Add docker/hf_{push,pull}.sh: simulate docker push/pull against HF Hub

HF Hub does not act as a generic OCI registry for arbitrary Docker
images -- the registry.hf.space endpoint only serves images that
Spaces have built, not images pushed by `docker push`. So we cannot
do `docker push huggingface.co/user/repo:tag` for an Unsloth image.

For cross-host testing where we want one canonical place to pull
from (and Docker Hub credentials are not yet configured), wrap the
manual flow into push/pull-shaped commands:

  hf_push.sh: docker save | pigz | huggingface-cli upload
  hf_pull.sh: huggingface-cli download | gunzip | docker load

This is approximation, not real OCI semantics -- every push uploads
the full ~4 GB blob, no layer dedup, no manifest negotiation. Good
for testing across A100 / H100 / RTX 6000 boxes; the real release
should go through .github/workflows/docker-publish.yml to Docker Hub,
which gets layer dedup, multi-arch manifest support, and standard
`docker pull` UX for users.

Usage:
  bash docker/hf_push.sh unsloth-blackwell:test danielhanchen/unsloth-blackwell-docker
  bash docker/hf_pull.sh danielhanchen/unsloth-blackwell-docker unsloth-blackwell-test.tar.gz unsloth-blackwell:test
This commit is contained in:
Daniel Han 2026-05-24 09:25:00 +00:00
commit 4bfb4b891a
2 changed files with 86 additions and 0 deletions

41
docker/hf_pull.sh Executable file
View file

@ -0,0 +1,41 @@
#!/usr/bin/env bash
# Simulate `docker pull <image>` against a Hugging Face Hub model repo.
#
# Counterpart to docker/hf_push.sh -- downloads the tar.gz blob from the HF
# repo and `docker load`s it.
#
# Usage:
# bash docker/hf_pull.sh <hf_repo> [<blob>] [<verify_tag>]
# bash docker/hf_pull.sh danielhanchen/unsloth-blackwell-docker unsloth-blackwell-test.tar.gz unsloth-blackwell:test
#
# Requires: docker, pigz (or gzip), huggingface-cli logged in (read scope is
# sufficient for public repos).
set -euo pipefail
REPO="${1:?usage: hf_pull.sh <hf_repo> [<blob>] [<verify_tag>]}"
BLOB="${2:-unsloth-blackwell.tar.gz}"
VERIFY="${3:-}"
WORK="${HF_PULL_TMP:-/tmp}"
command -v docker >/dev/null || { echo "ERROR: docker not on PATH"; exit 1; }
command -v huggingface-cli >/dev/null || { echo "ERROR: huggingface-cli not on PATH (pip install -U huggingface_hub)"; exit 1; }
DECOMPRESSOR=$(command -v pigz || command -v gzip) || { echo "ERROR: need pigz or gzip"; exit 1; }
DEST="${WORK}/$(basename "${BLOB}")"
echo ">> downloading ${REPO}/${BLOB} -> ${DEST}"
huggingface-cli download "${REPO}" "${BLOB}" --repo-type=model --local-dir "${WORK}"
ls -lh "${DEST}"
echo ">> loading into docker (using ${DECOMPRESSOR##*/})"
"${DECOMPRESSOR}" -d -c "${DEST}" | docker load
if [[ -n "${VERIFY}" ]]; then
if docker image inspect "${VERIFY}" >/dev/null 2>&1; then
echo ">> verified: ${VERIFY} is loaded"
docker image inspect --format 'image_id={{.Id}} size={{.Size}}' "${VERIFY}"
else
echo "WARN: expected tag ${VERIFY} not found after load. docker images:"
docker images
exit 1
fi
fi

45
docker/hf_push.sh Executable file
View file

@ -0,0 +1,45 @@
#!/usr/bin/env bash
# Simulate `docker push <image>` against a Hugging Face Hub model repo.
#
# HF Hub doesn't act as an OCI registry for arbitrary images (only Spaces have
# that). So we approximate the push by:
# 1. docker save | pigz -> single tar.gz blob
# 2. huggingface-cli upload to <repo>/{tag}.tar.gz
#
# This is good for cross-host testing where you want one canonical place to
# pull from. For the real release, use Docker Hub or GHCR with `docker push`,
# which gives you layer dedup, manifest negotiation, and standard `docker pull`
# UX -- see .github/workflows/docker-publish.yml in this repo.
#
# Usage:
# bash docker/hf_push.sh <image:tag> <hf_repo>
# bash docker/hf_push.sh unsloth-blackwell:test danielhanchen/unsloth-blackwell-docker
#
# Requires: docker, pigz (or gzip), huggingface-cli logged in with a write token.
set -euo pipefail
IMAGE="${1:?usage: hf_push.sh <image:tag> <hf_repo>}"
REPO="${2:?usage: hf_push.sh <image:tag> <hf_repo>}"
TAG="${IMAGE##*:}"
NAME="${IMAGE%:*}"
NAME="${NAME##*/}"
BLOB="${NAME}-${TAG}.tar.gz"
WORK="${HF_PUSH_TMP:-/tmp}"
command -v docker >/dev/null || { echo "ERROR: docker not on PATH"; exit 1; }
command -v huggingface-cli >/dev/null || { echo "ERROR: huggingface-cli not on PATH (pip install -U huggingface_hub)"; exit 1; }
COMPRESSOR=$(command -v pigz || command -v gzip) || { echo "ERROR: need pigz or gzip"; exit 1; }
OUT="${WORK}/${BLOB}"
echo ">> saving ${IMAGE} -> ${OUT} (using ${COMPRESSOR##*/})"
docker save "${IMAGE}" | "${COMPRESSOR}" > "${OUT}"
ls -lh "${OUT}"
echo ">> uploading to https://huggingface.co/${REPO}/blob/main/${BLOB}"
huggingface-cli upload "${REPO}" "${OUT}" "${BLOB}" \
--repo-type=model \
--commit-message="push ${IMAGE} ($(docker inspect --format '{{.Id}}' "${IMAGE}" | cut -c8-19))"
echo ">> pushed."
echo "On the consumer side, run:"
echo " bash docker/hf_pull.sh ${REPO} ${BLOB} ${IMAGE}"