docker_build/remote/README.md

MMSciCode env reconstruction on the domestic server

This bundle reconstructs the 204 MMSciCode conda envs on a domestic Chinese server using mirror sources (no need to download the 165 GB tarball repo). It then packs each env into a tarball and builds a Docker image per env.

Inputs (from HF: MMSciCode/mmsci-envs-tmp)

configs/
  <env>/
    env.yml           # conda env create -f
    explicit.txt      # conda list --explicit --md5 (URL-pinned fallback)
    pip-freeze.txt    # raw pip freeze from local env
    pip-clean.txt     # pip-freeze with conda-built file:// stripped
    editable.txt      # tab-separated: <egg-link name> \t <original source path>
    meta.json         # python ver, has_cuda hint, editable_count
editable_bundle.tar.gz  # only present if any env had editable installs

Domestic-server prerequisites

• conda (or miniconda) installed; conda >= 22.x
• mamba from the base env: conda install -n base -c conda-forge mamba conda-pack
• Docker (with nvidia-docker runtime if envs need GPU)
• ~200 GB free disk for staging tarballs + ~300 GB for Docker images

Step 1 — configure mirrors

cp configs_repo/remote/condarc.template ~/.condarc
mkdir -p ~/.config/pip
cp configs_repo/remote/pip.conf.template ~/.config/pip/pip.conf

Step 2 — fetch the artifacts

# example using huggingface_hub with HF_ENDPOINT=https://hf-mirror.com if needed
huggingface-cli download MMSciCode/mmsci-envs-tmp \
    --repo-type dataset \
    --local-dir /home/mmsci/repo \
    --include 'configs/**' 'editable_bundle.tar.gz'

cd /home/mmsci/repo
[ -f editable_bundle.tar.gz ] && tar -xzf editable_bundle.tar.gz   # -> editable_sources/

Step 3 — build all envs (parallel) and pack each as a tarball

CONFIGS_DIR=/home/mmsci/repo/configs \
OUT_DIR=/home/mmsci/conda_packs \
EDITABLE_SRC=/home/mmsci/repo/editable_sources \
MAMBA_BIN=/opt/conda/bin/mamba \
CONDA_PACK=/opt/conda/bin/conda-pack \
JOBS=4 \
bash /home/mmsci/configs_repo/remote/build_envs_remote.sh

Outputs:

• $OUT_DIR/<env>.tar.gz — per-env conda-pack tarball
• $OUT_DIR/build.state.tsv — per-env status (done/skip/fail)
• $OUT_DIR/build.failed.txt — failed envs (re-run script to retry; idempotent)
• $OUT_DIR/_logs/<env>.log — per-env mamba + pack log

Expect 5–15% of envs to fail on first pass (mirror gaps, conda-pack errors, editable misses). Inspect the failed logs and either:

• Adjust env.yml (drop the offending pin, add a missing channel), or
• Fall back to the original tarball from MMSciCode/mmsci-envs-tmp root: huggingface-cli download MMSciCode/mmsci-envs-tmp --include '<env>.tar.gz' --local-dir $OUT_DIR

Step 4 — build Docker images

Get the Dockerfile (from the source repo or the HF docker_build/ upload):

PACKS_DIR=/home/mmsci/conda_packs \
DOCKERFILE=/home/mmsci/configs_repo/Dockerfile.conda-env \
IMAGE_PREFIX=mmsci-py \
JOBS=2 \
bash /home/mmsci/configs_repo/remote/build_docker_images.sh

Each env -> mmsci-py-<sanitized-name>:latest. The Dockerfile is the unmodified Dockerfile.conda-env — it only does COPY tarball; tar -xzf; conda-unpack, so build time per image is whatever your disk can write the tarball at (~30s–2min per image).

Troubleshooting

PackageNotFoundError during mamba env create The Tsinghua/Aliyun mirrors don't have every conda-forge package. Try:

mamba env create -n <env> -f configs/<env>/env.yml \
    --override-channels -c conda-forge -c defaults

If still missing, the explicit URL list often works:

conda create -n <env> --file configs/<env>/explicit.txt

(URLs point to anaconda.org which is reachable but slow; consider setting CONDA_SUBDIR=linux-64.)

pip install connection errors Confirm ~/.config/pip/pip.conf is in effect: pip config list. Some git+https URLs in pip-freeze.txt may need a proxy; consider running the build under https_proxy=... if you have a corporate proxy.

conda-pack editable-packages error build_envs_remote.sh passes --ignore-editable-packages so this should not happen. If it does, the env has an editable install that wasn't recorded in editable.txt — re-export from the source machine.