From d7325f930683be629d9179dc07e6a0b9545feffe Mon Sep 17 00:00:00 2001 From: Hayden Free Date: Sat, 25 Apr 2026 22:23:45 -0400 Subject: [PATCH] Add generic RunPod runner harness + d12 baseline --- runs/runpod/d12.sh | 119 +++++++++++++++++++++++++++++++++++++++++ runs/runpod/kickoff.sh | 92 +++++++++++++++++++++++++++++++ 2 files changed, 211 insertions(+) create mode 100755 runs/runpod/d12.sh create mode 100755 runs/runpod/kickoff.sh diff --git a/runs/runpod/d12.sh b/runs/runpod/d12.sh new file mode 100755 index 00000000..9119c87d --- /dev/null +++ b/runs/runpod/d12.sh @@ -0,0 +1,119 @@ +#!/usr/bin/env bash +# d12 baseline runner. Runs INSIDE a RunPod pod. +# Pipeline: tokenizer -> base_train -> base_eval -> SFT -> chat_eval -> report. +# On exit: +# success -> upload final cache to HF, self-delete pod +# failure -> upload logs + report dir to HF under _failures/, self-delete pod +# (set UPLOAD_FAILURE_CACHE=1 to also dump partial cache for offline debug) +# +# Required env (passed via runpodctl --env at pod-create): +# HF_TOKEN, WANDB_API_KEY +# Optional env: +# WANDB_RUN default: d12 +# NANOCHAT_REPO default: Team-XSA/nanochat +# NANOCHAT_REF default: dev +# HF_REPO default: haydenfree/nanochat-d12-baseline +# BACKUP_INTERVAL default: 300 (seconds between background HF uploads) +# UPLOAD_FAILURE_CACHE default: 0 +# Auto-set by RunPod: +# RUNPOD_POD_ID, RUNPOD_API_KEY (pod-scoped) + +set -euo pipefail + +NANOCHAT_REPO="${NANOCHAT_REPO:-Team-XSA/nanochat}" +NANOCHAT_REF="${NANOCHAT_REF:-dev}" +HF_REPO="${HF_REPO:-haydenfree/nanochat-d12-baseline}" +WANDB_RUN="${WANDB_RUN:-d12}" +BACKUP_INTERVAL="${BACKUP_INTERVAL:-300}" +UPLOAD_FAILURE_CACHE="${UPLOAD_FAILURE_CACHE:-0}" + +WORKDIR="/workspace/nanochat" +LOG_FILE="/workspace/runner.log" +NANOCHAT_BASE_DIR="$HOME/.cache/nanochat" +BACKUP_PID="" + +mkdir -p /workspace +exec > >(tee -a "$LOG_FILE") 2>&1 + +echo "[runner] $(date -Iseconds) starting on pod=$RUNPOD_POD_ID" +echo "[runner] repo=$NANOCHAT_REPO ref=$NANOCHAT_REF hf_repo=$HF_REPO wandb_run=$WANDB_RUN" + +cleanup() { + local rc=$? + set +e + echo "[runner] cleanup: exit code $rc at $(date -Iseconds)" + if [ -n "$BACKUP_PID" ] && kill -0 "$BACKUP_PID" 2>/dev/null; then + kill "$BACKUP_PID" 2>/dev/null || true + fi + + local TS + TS=$(date -u +%Y%m%dT%H%M%SZ) + + if [ "$rc" -eq 0 ]; then + echo "[runner] success — final upload to $HF_REPO" + if [ -d "$NANOCHAT_BASE_DIR" ]; then + huggingface-cli upload "$HF_REPO" "$NANOCHAT_BASE_DIR" . \ + --repo-type model --commit-message "final rc=0 $TS" || \ + echo "[runner] WARN: final upload failed" + fi + else + echo "[runner] failure rc=$rc — dumping logs to HF for offline debug" + mkdir -p /tmp/failure + cp /workspace/*.log /tmp/failure/ 2>/dev/null || true + [ -d "$NANOCHAT_BASE_DIR/report" ] && cp -r "$NANOCHAT_BASE_DIR/report" /tmp/failure/ 2>/dev/null || true + [ -d "$WORKDIR" ] && (cd "$WORKDIR" && git rev-parse HEAD 2>/dev/null > /tmp/failure/git-head.txt || true) + + huggingface-cli upload "$HF_REPO" /tmp/failure "_failures/${TS}-rc${rc}/logs" \ + --repo-type model --commit-message "failure rc=$rc logs $TS" || \ + echo "[runner] WARN: log upload failed" + + if [ "$UPLOAD_FAILURE_CACHE" = "1" ] && [ -d "$NANOCHAT_BASE_DIR" ]; then + echo "[runner] UPLOAD_FAILURE_CACHE=1 — also dumping partial cache (may be slow)" + huggingface-cli upload "$HF_REPO" "$NANOCHAT_BASE_DIR" "_failures/${TS}-rc${rc}/cache" \ + --repo-type model --commit-message "failure rc=$rc cache $TS" || true + fi + echo "[runner] failure artifacts: https://huggingface.co/$HF_REPO/tree/main/_failures/${TS}-rc${rc}" + fi + + echo "[runner] self-deleting pod $RUNPOD_POD_ID" + runpodctl pod delete "$RUNPOD_POD_ID" 2>&1 || \ + curl -sS -X DELETE -H "Authorization: Bearer ${RUNPOD_API_KEY:-}" \ + "https://rest.runpod.io/v1/pods/$RUNPOD_POD_ID" + exit "$rc" +} +trap cleanup EXIT + +: "${HF_TOKEN:?HF_TOKEN must be set}" +: "${WANDB_API_KEY:?WANDB_API_KEY must be set}" +: "${RUNPOD_POD_ID:?RUNPOD_POD_ID must be set (auto by RunPod)}" + +rm -rf "$WORKDIR" +git clone "https://github.com/${NANOCHAT_REPO}.git" "$WORKDIR" +cd "$WORKDIR" +git checkout "$NANOCHAT_REF" +echo "[runner] HEAD = $(git rev-parse HEAD)" + +sed -i 's/--depth=24/--depth=12/' runs/speedrun.sh +sed -i 's/ --target-param-data-ratio=8//' runs/speedrun.sh +echo "[runner] speedrun.sh edits applied:" +grep -n 'depth\|target-param' runs/speedrun.sh || true + +pip install --quiet --upgrade huggingface_hub + +( + while true; do + sleep "$BACKUP_INTERVAL" + if [ -d "$NANOCHAT_BASE_DIR" ]; then + huggingface-cli upload "$HF_REPO" "$NANOCHAT_BASE_DIR" . \ + --repo-type model \ + --commit-message "checkpoint $(date -Iseconds)" >> /workspace/backup.log 2>&1 || true + fi + done +) & +BACKUP_PID=$! +echo "[runner] backup loop pid=$BACKUP_PID interval=${BACKUP_INTERVAL}s" + +export WANDB_RUN +WANDB_RUN="$WANDB_RUN" bash runs/speedrun.sh + +echo "[runner] $(date -Iseconds) pipeline complete" diff --git a/runs/runpod/kickoff.sh b/runs/runpod/kickoff.sh new file mode 100755 index 00000000..a0e4e2fb --- /dev/null +++ b/runs/runpod/kickoff.sh @@ -0,0 +1,92 @@ +#!/usr/bin/env bash +# Generic local kickoff for RunPod runs. +# Picks a runner script in this repo (runs/runpod/.sh) and spins up a pod. +# +# Prereqs: +# 1. ~/.config/team-xsa/runpod.env exports HF_TOKEN, WANDB_API_KEY, RUNPOD_TEMPLATE_ID +# 2. The template referenced by RUNPOD_TEMPLATE_ID has docker-start-cmd: +# bash,-lc,curl -fsSL "$RUNNER_URL" | bash >> /workspace/runner.log 2>&1 +# 3. The runner script for this experiment has been pushed to Team-XSA/nanochat +# +# Usage: +# source ~/.config/team-xsa/runpod.env +# bash runs/runpod/kickoff.sh d12 # uses runs/runpod/d12.sh +# bash runs/runpod/kickoff.sh d24 # uses runs/runpod/d24.sh +# bash runs/runpod/kickoff.sh xsa_d12 # uses runs/runpod/xsa_d12.sh +# +# Optional env overrides: +# GPU_ID default: "NVIDIA H100 80GB HBM3" +# GPU_COUNT default: 8 +# CLOUD_TYPE default: COMMUNITY (SECURE for guaranteed availability) +# DISK_GB default: 200 +# NANOCHAT_REPO default: Team-XSA/nanochat +# NANOCHAT_REF default: dev +# WANDB_RUN default: +# POD_NAME default: - + +set -euo pipefail + +RUNNER="${1:-}" +if [ -z "$RUNNER" ]; then + echo "Usage: bash runs/runpod/kickoff.sh " + echo " e.g. bash runs/runpod/kickoff.sh d12" + exit 1 +fi + +: "${HF_TOKEN:?HF_TOKEN not set — source ~/.config/team-xsa/runpod.env}" +: "${WANDB_API_KEY:?WANDB_API_KEY not set — source ~/.config/team-xsa/runpod.env}" +: "${RUNPOD_TEMPLATE_ID:?RUNPOD_TEMPLATE_ID not set — create the template once and add it to ~/.config/team-xsa/runpod.env}" + +NANOCHAT_REPO="${NANOCHAT_REPO:-Team-XSA/nanochat}" +NANOCHAT_REF="${NANOCHAT_REF:-dev}" +WANDB_RUN="${WANDB_RUN:-$RUNNER}" +RUNNER_URL="${RUNNER_URL:-https://raw.githubusercontent.com/${NANOCHAT_REPO}/${NANOCHAT_REF}/runs/runpod/${RUNNER}.sh}" + +GPU_ID="${GPU_ID:-NVIDIA H100 80GB HBM3}" +GPU_COUNT="${GPU_COUNT:-8}" +CLOUD_TYPE="${CLOUD_TYPE:-COMMUNITY}" +DISK_GB="${DISK_GB:-200}" +POD_NAME="${POD_NAME:-${RUNNER}-$(date +%Y%m%d-%H%M)}" + +echo "Verifying runner URL is reachable: $RUNNER_URL" +if ! curl -sfI "$RUNNER_URL" >/dev/null; then + echo "ERROR: runner not reachable at $RUNNER_URL" + echo " - Did you push runs/runpod/${RUNNER}.sh to ${NANOCHAT_REPO}@${NANOCHAT_REF}?" + echo " - Is the repo public?" + exit 1 +fi + +export HF_TOKEN WANDB_API_KEY WANDB_RUN RUNNER_URL NANOCHAT_REPO NANOCHAT_REF +ENV_JSON=$(python3 - <<'PY' +import json, os +keys = ["HF_TOKEN","WANDB_API_KEY","WANDB_RUN","RUNNER_URL","NANOCHAT_REPO","NANOCHAT_REF"] +print(json.dumps({k: os.environ[k] for k in keys if k in os.environ})) +PY +) + +echo "Creating pod:" +echo " name = $POD_NAME" +echo " template = $RUNPOD_TEMPLATE_ID" +echo " runner = $RUNNER_URL" +echo " gpu = $GPU_COUNT × $GPU_ID" +echo " cloud = $CLOUD_TYPE" +echo " disk = ${DISK_GB} GB" + +runpodctl pod create \ + --name "$POD_NAME" \ + --template-id "$RUNPOD_TEMPLATE_ID" \ + --gpu-id "$GPU_ID" \ + --gpu-count "$GPU_COUNT" \ + --cloud-type "$CLOUD_TYPE" \ + --container-disk-in-gb "$DISK_GB" \ + --env "$ENV_JSON" + +echo +echo "Logs (after pod boots):" +echo " POD_ID=\$(runpodctl pod list --name '$POD_NAME' -o json | jq -r '.[0].id')" +echo " runpodctl ssh info \$POD_ID" +echo " ssh @ 'tail -f /workspace/runner.log'" +echo +echo "Wandb: project=nanochat / nanochat-sft, run name: $WANDB_RUN" +echo "HF (success): https://huggingface.co/haydenfree/nanochat-d12-baseline" +echo "HF (failure): https://huggingface.co/haydenfree/nanochat-d12-baseline/tree/main/_failures"