/model:ro \
#     -v $(pwd)/.cache/hf:/root/.cache/huggingface \
#     qwen3.5-ocr-jp-2b-local
# Open http://localhost:7860

FROM nvidia/cuda:12.8.1-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 cu12.8 index so the wheel ABI matches our cuda image.
RUN pip install --no-cache-dir \
        --extra-index-url https://download.pytorch.org/whl/cu128 \
        torch==2.8.0 torchvision==0.23.0

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 2.8 / cxx11abi=TRUE).
# 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/v1.6.2.post1/causal_conv1d-1.6.2.post1+cu12torch2.8cxx11abiTRUE-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"]