When the secret is present, base_train --run=smoke-$run_number gets logged to wandb (project=nanochat). Without the secret it falls back to dummy / DummyWandb, so local runs stay offline.
- scripts/smoke.sh: end-to-end nanochat pipeline (dataset → tokenizer → tiny d=6 base_train, 50 iters); idempotent, caches under /data/nanochat-smoke (venv, uv-cache, shards). - CN mirrors baked in: aliyun PyPI, sjtu pytorch-wheels/cu128, hf-mirror. - workflow targets gpu label (ailab). drop earlier dummy workflow.
Fam Zheng