itrialspace/docker/Dockerfile.gpu

# iTrialSpace — GPU image for synthesis (NodMAISI) and VLM evaluation.
# Pinned to the development environment: CUDA 12.1, torch 2.4.1, git-pinned MONAI.
#
#   docker build -t itrialspace-gpu -f docker/Dockerfile.gpu .
#
#   # Step 3 synthesis (mount data, outputs, and the NodMAISI weights):
#   docker run --rm --gpus all \
#       -e ITRIALSPACE_DATA_DIR=/data -e ITRIALSPACE_OUTPUT_DIR=/out \
#       -e NODMAISI_MODELS_DIR=/models \
#       -v /host/iTrialSpace:/data -v /host/outputs:/out -v /host/nodmaisi/models:/models \
#       --entrypoint python itrialspace-gpu \
#       -m itrialspace.synthesis.tools.run_itrialspace_to_ct \
#           --audit /out/inserted_masks/mode1_controlled_prevalence/audit.json \
#           --config src/itrialspace/synthesis/tools/integration_config.yaml \
#           --outdir /out/generated_cts/mode1_controlled_prevalence
#
# For a byte-for-byte clone of the dev env instead, build from requirements.lock.txt
# (see docs/installation.md). Base matches torch 2.4.1+cu121.
FROM nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04

ENV DEBIAN_FRONTEND=noninteractive \
    PYTHONUNBUFFERED=1 \
    PIP_NO_CACHE_DIR=1 \
    ITRIALSPACE_DATA_DIR=/data \
    ITRIALSPACE_OUTPUT_DIR=/out \
    NODMAISI_MODELS_DIR=/models

# System Python (3.10 on 22.04; package requires >=3.10) + libs for nibabel/SimpleITK/matplotlib
RUN apt-get update && apt-get install -y --no-install-recommends \
        python3 python3-pip python3-dev build-essential git \
        libgl1 libglib2.0-0 curl \
    && ln -sf /usr/bin/python3 /usr/bin/python \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /app
COPY . .

# Match the dev env: torch 2.4.1 (cu121) and the git-pinned MONAI build, then the
# package (editable so the synthesis subprocess path model resolves like dev).
RUN pip install --upgrade pip \
    && pip install torch==2.4.1 --index-url https://download.pytorch.org/whl/cu121 \
    && pip install "monai @ git+https://github.com/Project-MONAI/MONAI.git@c3a317d2bcb486199f40bda0d722a41e3869712a" \
    && pip install -e ".[imaging,synthesis,vlm,apps]"

# Pin the VLM/transformers stack to the dev-env (requirements.lock.txt) versions.
# The unpinned [vlm] extra otherwise pulls transformers 5.x, which references a
# torch-2.7-only dtype (float8_e8m0fnu) and crashes BiomedCLIP/MedGemma on torch 2.6.
RUN pip install --no-deps \
        transformers==4.51.3 \
        tokenizers==0.21.4 \
        huggingface_hub==0.36.2 \
        open_clip_torch==2.32.0 \
        timm==1.0.27 \
        safetensors==0.7.0

# Non-root user; create the runtime mount points.
# Make /app world-readable so the image can also be run with `--user <host uid>`
# (needed when the mounted data / weights are owned by a non-1000 host user, e.g.
# data on a group-restricted share or weights under a home dir). The `its` user
# remains the default for simple runs.
RUN useradd -m -u 1000 its && mkdir -p /data /out /models \
    && chown -R its:its /app /data /out /models \
    && chmod -R a+rX /app
USER its

# Default to the umbrella CLI; override --entrypoint for synthesis/VLM module runs.
ENTRYPOINT ["its"]
CMD ["config"]