install-miles-diffusion

name	install-miles-diffusion
description	One-click install of miles-diffusion on a fresh Linux GPU machine. Sets up the conda env, sglang-diffusion (PR

install-miles-diffusion

Drives the install helper at .claude/skills/install-miles-diffusion/install.sh and verifies the result. The helper is idempotent — re-running it skips steps that already succeeded (conda env exists, sglang at pinned commit, etc.).

All package versions are pinned, including torch, sglang (commit SHA), transformers, accelerate, ray, diffusers, peft, paddleocr, paddlepaddle-gpu, and the entire flow_grpo dep set. Pins reflect the currently-validated working environment.

When to invoke

User says anything like: "install miles-diffusion", "set up this repo on a new machine", "bootstrap the env for run-diffusion-grpo-ocr.sh", "装一下", etc.

What it installs

End goal: bash scripts/run-diffusion-grpo-ocr.sh boots cleanly on this host with bit-reproducible package versions.

Components (every version pinned):

1. apt system libs — libglib2.0-0 libgl1 (paddleocr / cv2 runtime).
2. Conda env — Python 3.11 (configurable via PY_VER).
3. Tooling — pip==26.0.1, wheel==0.45.1, setuptools==82.0.1 (resolver behaviour depends on these).
4. PyTorch — torch==2.9.1 on cu129 (override via TORCH_VER / CUDA_VER).
5. sglang-diffusion — clones Rockdu/sglang @ sglang-diffusion-rollout-test into $SGLANG_DIR (default ../sglang) and git checkout --detach $SGLANG_COMMIT (default 0372158dd66bc7cb0740c733bd60047db790ec7d). Installed editable as python[all]. Pinning to a SHA (not just the branch tip) is required for bit-exact rollout reproducibility. Override SGLANG_REPO / SGLANG_BRANCH / SGLANG_COMMIT only if you know what you're doing.
6. miles package — pip install -r requirements.txt (all ==-pinned: transformers 5.5.4, accelerate 1.12.0, ray 2.53.0, datasets 4.4.2, safetensors 0.7.0, wandb 0.23.1, …) plus pip install -e . --no-deps.
7. flow_grpo OCR deps — runs flow_grpo/setup.sh (every line --no-deps and ==-pinned: paddleocr 2.9.1, paddlepaddle-gpu 2.6.2, peft 0.18.1, diffusers 0.37.0, opencv 4.11.0.86, etc.).
8. torch_memory_saver — pinned to 0.0.9, skipped silently on failure.
9. Smoke test — nvidia-smi, then python -c "import train_diffusion".

How to run

Before doing anything, surface these to the user and let them override:

• ENV_NAME (default miles-diffusion)
• PY_VER (default 3.11)
• SGLANG_DIR (default $(dirname "$PWD")/sglang)
• SGLANG_REPO (default https://github.com/Rockdu/sglang.git)
• SGLANG_BRANCH (default sglang-diffusion-rollout-test)
• SGLANG_COMMIT (default 0372158dd66bc7cb0740c733bd60047db790ec7d)
• CUDA_VER (default 12.9)
• TORCH_VER (default 2.9.1)

Then:

bash .claude/skills/install-miles-diffusion/install.sh

It's long — run with run_in_background: true and stream with Monitor, or let it block in foreground if the user is OK with a 5–15 min wait.

What can go wrong

• No conda/mamba — helper aborts with a message telling the user to install miniforge. Don't auto-install conda.
• No CUDA toolkit / no GPU — nvidia-smi fails at smoke-test; install still succeeds but warn the user.
• sglang branch missing / renamed — if Rockdu/sglang no longer has sglang-diffusion-rollout-test, the clone/fetch fails. Do not silently fall back to upstream sgl-project/sglang: the required changes (multimodal_gen + weight-sync RPC) only live on the Rockdu fork. Surface the failure to the user and ask which branch to pin to instead.
• paddlepaddle-gpu wheel mismatch — pinned to 2.6.2 in flow_grpo/setup.sh. If the machine's CUDA is too new, you may need to swap the pin. Report the mismatch; don't silently change the pin.
• System apt missing sudo — fall back to apt-get without sudo (works in containers). If both fail, tell the user which .so is missing.

After install

Tell the user:

1. conda activate <ENV_NAME>
2. export WANDB_API_KEY=... (optional)
3. bash scripts/run-diffusion-grpo-ocr.sh

Do not kick off the training run yourself — the script does pkill -9 python at the top, which would kill the Claude Code session.