src/submission/submit.py

import json
from datetime import datetime, timezone

from src.display.formatting import styled_error, styled_message, styled_warning
from src.display.utils import EvalQueuedModel, LLMJpEvalVersion, VllmVersion
from src.envs import API, EVAL_REQUESTS_PATH, HF_TOKEN, QUEUE_REPO
from src.submission.check_validity import already_submitted_models, check_model_card, is_model_on_hub

REQUESTED_MODELS: set[EvalQueuedModel] = set()

LLM_JP_EVAL_VERSION = LLMJpEvalVersion.current.value.name
VLLM_VERSION = VllmVersion.current.value.name


def add_new_eval(
    model_id: str,
    revision: str,
    precision: str,
    model_type: str,
    add_special_tokens: str,
):
    global REQUESTED_MODELS
    if not REQUESTED_MODELS:
        REQUESTED_MODELS = already_submitted_models(EVAL_REQUESTS_PATH)

    revision = revision or "main"

    model_data = EvalQueuedModel(
        model=model_id,
        revision=revision,
        precision=precision,
        add_special_tokens=add_special_tokens,
        llm_jp_eval_version=LLM_JP_EVAL_VERSION,
        vllm_version=VLLM_VERSION,
    )

    if model_data in REQUESTED_MODELS:
        return styled_warning("This model has already been submitted with the same configuration.")

    if "/" in model_id:
        user_or_org, model_name = model_id.split("/")
    else:
        user_or_org, model_name = "", model_id

    current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")

    if model_type is None or model_type == "":
        return styled_error("Please select a model type.")

    # Is the model on the hub?
    model_on_hub, error, _ = is_model_on_hub(
        model_name=model_id, revision=revision, token=HF_TOKEN, test_tokenizer=True
    )
    if not model_on_hub:
        return styled_error(f'Model "{model_id}" {error}')

    # Is the model info correctly filled?
    try:
        model_info = API.model_info(repo_id=model_id, revision=revision)
    except Exception:
        return styled_error("Could not get your model information. Please fill it up properly.")

    # Were the model card and license filled?
    try:
        _ = model_info.cardData["license"]
    except Exception:
        return styled_error("Please select a license for your model")

    modelcard_OK, error_msg = check_model_card(model_id)
    if not modelcard_OK:
        return styled_error(error_msg)

    # Seems good, creating the eval
    print("Adding new eval")

    eval_entry = {
        "model_type": model_type,
        "model": model_id,
        "precision": precision,
        "revision": revision,
        "add_special_tokens": add_special_tokens,
        "llm_jp_eval_version": LLM_JP_EVAL_VERSION,
        "vllm_version": VLLM_VERSION,
        "status": "PENDING",
        "submitted_time": current_time,
    }

    print("Creating eval file")
    OUT_DIR = EVAL_REQUESTS_PATH / user_or_org
    OUT_DIR.mkdir(parents=True, exist_ok=True)
    out_file_name = f"{model_name}_eval_request_False_{precision}_{add_special_tokens}_{VLLM_VERSION}.json"
    out_path = OUT_DIR / out_file_name

    with out_path.open("w") as f:
        f.write(json.dumps(eval_entry))

    print("Uploading eval file")
    API.upload_file(
        path_or_fileobj=out_path,
        path_in_repo=out_path.relative_to(EVAL_REQUESTS_PATH).as_posix(),
        repo_id=QUEUE_REPO,
        repo_type="dataset",
        commit_message=f"Add {model_id} to eval queue",
    )
    REQUESTED_MODELS.add(model_data)

    # Remove the local file
    out_path.unlink()

    return styled_message(
        "Your request has been submitted to the evaluation queue!\nPlease wait for up to an hour for the model to show in the PENDING list."
    )