Fam Zheng
b585e07dc2
smoke / nanochat-smoke (push) Successful in 2m30s
- .gitea/workflows/smoke.yml: gitea CI on ailab gpu runner (manual git clone since actions/checkout@v4 mis-resolves subpath gitea); injects WANDB_API_KEY + CI_RUN_TAG=smoke-$run_number - scripts/smoke.sh: in-place smoke (uv sync + 1 shard + tokenizer + d=6 50-step base_train); idempotent cache at /data/nanochat-smoke/ - doc/research_feasibility.md: voice-first multimodal feasibility study (mochi) - doc/todo.md: phase-by-phase roadmap (W1 Whisper smoke → W4 MVP) - README.md: omni preamble pointing at upstream nanochat README - .gitignore: exclude .claude/ runtime files
56 lines
1.8 KiB
Bash
Executable File
56 lines
1.8 KiB
Bash
Executable File
#!/bin/bash
|
|
# End-to-end smoke for nanochat-omni: dataset → tokenizer → tiny base_train.
|
|
# Runs in-place at the repo root (post-monorepo-fork).
|
|
# Idempotent: caches venv + uv-cache + dataset shards under /data/nanochat-smoke/.
|
|
# Targets ailab (CN, RTX 5090). CN mirrors (sjtu pytorch, aliyun PyPI, hf-mirror)
|
|
# are committed directly into pyproject.toml/uv.lock/nanochat/dataset.py.
|
|
set -euo pipefail
|
|
export PATH="$HOME/.local/bin:$PATH"
|
|
|
|
CACHE_ROOT=${CACHE_ROOT:-/data/nanochat-smoke}
|
|
export NANOCHAT_BASE_DIR=$CACHE_ROOT/cache
|
|
export UV_CACHE_DIR=$CACHE_ROOT/uv-cache
|
|
export UV_DEFAULT_INDEX=${UV_DEFAULT_INDEX:-https://mirrors.aliyun.com/pypi/simple/}
|
|
export UV_INDEX_STRATEGY=unsafe-best-match
|
|
export HF_ENDPOINT=${HF_ENDPOINT:-https://hf-mirror.com}
|
|
export OMP_NUM_THREADS=1
|
|
|
|
mkdir -p "$CACHE_ROOT" "$NANOCHAT_BASE_DIR"
|
|
|
|
# wandb: real run if WANDB_API_KEY is set, otherwise fall back to dummy (DummyWandb).
|
|
if [ -n "${WANDB_API_KEY:-}" ]; then
|
|
RUN_TAG=${CI_RUN_TAG:-smoke-$(date +%Y%m%d-%H%M%S)}
|
|
else
|
|
RUN_TAG=dummy
|
|
fi
|
|
|
|
[ -d .venv ] || uv venv
|
|
uv sync --extra gpu --index-strategy unsafe-best-match
|
|
source .venv/bin/activate
|
|
|
|
echo "=== [1/4] download 1 climbmix shard ==="
|
|
time python -m nanochat.dataset -n 1
|
|
|
|
echo "=== [2/4] train tokenizer (50M chars) ==="
|
|
time python -m scripts.tok_train --max-chars=50000000
|
|
|
|
echo "=== [3/4] tok_eval ==="
|
|
time python -m scripts.tok_eval
|
|
|
|
echo "=== [4/4] base_train d=6 50 iters ==="
|
|
time python -m scripts.base_train \
|
|
--depth=6 \
|
|
--head-dim=64 \
|
|
--window-pattern=L \
|
|
--max-seq-len=512 \
|
|
--device-batch-size=8 \
|
|
--total-batch-size=4096 \
|
|
--eval-every=25 \
|
|
--eval-tokens=131072 \
|
|
--core-metric-every=-1 \
|
|
--sample-every=25 \
|
|
--num-iterations=50 \
|
|
--run="$RUN_TAG"
|
|
|
|
echo "=== smoke done ==="
|