main_backend.py

import argparse
import logging
import pprint
import os

from huggingface_hub import snapshot_download

import src.backend.run_eval_suite as run_eval_suite
import src.backend.manage_requests as manage_requests
import src.backend.sort_queue as sort_queue
import src.envs as envs

os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'

logging.basicConfig(level=logging.ERROR)
pp = pprint.PrettyPrinter(width=80)

PENDING_STATUS = "PENDING"
RUNNING_STATUS = "RUNNING"
FINISHED_STATUS = "FINISHED"
FAILED_STATUS = "FAILED"

snapshot_download(repo_id=envs.RESULTS_REPO, revision="main", local_dir=envs.EVAL_RESULTS_PATH_BACKEND, repo_type="dataset", max_workers=60)
snapshot_download(repo_id=envs.QUEUE_REPO, revision="main", local_dir=envs.EVAL_REQUESTS_PATH_BACKEND, repo_type="dataset", max_workers=60)


def run_auto_eval(args):
    if not args.reproduce:
        current_pending_status = [PENDING_STATUS]

        manage_requests.check_completed_evals(
            api=envs.API,
            checked_status=RUNNING_STATUS,
            completed_status=FINISHED_STATUS,
            failed_status=FAILED_STATUS,
            hf_repo=envs.QUEUE_REPO,
            local_dir=envs.EVAL_REQUESTS_PATH_BACKEND,
            hf_repo_results=envs.RESULTS_REPO,
            local_dir_results=envs.EVAL_RESULTS_PATH_BACKEND
        )
        logging.info("Checked completed evals")
        eval_requests = manage_requests.get_eval_requests(
            job_status=current_pending_status, hf_repo=envs.QUEUE_REPO, local_dir=envs.EVAL_REQUESTS_PATH_BACKEND
        )
        logging.info("Got eval requests")
        eval_requests = sort_queue.sort_models_by_priority(api=envs.API, models=eval_requests)
        logging.info("Sorted eval requests")

        print(f"Found {len(eval_requests)} {','.join(current_pending_status)} eval requests")
        print(eval_requests)
        if len(eval_requests) == 0:
            print("No eval requests found. Exiting.")
            return

        if args.model is not None:
            eval_request = manage_requests.EvalRequest(
                model=args.model,
                status=PENDING_STATUS,
                precision=args.precision
            )
            pp.pprint(eval_request)
        else:
            eval_request = eval_requests[0]
            pp.pprint(eval_request)

        run_eval_suite.run_evaluation(
            eval_request=eval_request,
            local_dir=envs.EVAL_RESULTS_PATH_BACKEND,
            results_repo=envs.RESULTS_REPO,
            batch_size=args.batch_size,
            device=envs.DEVICE,
            no_cache=True,
            need_check=not args.publish,
            write_results=args.update
        )
        logging.info("Eval finished, now setting status to finished")
    else:
        eval_request = manage_requests.EvalRequest(
            model=args.model,
            model_path=args.model_path,
            status=PENDING_STATUS,
            precision=args.precision
        )
        pp.pprint(eval_request)
        logging.info("Running reproducibility eval")

        run_eval_suite.run_evaluation(
            eval_request=eval_request,
            local_dir=envs.EVAL_RESULTS_PATH_BACKEND,
            results_repo=envs.RESULTS_REPO,
            batch_size=args.batch_size,
            device=envs.DEVICE,
            need_check=not args.publish,
            write_results=args.update,
            limit=args.limit,
            use_vllm=args.use_vllm,
            tensor_parallel_size=args.tensor_parallel_size,
        )
        logging.info("Reproducibility eval finished")


def main():
    parser = argparse.ArgumentParser(description="Run auto evaluation with optional reproducibility feature")

    # Optional arguments
    parser.add_argument("--reproduce", type=bool, default=False, help="Reproduce the evaluation results")
    parser.add_argument("--model", type=str, default=None, help="Your Model ID")
    parser.add_argument("--model_path", type=str, default=None, help="Full path of model")
    parser.add_argument("--precision", type=str, default="float16", help="Precision of your model")
    parser.add_argument("--publish", type=bool, default=False, help="whether directly publish the evaluation results on HF")
    parser.add_argument("--update", type=bool, default=False, help="whether to update google drive files")
    parser.add_argument("--limit", type=int, default=None, help="Limit on the number of items to process")
    parser.add_argument("--use_vllm", type=bool, default=False, help="Whether to infer with vllm or not")
    parser.add_argument("--tensor_parallel_size", type=int, default=1)
    parser.add_argument("--batch_size", type=int, default=1)

    args = parser.parse_args()

    run_auto_eval(args)


if __name__ == "__main__":
    main()