feat(docker): add Dockerfile.arm64 for Linux aarch64 / SBC builds

[Steve235lab]

Summary

• Adds docker/Dockerfile.arm64 to support building InvokeAI on Linux ARM64 devices (Raspberry Pi, SBCs, etc.)
• Updates docker/README.md with build instructions and corrects the "only x86_64 is supported" note

Why it's needed: The existing Dockerfile fails on linux/aarch64 because PyTorch's +cpu wheels on download.pytorch.org/whl/cpu are only published for manylinux_2_28_x86_64 and win_amd64. There are no aarch64 wheels on that index.

How it works:

The new Dockerfile.arm64 diverges from the upstream Dockerfile in one key section — the dependency install step:

1. pyproject.toml is COPYed as a writable file rather than a read-only bind-mount.
2. awk strips the explicit [tool.uv.sources] index overrides for torch, torchvision, and pytorch-triton-rocm, so uv falls back to PyPI for those packages.
3. sed replaces all platform-specific version pins (+cpu / +cu128 / +rocm*) across all optional extras with bare PyPI-compatible specifiers. This is necessary because uv sync resolves all extras simultaneously — any unresolvable extra (e.g. rocm still pinning torch==2.7.1+rocm6.3) fails the entire build even if that extra isn't being installed.
4. uv sync --extra cpu is run without --frozen since the lock file was generated on x86_64.

PyPI ships torch==2.7.1 and torchvision==0.22.1 with linux_aarch64 wheels, so the install succeeds cleanly.

Tested on: Raspberry Pi 5, linux/aarch64, Ubuntu 24.04, Docker 27.x. Full build succeeds and InvokeAI starts correctly.

Scope note: This PR covers Linux ARM64 only. macOS Apple Silicon has different Docker/PyTorch considerations and is out of scope here.

Related Issues / Discussions

The upstream Dockerfile already contains this comment acknowledging the gap:

# NOTE: there are no pytorch builds for arm64 + cuda, only cpu

QA Instructions

On any linux/aarch64 machine with Docker installed:

git clone https://github.com/invoke-ai/InvokeAI.git
cd InvokeAI
docker build -f docker/Dockerfile.arm64 -t invokeai:arm64 .
docker run --publish 9090:9090 --volume ~/invokeai:/invokeai invokeai:arm64

Navigate to http://localhost:9090 and verify the UI loads. API-backed models (GPT Image, Gemini) can be configured and used without a GPU.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)