Runbook

Current DeepSeek RL launch and qualification entry: Miles baseline.

Before launch

Check	Required evidence
Source	Recorded commit, implementation hashes, clean launch tree where required
Runtime	Container identity and resolved package versions
Data	Complete manifest, tokenizer identity, approved split and cursor
Model	Named recipe, exact parent checkpoint, trainable set
Compute	Live allocation identity, mounts, GPU ownership and memory budget
Numerics	Relevant load/forward/backward/replay and distributed checks
Output	Fresh attempt namespace; no conflicting writer

Exclusive execution requires an explicitly available GPU allocation. Shared-GPU execution requires a declared sharing policy and measured memory admission.

Inspect before executing

PYTHONPATH=src python -m archlab.cli verify
PYTHONPATH=src python -m archlab.cli render \
  --config recipes/reference/speedrun_100m.yaml#qwen_gdn \
  --path data_root=/path/to/prepared/data

Rendering prints a command. It does not launch training.

Resume or restart

Operation	State handling
Resume same contract	Restore model, optimizer, scheduler, RNG and data cursor
New scientific contract	Record the changed axes and initialization explicitly
Graceful stop	Use the trainer's documented stop protocol; verify a complete checkpoint
Operational retry	New attempt identity, same scientific run identity

A source update does not change modules already imported by a live process. A running allocation is not proof of a healthy trainer.

During a run

Watch fresh progress, finite losses/gradients, memory peaks, data continuity and complete checkpoint markers. Compare throughput using the declared measurement window; report shared-resource contention.

Next: Data and Provenance · Tracking and Storage