Install Miles Diffusion

install-miles-diffusion

Docker is the official way to run miles_diffusion — pull OFFICIAL_IMAGE and use it. This skill is a fallback for machines that cannot, and it is not recommended: it reconstructs the image's package set, it is not the image. deep_ep and flash_mla cannot be reproduced at all, and every apt-sourced dist depends on the host being Ubuntu 24.04.

Reproduces the official Docker environment on a box that has only CUDA 12.9 and Ubuntu 24.04. The target is not "an env that works" but "the image's env": snapshot/ is a capture of OFFICIAL_IMAGE, install.sh replays it, and verify_env.py fails if the result drifts.

Files

Run

A bare CUDA image has no git, and some clusters hand out a resolver that can't answer for outside names:

grep -q nameserver /etc/resolv.conf || echo "nameserver 8.8.8.8" >> /etc/resolv.conf
apt-get update -qq && apt-get install -y --no-install-recommends git ca-certificates
git clone https://github.com/radixark/miles_diffusion.git && cd miles_diffusion
bash .claude/skills/install-miles-diffusion/install.sh

~20 min cold, ~4 with a warm pip cache. Run it in the background and poll the log. --from STEP resumes; every step is idempotent. To check an existing box instead: python3 .claude/skills/install-miles-diffusion/verify_env.py.

Why it replays instead of resolving

The image does not satisfy its own dependency metadata — pip check there reports ~20 violations, because sglang is installed --no-deps from a commit newer than the base image's package set, and nvidia-modelopt wants setuptools>=80 against a pinned 70.2.0. Resolving the freeze returns ResolutionImpossible, and any relaxation that resolves lands on versions the image does not have. So install.sh installs packages.txt with --no-deps.

Two packages cannot be reproduced at all: deep_ep and flash_mla are compiled inside lmsysorg/sglang and published nowhere. Neither is on a miles_diffusion import path.

flash_attn_3 comes from a wheel the image installs from a local copy, so pip freeze reports an unresolvable file:// path. refresh.sh rewrites it to the release asset the Dockerfile pulls from (FA3_WHEELS_REPO / FA3_WHEELS_TAG), so pip installs it like any other pin.

sglang and miles are anchored to the refs the Dockerfile builds from

sglang comes from the sglang-miles-h3 integration branch and miles from main, as in the Dockerfile (SGLANG_DIFFUSION_BRANCH=sglang-miles-h3, MILES_DIFFUSION_COMMIT=main). The difference is that the Dockerfile follows the branch tip at build time while this pins a commit and checks it is an ancestor of the branch. That check matters: the first capture pinned a local cherry-pick that exists in no remote.

Because the ancestor check needs real history, the checkout is a blob-filtered full clone rather than the image's --depth=1, so setuptools_scm reports a different commit count for the same sha; verify_env.py compares sglang and miles by presence, not by that string.

Refreshing the snapshot

bash .claude/skills/install-miles-diffusion/refresh.sh <devbox-on-the-new-image>

Capture from a box running the image itself, not from whichever tag a devbox happens to be on — the first snapshot here was taken from a devbox two image releases behind, which pinned diffusers 0.37.0 against a repo that had moved to 0.38.0.

A devbox people have worked on is also not a pristine image. refresh.sh guards both drift classes it has hit: the sglang commit comes from setuptools_scm rather than git rev-parse HEAD, and packages dated after the image's build day are marked #skip[capture-drift]. If the box ran :latest, replace OFFICIAL_IMAGE with the concrete tag it resolved to.

What goes wrong

• "Cannot uninstall X, RECORD file not found" — apt dists have no RECORD. step_pip pre-seeds pip, PyJWT and wheel with --ignore-installed; a fourth needs the same flag.
• "Device or resource busy" on a file in /usr/local/bin — the host mounted a binary read-only there (rx devboxes do this for uv, gh, claude). step_packages drops the matching packages and verify treats them as expected absences.
• Not Ubuntu 24.04 — the apt-sourced python dists then come from different pockets and verify reports drift on them.

Running SD3 after install

export HF_TOKEN=...              # SD3.5 is gated
export NCCL_NVLS_ENABLE=0        # partial-node containers have no IMEX multicast capability
python3 scripts/run_diffusion_grpo_sd3_ocr_sglang.py --cuda-visible-devices 0,1

Without NCCL_NVLS_ENABLE=0, execute_train enables NVLink SHARP whenever it sees NVLink and NCCL dies in FSDPTrainRayActor.init with Failed to bind NVLink SHARP (NVLS) Multicast memory. On a fresh HF cache the SD3.5 download pulls both model.safetensors and model.fp16.safetensors for the CLIP encoders; sglang's fast loader refuses the duplicate tensor names, logs a traceback and falls back to the native loader. The run continues.

Verified on nvidia/cuda:12.9.1-cudnn-devel-ubuntu24.04, 4×H100, no python preinstalled: 369 matched / 0 mismatched / 0 missing, then a completed GRPO step of the SD3 recipe.