Getting Started

Install

Build from source (requires Go):

git clone https://github.com/laiaqwq/sandwichcake
cd sandwichcake
make build          # produces bin/sandwichcake with version stamped in
sudo make install   # installs to $(PREFIX)/bin plus a `swcake` symlink

Or without the Makefile:

go build ./... && go install ./cmd/sandwichcake

Quickstart

1. Pack a directory into a slice

sandwichcake pack weight ./models/qwen3-8b \
  --root /models/qwen3-8b \
  --to registry.example.com/ai/weights:qwen3-8b \
  --source hf://Qwen/Qwen3-8B@main

--root is required: it is the absolute path the directory contents take inside the composed image. Everything else (compression, layer splitting, ownership) has a per-type default you can override.

2. Write a cake.yaml

apiVersion: sandwichcake.io/v1
kind: Cake
target:
  ref: registry.example.com/ai/infer:v1
base:
  ref: python:3.12-slim
slices:
  - ref: registry.example.com/ai/venv:torch-2.8
  - ref: registry.example.com/ai/weights:qwen3-8b
config:
  env:
    PATH:
      prepend: /opt/venv/bin
  entrypoint: ["python", "serve.py"]

3. Preview the transfer plan

sandwichcake plan -f cake.yaml

Shows per-blob classification — which layers are already in the target repo, which can be cross-repo mounted, and which must be copied — before anything is pushed.

4. Compose and push

sandwichcake compose -f cake.yaml

Confirm the plan and the composed manifest is pushed to target.ref, along with metadata artifacts for each slice. A cake.lock with pinned digests is written next to cake.yaml; replay it bit-identically with compose --locked.

5. Inspect the result

sandwichcake inspect registry.example.com/ai/infer:v1
sandwichcake diff registry.example.com/ai/infer:v1 :v2

Registry auth

sandwichcake uses the standard Docker credential chain (~/.docker/config.json, credential helpers). For plain-HTTP registries add --insecure-registry <host> (repeatable).