code/self_healing_finetuning/inference/deploy.py

from __future__ import annotations

from pathlib import Path
import json

from n21.config import write_json
from observability.audit_log import AuditLogger, utc_now
from observability.heartbeat import heartbeat
from observability.progress import ProgressState
from orchestrator.provider_routing import get_provider


def deploy(run_dir: Path, *, run_id: str, infer_provider: str, env: str, model_candidate: str, live: bool = False) -> dict[str, object]:
    logger = AuditLogger(run_dir / "logs" / "audit.jsonl")
    run_manifest_path = run_dir / "manifests" / "run_manifest.json"
    train_provider = "unknown"
    if run_manifest_path.exists():
        run_manifest = json.loads(run_manifest_path.read_text(encoding="utf-8"))
        train_provider = run_manifest.get("execution", {}).get("train_provider", "unknown")
    heartbeat(logger, ProgressState(run_id, 1, "deploy", "prepare_endpoint", 25, 25, 75, train_provider, infer_provider, model_candidate, next_action="create deployment manifest"))
    provider = get_provider(infer_provider)
    model_manifest = {"run_id": run_id, "model_id": f"linvest21-fin-fingpt-qa-{infer_provider}-core-v0.1.0"}
    handle = provider.deploy_endpoint(model_manifest, env)
    manifest = {
        "deployment_id": handle["deployment_id"],
        "run_id": run_id,
        "env": env,
        "infer_provider": infer_provider,
        "status": handle["status"],
        "endpoint_uri": handle["endpoint_uri"],
        "dry_run": not live,
        "created_at": utc_now(),
    }
    write_json(run_dir / "manifests" / "deployment_manifest.json", manifest)
    heartbeat(logger, ProgressState(run_id, 1, "deploy", "deployment_manifest", 100, 100, 85, train_provider, infer_provider, model_candidate, status="planned", next_action="run promote or rollback"))
    return manifest