CIS490/tools/ingest_lambda_artifacts.py
Max 308140c6ce training: lambda-cloud one-shot training integration
External-GPU path for the time-pressured first round, before the
Windows desktop joins the WG fleet. Lambda is treated as an "external
worker" whose output lands in the same /var/lib/cis490/models/ tree
the receiver-coordinated fleet uses, so cis490-jobs status reflects
Lambda runs identically to fleet runs.

Three scripts + one ingest tool:

  scripts/build-lambda-bundle.sh
    Tarball at /tmp/cis490-lambda/lambda-bundle-<short>.tar.zst with:
      - the repo (sans .git, sans data/, sans artifacts*)
      - data/processed/{validation_v1,features_window_v1}.parquet
      - data/processed/feature_schema_v1.json
      - data/processed/tensor_window_v1/   (npz shards)
      - bootstrap.sh (entrypoint)
      - training_manifest.toml (the canonical job list)
      - BUNDLE_MANIFEST.json (commit hash + counts + build stamp)
    Verifies all four data inputs exist BEFORE compressing 5+ GB.

  scripts/run-on-lambda.sh ubuntu@<ip>
    rsync bundle up → ssh + run bootstrap → rsync artifacts +
    reports/eval back to artifacts-lambda/ + reports/lambda/.
    Resumable rsync; sha256-verified.

  scripts/lambda-bootstrap.sh   (runs ON the Lambda instance)
    Creates .venv with cu121 torch + xgboost + the [training] deps,
    iterates the manifest's job list in priority order (highest first),
    runs trainer/run.py (or run_ssl.py for transformer_ssl) per job,
    skips jobs whose .ckpt.json already exists (idempotent on re-run),
    writes per-job logs/<model>_<mode>.log, runs eval suite at the end,
    stamps artifacts/RUN_SUMMARY.json with counts + failed-job list.

  tools/ingest_lambda_artifacts.py
    Bundles each (ckpt.json + sidecar + train.json) trio into a
    .tar.zst, sha256, PUTs to the local trainer-receiver's
    /v1/model/{job_id}, marks the job complete. Maps (model, mode) →
    job_id by re-reading the canonical manifest. Handles the queue
    state churn (requeue if completed, claim if pending, fail-back
    on race losses).

End-to-end smoke verified on the A100 instance just provisioned:
  - SSH from Pi via ed25519 keypair (cis490-trainer-pi)
  - GPU: A100-SXM4-40GB, driver 580.105.08
  - venv warmed: torch 2.5.1+cu121, xgboost 3.2.0
  - 464 GB ephemeral disk available

Pi-side feature build (build_features.py + build_tensors.py against
all 72,952 accepted+degraded episodes) is in progress; bundle build
gates on its completion. Estimated wall-clock for the full Lambda
training run on A100: ~2.5 hours for 12 supervised + 2 SSL models +
eval suite.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-08 12:32:04 -05:00

241 lines
9.3 KiB
Python

"""Ingest Lambda-trained artifacts into the local trainer-receiver.
Lambda finishes a run; run-on-lambda.sh rsyncs artifacts/ + reports/eval/
back to the Pi at $REPO/artifacts-lambda/ + $REPO/reports/lambda/.
This script bundles each (ckpt.json + sidecar + train.json) trio into a
.tar.zst, computes its sha256, and POSTs it to the trainer-receiver via
PUT /v1/model/{job_id} so the model store on the Pi looks identical to
what a real fleet worker would have produced.
This preserves the fleet abstraction: Lambda is just an "external worker"
whose output lands in the same /var/lib/cis490/models/ tree. The
operator's `cis490-jobs status` shows them as completed jobs with
proper artifact_ids.
Usage:
python -m tools.ingest_lambda_artifacts \\
--artifacts ./artifacts-lambda \\
--reports ./reports/lambda/eval \\
--receiver http://127.0.0.1:8445 \\
--validation data/processed/validation_v1.parquet \\
--as-host lambda-a100
"""
from __future__ import annotations
import argparse
import hashlib
import io
import json
import logging
import re
import sys
import tarfile
from pathlib import Path
import zstandard as zstd
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from training.fleet.client import FleetClient
from training.fleet.manifest import load as load_manifest
log = logging.getLogger("cis490.ingest_lambda")
# Map ckpt filename → (model, mode). e.g. "gbt_realistic.ckpt.json" → ("gbt", "realistic").
_CKPT_RE = re.compile(r"^(?P<model>[a-z_]+)_(?P<mode>realistic|oracle)\.ckpt\.json$")
def _pair_files(artifacts_dir: Path, reports_dir: Path
) -> list[tuple[str, str, Path, Path, Path | None]]:
"""Find (model, mode, ckpt.json, sidecar, train.json or None) tuples."""
out: list[tuple[str, str, Path, Path, Path | None]] = []
for ckpt in sorted(artifacts_dir.glob("*.ckpt.json")):
m = _CKPT_RE.match(ckpt.name)
if not m:
log.warning("skipping unexpected file: %s", ckpt.name)
continue
model, mode = m.group("model"), m.group("mode")
# Sidecar: .pt for NN, .xgb.json for GBT
for suf in (".pt", ".xgb.json"):
sidecar = ckpt.with_suffix("").with_suffix(suf)
if sidecar.exists():
break
else:
log.warning("no sidecar (.pt or .xgb.json) found for %s; skipping",
ckpt.name)
continue
# Training report (optional)
train_json: Path | None = None
for cand in (
reports_dir / f"{model}_{mode}_train.json",
reports_dir / f"{model}_{mode}_pretrain.json",
):
if cand.exists():
train_json = cand
break
out.append((model, mode, ckpt, sidecar, train_json))
return out
def _bundle(model: str, mode: str, ckpt: Path, sidecar: Path,
train_json: Path | None) -> tuple[bytes, str]:
"""Produce the .tar.zst payload + its sha256."""
buf = io.BytesIO()
cctx = zstd.ZstdCompressor(level=10)
with cctx.stream_writer(buf) as zw:
with tarfile.open(fileobj=zw, mode="w|") as tar:
tar.add(ckpt, arcname=ckpt.name)
tar.add(sidecar, arcname=sidecar.name)
if train_json is not None:
tar.add(train_json, arcname=train_json.name)
payload = buf.getvalue()
return payload, hashlib.sha256(payload).hexdigest()
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--artifacts", required=True, type=Path,
help="dir holding *.ckpt.json + sidecar files (e.g. "
"./artifacts-lambda)")
ap.add_argument("--reports", required=True, type=Path,
help="dir holding *_train.json + *_pretrain.json")
ap.add_argument("--receiver", default="http://127.0.0.1:8445",
help="trainer-receiver base URL")
ap.add_argument("--manifest", type=Path,
default=Path("etc/training_manifest.toml.example"),
help="canonical manifest — used to map (model, mode) → job_id")
ap.add_argument("--as-host", default="lambda-a100",
help="X-Lab-Host value for the upload (the worker name "
"shown in cis490-jobs)")
ap.add_argument("--operator-token-env", default="CIS490_OPERATOR_TOKEN")
ap.add_argument("--dry-run", action="store_true")
ap.add_argument("--log-level", default="INFO")
args = ap.parse_args()
logging.basicConfig(
level=args.log_level,
format="%(asctime)s %(levelname)s %(name)s %(message)s",
)
pairs = _pair_files(args.artifacts, args.reports)
if not pairs:
log.error("no (ckpt.json, sidecar) pairs under %s", args.artifacts)
return 1
log.info("found %d trained (model, mode) pairs", len(pairs))
# Map (model, mode, hyper, …) → job_id by reading the manifest.
man = load_manifest(args.manifest)
job_index: dict[tuple[str, str], str] = {}
for j in man.jobs:
job_index[(j.model, j.mode)] = j.job_id
import os
operator_token = os.environ.get(args.operator_token_env)
client = FleetClient(args.receiver, host_id=args.as_host,
operator_token=operator_token)
# Pre-flight: receiver alive?
try:
rc, _ = client._request("GET", "/v1/health")
except Exception as e:
log.error("trainer-receiver not reachable at %s: %s",
args.receiver, e)
return 1
n_ok = 0
n_skipped = 0
n_failed = 0
for model, mode, ckpt, sidecar, train_json in pairs:
key = (model, mode)
job_id = job_index.get(key)
if job_id is None:
log.warning("no manifest entry for (%s, %s); skipping", model, mode)
n_skipped += 1
continue
# Make sure the queue knows about this job. If it's currently
# completed by an earlier run, requeue first so we can re-upload.
existing = next((r for r in client.list_jobs()
if r.get("job_id") == job_id), None)
if existing is None:
log.warning("queue has no row for job_id=%s — did you "
"`cis490-jobs reload` after editing the manifest?", job_id)
n_skipped += 1
continue
if existing.get("status") == "completed":
log.info("job %s already completed; requeueing for re-upload", job_id)
if not args.dry_run:
if not client.requeue(job_id):
log.warning("requeue failed for %s", job_id)
n_failed += 1
continue
# Claim it as if we were a worker
capability = {"cuda_available": True, "cpu_cores": 8,
"ram_available_gib": 32, "cuda_devices": [
{"name": "Lambda A100", "vram_total_gib": 40,
"vram_free_gib": 35}
]}
if not args.dry_run:
claim = client.claim(capability)
if not claim or claim.get("job_id") != job_id:
# claim_next picks by priority, not by id, so we may have
# claimed a different job. Manually mark this one in flight.
if claim and claim.get("job_id"):
# Release the unrelated claim by failing it back —
# cleanest way to put it back in pending.
client.fail(claim["job_id"], error="lambda-ingest race; releasing")
# Try once more with a more targeted approach: requeue + claim
client.requeue(job_id)
claim = client.claim(capability)
if not claim or claim.get("job_id") != job_id:
log.warning("could not claim job_id=%s after requeue; "
"skipping", job_id)
n_skipped += 1
continue
# Bundle + upload + complete
payload, sha = _bundle(model, mode, ckpt, sidecar, train_json)
log.info("uploading %s_%s job_id=%s sha=%s size=%dKiB",
model, mode, job_id, sha[:12], len(payload) // 1024)
if args.dry_run:
log.info(" (dry-run; skipping HTTP)")
n_ok += 1
continue
# PUT /v1/model/{job_id}
code, body = client._request(
"PUT", f"/v1/model/{job_id}",
body=payload,
extra_headers={
"x-content-sha256": sha,
"content-length": str(len(payload)),
"content-type": "application/octet-stream",
},
expect_status=(200, 201),
)
if code not in (200, 201):
log.error("upload failed: code=%d body=%r", code, body)
n_failed += 1
continue
artifact_id = (body or {}).get("artifact_id", sha)
# Mark complete
if not client.complete(job_id, artifact_id=artifact_id):
log.warning("complete() returned false for %s", job_id)
n_failed += 1
continue
n_ok += 1
print()
print(f" uploaded: {n_ok}")
print(f" skipped: {n_skipped}")
print(f" failed: {n_failed}")
print()
print("status: cis490-jobs status")
print("models: ls /var/lib/cis490/models/")
return 0 if n_failed == 0 else 2
if __name__ == "__main__":
raise SystemExit(main())