File size: 4,903 Bytes
2a5f9fb
df66f6e
2a5f9fb
 
df66f6e
 
 
6e56e0d
df66f6e
6e56e0d
 
df66f6e
6e56e0d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2a5f9fb
6e56e0d
2a5f9fb
6e56e0d
 
 
 
 
 
 
 
 
 
 
 
2a5f9fb
 
 
 
b1a1395
2a5f9fb
 
 
 
 
 
 
 
 
 
 
fc1e99b
 
2a5f9fb
6e56e0d
 
 
 
 
 
 
 
 
 
df66f6e
 
 
6e56e0d
 
 
 
 
 
 
 
 
 
2a5f9fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
import json
import os
import re
from collections import defaultdict
from datetime import datetime, timedelta, timezone

import huggingface_hub
from huggingface_hub import ModelCard
from huggingface_hub.hf_api import ModelInfo
from transformers import AutoConfig

from src.envs import HAS_HIGHER_RATE_LIMIT


# ht to @Wauplin, thank you for the snippet!
# See https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/317
def check_model_card(repo_id: str) -> tuple[bool, str]:
    # Returns operation status, and error message
    try:
        card = ModelCard.load(repo_id)
    except huggingface_hub.utils.EntryNotFoundError:
        return False, "Please add a model card to your model to explain how you trained/fine-tuned it."

    # Enforce license metadata
    if card.data.license is None:
        if not ("license_name" in card.data and "license_link" in card.data):
            return False, (
                "License not found. Please add a license to your model card using the `license` metadata or a"
                " `license_name`/`license_link` pair."
            )

    # Enforce card content
    if len(card.text) < 200:
        return False, "Please add a description to your model card, it is too short."

    return True, ""


def is_model_on_hub(model_name: str, revision: str, token: str = None, trust_remote_code=False) -> tuple[bool, str]:
    try:
        AutoConfig.from_pretrained(model_name, revision=revision, trust_remote_code=trust_remote_code, token=token)
        return True, None

    except ValueError:
        return (
            False,
            "needs to be launched with `trust_remote_code=True`. For safety reason, we do not allow these models to be automatically submitted to the leaderboard.",
        )

    except Exception:
        return False, "was not found on hub!"


def get_model_size(model_info: ModelInfo, precision: str):
    size_pattern = size_pattern = re.compile(r"(\d\.)?\d+(b|m)")
    try:
        model_size = round(model_info.safetensors["total"] / 1e9, 3)
    except (AttributeError, TypeError ):
        try:
            size_match = re.search(size_pattern, model_info.modelId.lower())
            model_size = size_match.group(0)
            model_size = round(float(model_size[:-1]) if model_size[-1] == "b" else float(model_size[:-1]) / 1e3, 3)
        except AttributeError:
            return 0  # Unknown model sizes are indicated as 0, see NUMERIC_INTERVALS in app.py

    size_factor = 8 if (precision == "GPTQ" or "gptq" in model_info.modelId.lower()) else 1
    model_size = size_factor * model_size
    return model_size

def get_model_arch(model_info: ModelInfo):
    return model_info.config.get("architectures", "Unknown")

def user_submission_permission(submission_name, users_to_submission_dates, rate_limit_period, rate_limit_quota):
    org_or_user, _ = submission_name.split("/")
    if org_or_user not in users_to_submission_dates:
        return True, ""
    submission_dates = sorted(users_to_submission_dates[org_or_user])

    time_limit = (datetime.now(timezone.utc) - timedelta(days=rate_limit_period)).strftime("%Y-%m-%dT%H:%M:%SZ")
    submissions_after_timelimit = [d for d in submission_dates if d > time_limit]

    num_models_submitted_in_period = len(submissions_after_timelimit)
    if org_or_user in HAS_HIGHER_RATE_LIMIT:
        rate_limit_quota = 2 * rate_limit_quota

    if num_models_submitted_in_period > rate_limit_quota:
        error_msg = f"Organisation or user `{org_or_user}`"
        error_msg += f"already has {num_models_submitted_in_period} model requests submitted to the leaderboard "
        error_msg += f"in the last {rate_limit_period} days.\n"
        error_msg += (
            "Please wait a couple of days before resubmitting, so that everybody can enjoy using the leaderboard 🤗"
        )
        return False, error_msg
    return True, ""


def already_submitted_models(requested_models_dir: str) -> set[str]:
    depth = 1
    file_names = []
    users_to_submission_dates = defaultdict(list)

    for root, _, files in os.walk(requested_models_dir):
        current_depth = root.count(os.sep) - requested_models_dir.count(os.sep)
        if current_depth == depth:
            for file in files:
                if not file.endswith(".json"):
                    continue
                with open(os.path.join(root, file), "r") as f:
                    info = json.load(f)
                    file_names.append(f"{info['model']}_{info['revision']}_{info['precision']}")

                    # Select organisation
                    if info["model"].count("/") == 0 or "submitted_time" not in info:
                        continue
                    organisation, _ = info["model"].split("/")
                    users_to_submission_dates[organisation].append(info["submitted_time"])

    return set(file_names), users_to_submission_dates