src/submission/submit.py

import json
import os
import ast
from datetime import datetime, timezone

from src.display.formatting import styled_error, styled_message, styled_warning
from src.envs import API, EVAL_REQUESTS_PATH, TOKEN, QUEUE_REPO
from src.submission.check_validity import (
    already_submitted_models,
    check_model_card,
    get_model_size,
    is_model_on_hub,
)
from src.display.utils import PromptTemplateName

REQUESTED_MODELS = None
USERS_TO_SUBMISSION_DATES = None

PLACEHOLDER_DATASET_WISE_NORMALIZATION_CONFIG = """{
    "NCBI" : {
        "" : "condition"
    },
    "CHIA" : {
        "" : "condition"
        "" : "drug"
        "" : "procedure"
        "" : "measurement"
    },
    "BIORED" : {
        "" : "condition"
        "" : "drug"
        "" : "gene"
        "" : "gene variant"
    },
    "BC5CDR" : {
        "" : "condition"
        "" : "drug"
    }
}

"""

def add_new_eval(
    model: str,
    # base_model: str,
    revision: str,
    # precision: str,
    # weight_type: str,
    model_arch: str,
    label_normalization_map: str,
    gliner_threshold:str,
    gliner_tokenizer_bool:str,
    prompt_template_name:str,
    model_type: str,
):
    """
    Saves request if valid else returns the error.
    Validity is checked based on -
        - model's existence on hub
        - necessary info on the model's card
        - label normalization is a valid python dict and contains the keys for all datasets
        - threshold for gliner is a valid float

    """
    global REQUESTED_MODELS
    global USERS_TO_SUBMISSION_DATES
    if not REQUESTED_MODELS:
        REQUESTED_MODELS, USERS_TO_SUBMISSION_DATES = already_submitted_models(EVAL_REQUESTS_PATH)

    user_name = ""
    model_path = model
    if "/" in model:
        user_name = model.split("/")[0]
        model_path = model.split("/")[1]

    # precision = precision.split(" ")[0]
    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.")
    
    model_type = model_type.split(":")[-1].strip()

    # Does the model actually exist?
    if revision == "":
        revision = "main"

    # # Is the model on the hub?
    # if weight_type in ["Delta", "Adapter"]:
    #     base_model_on_hub, error, _ = is_model_on_hub(model_name=base_model, revision=revision, token=TOKEN, test_tokenizer=True)
    #     if not base_model_on_hub:
    #         return styled_error(f'Base model "{base_model}" {error}')

    if not model_arch == "GLiNER Encoder":
        model_on_hub, error, _ = is_model_on_hub(model_name=model, revision=revision, token=TOKEN, test_tokenizer=True)
        if not model_on_hub:
            return styled_error(f'Model "{model}" {error}')
    else:
        if len(list(API.list_models(model_name=model))) !=1:
            return styled_error(f'Model "{model}" does not exist on the hub!')        
    

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

    model_size = get_model_size(model_info=model_info)

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

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

    # Verify the inference config now
    try:
        label_normalization_map = ast.literal_eval(label_normalization_map)
    except Exception as e:
        return styled_error("Please enter a valid json for the labe; normalization map")

    inference_config = {
        # "model_arch" : model_arch,
        "label_normalization_map": label_normalization_map,
    }

    match model_arch:
        case "Encoder":
            pass
        case "Decoder":
            if not prompt_template_name in [prompt_template.value for prompt_template in PromptTemplateName]:
                return styled_error("Prompt template name is invalid")
            inference_config = {
                **inference_config,
                "prompt_template_name": prompt_template_name,
            }
        case "GLiNER Encoder":
                try:
                    gliner_threshold = float(gliner_threshold)
                    gliner_tokenizer_bool = ast.literal_eval(gliner_tokenizer_bool)
                    inference_config = {
                        **inference_config,
                        "gliner_threshold": gliner_threshold,
                        "gliner_tokenizer_bool" : gliner_tokenizer_bool
                    }
                except Exception as e:
                    return styled_error("Please enter a valid float for the threshold")
        case _:
            return styled_error("Model Architecture is invalid")

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


    eval_entry = {
        "model": model,
        # "base_model": base_model,
        "revision": revision,
        # "precision": precision,
        # "weight_type": weight_type,
        "model_architecture": model_arch,
        "status": "PENDING",
        "submitted_time": current_time,
        "model_type": model_type,
        "likes": model_info.likes,
        "params": model_size,
        "license": license,
        "private": False,
        "inference_config":inference_config,
    }

    # Check for duplicate submission

    if f"{model}_{revision}" in REQUESTED_MODELS:
        return styled_warning("This model has been already submitted. Add the revision if the model has been updated.")

    print("Creating eval file")
    OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
    os.makedirs(OUT_DIR, exist_ok=True)
    out_path = f"{OUT_DIR}/{model_path}_{revision}_eval_request.json"

    with open(out_path, "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.split("eval-queue/")[1],
        repo_id=QUEUE_REPO,
        repo_type="dataset",
        commit_message=f"Add {model} to eval queue",
    )

    # Remove the local file
    os.remove(out_path)

    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."
    )