run_suite.py

"""Run disjoint checkpoint/task assignments on Kaggle, Colab, or another GPU."""
import argparse
from dataclasses import asdict
import hashlib
import json
import os
from pathlib import Path
import subprocess
import sys
import time
from runner import DiffusionEvalConfig, evaluation_units, write_json


def assignments(catalog, worker_index=0, num_workers=1, tasks=None, run_ppl=True):
    if num_workers < 1 or not 0 <= worker_index < num_workers:
        raise ValueError("Require 0 <= worker-index < num-workers")
    cfg = DiffusionEvalConfig()
    cfg.skip_ppl = not run_ppl
    work = []
    for item in catalog:
        if not item.get("ready", True):
            continue
        family = "dream" if item.get("model", "").startswith("Dream-org/") or "structured-dream-" in item["repo_id"] else "llada"
        work.extend({"repo_id": item["repo_id"], "revision": item.get("revision"),
                     "unit": unit, "family": family}
                    for unit in evaluation_units(cfg, family, selected_tasks=tasks))
    return work[worker_index::num_workers]


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--catalog", default=str(Path(__file__).with_name("models.json")))
    parser.add_argument("--repos", nargs="+", help="Select one or more original or fine-tuned public repositories")
    parser.add_argument("--worker-index", type=int, default=0)
    parser.add_argument("--num-workers", type=int, default=1)
    parser.add_argument("--tasks", nargs="+")
    parser.add_argument("--skip-ppl", action="store_true")
    parser.add_argument("--plan", action="store_true")
    parser.add_argument("--batch-size", type=int)
    parser.add_argument("--dtype", choices=["auto", "bfloat16", "float16", "float32"])
    parser.add_argument("--device", default="cuda")
    parser.add_argument("--device-map", choices=["auto", "single"], default="auto")
    parser.add_argument("--gpu-memory-ratio", type=float, default=.65)
    parser.add_argument("--output", default="results")
    parser.add_argument("--cache-dir", default=".eval_cache")
    parser.add_argument("--offload-dir", default=".eval_offload")
    parser.add_argument("--time-budget-seconds", type=float, default=8*3600)
    parser.add_argument("--request-chunk-size", type=int, default=1)
    parser.add_argument("--require-finetuned-slice", action="store_true")
    args = parser.parse_args()
    catalog = json.loads(Path(args.catalog).read_text())
    if args.repos:
        selected = {r["repo_id"]: r for r in catalog}
        catalog = [selected.get(repo, {"repo_id": repo, "ready": True}) for repo in args.repos]
    work = assignments(catalog, args.worker_index, args.num_workers, args.tasks, not args.skip_ppl)
    if args.plan:
        print(json.dumps({"assignments": work, "units": len(work),
                          "shared_config": asdict(DiffusionEvalConfig())}, indent=2))
        return
    deadline = time.monotonic() + args.time_budget_seconds
    grouped = {}
    for item in work:
        grouped.setdefault((item["repo_id"], item["revision"]), []).append(item["unit"])
    status = {"worker_index": args.worker_index, "num_workers": args.num_workers,
              "assigned_units": len(work), "models_attempted": [], "errors": []}
    runner = str(Path(__file__).with_name("runner.py"))
    for (repo, revision), units in grouped.items():
        remaining = deadline - time.monotonic()
        if remaining <= 0:
            break
        cmd = [sys.executable, runner, "--repo", repo, "--units", *units,
               "--device", args.device, "--device-map", args.device_map,
               "--gpu-memory-ratio", str(args.gpu_memory_ratio), "--output", args.output,
               "--cache-dir", args.cache_dir, "--offload-dir", args.offload_dir,
               "--time-budget-seconds", str(remaining), "--request-chunk-size", str(args.request_chunk_size)]
        if revision:
            cmd += ["--revision", revision]
        if args.batch_size is not None:
            cmd += ["--batch-size", str(args.batch_size)]
        if args.dtype:
            cmd += ["--dtype", args.dtype]
        if args.skip_ppl:
            cmd += ["--skip-ppl"]
        if args.require_finetuned_slice:
            cmd += ["--require-finetuned-slice"]
        status["models_attempted"].append(repo)
        write_json(Path(args.output) / ("worker_" + str(args.worker_index) + "_status.json"), status)
        result = subprocess.run(cmd)
        if result.returncode:
            status["errors"].append({"repo_id": repo, "exit_code": result.returncode})
            break
    write_json(Path(args.output) / ("worker_" + str(args.worker_index) + "_status.json"), status)
    if status["errors"]:
        raise SystemExit(status["errors"][-1]["exit_code"])


if __name__ == "__main__":
    main()