src/display/utils.py

from dataclasses import dataclass, make_dataclass
from enum import Enum

import pandas as pd

from src.display.about import Tasks


def fields(raw_class):
    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]


# These classes are for user facing column names,
# to avoid having to change them all around the code
# when a modif is needed
@dataclass
class ColumnContent:
    name: str
    type: str
    displayed_by_default: bool
    hidden: bool = False
    never_hidden: bool = False
    dummy: bool = False


## Leaderboard columns
auto_eval_column_dict = [["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)],
                         ["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)],
                         ["model_report", ColumnContent, ColumnContent("Report", "markdown", True, never_hidden=True)]
                         ]
# Init
# Scores
for task in Tasks:
    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])
# Model information
auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
# auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])
# auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])
# auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])
# auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False, dummy=True)])
# auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("Params (B)", "number", False)])

# auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False, dummy=True)])
auto_eval_column_dict.append(["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False, dummy=True)])
# auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False, dummy=True)])
# Dummy column for the search bar (hidden by the custom CSS)
auto_eval_column_dict.append(["dummy", ColumnContent, ColumnContent("model_name_for_query", "str", False, dummy=True)])

# We use make dataclass to dynamically fill the scores from Tasks
AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)


# For the queue columns in the submission tab
@dataclass(frozen=True)
class EvalQueueColumn:  # Queue column
    model = ColumnContent("model", "markdown", True)
    revision = ColumnContent("revision", "str", True)
    # private = ColumnContent("private", "bool", True)
    # precision = ColumnContent("precision", "str", True)
    # weight_type = ColumnContent("weight_type", "str", "Original")
    status = ColumnContent("status", "str", True)


# All the model information that we might need
@dataclass
class ModelDetails:
    name: str
    display_name: str = ""
    symbol: str = ""  # emoji


class ModelType(Enum):
    OPEN = ModelDetails(name="Publicly Available", symbol="🟢")
    Unknown = ModelDetails(name="Private", symbol="🔒")

    def to_str(self, separator=" "):
        return f"{self.value.symbol}{separator}{self.value.name}"

    @staticmethod
    def from_str(type):
        if "open" in type or "🟢" in type:
            return ModelType.OPEN
        return ModelType.Unknown


class WeightType(Enum):
    Adapter = ModelDetails("Adapter")
    Original = ModelDetails("Original")
    Delta = ModelDetails("Delta")


class Precision(Enum):
    float16 = ModelDetails("float16")
    bfloat16 = ModelDetails("bfloat16")

    qt_gptq_3bit = ModelDetails("GPTQ-3bit")
    qt_gptq_4bit = ModelDetails("GPTQ-4bit")
    qt_gptq_8bit = ModelDetails("GPTQ-8bit")
    qt_awq_3bit = ModelDetails("AWQ-3bit")
    qt_awq_4bit = ModelDetails("AWQ-4bit")
    qt_awq_8bit = ModelDetails("AWQ-8bit")

    Unknown = ModelDetails("🔒")

    def from_str(precision):
        if precision in ["torch.float16", "float16"]:
            return Precision.float16
        if precision in ["torch.bfloat16", "bfloat16"]:
            return Precision.bfloat16
        if precision in ["GPTQ-3bit"]:
            return Precision.qt_gptq_3bit
        if precision in ["GPTQ-4bit"]:
            return Precision.qt_gptq_4bit
        if precision in ["GPTQ-8bit"]:
            return Precision.qt_gptq_8bit
        if precision in ["AWQ-3bit"]:
            return Precision.qt_awq_3bit
        if precision in ["AWQ-4bit"]:
            return Precision.qt_awq_4bit
        if precision in ["AWQ-8bit"]:
            return Precision.qt_awq_8bit
        return Precision.Unknown


# Column selection
COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden]
COLS_LITE = [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]
TYPES_LITE = [c.type for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]

EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]

BENCHMARK_COLS = [t.value.col_name for t in Tasks]

NUMERIC_INTERVALS = {
    "🔒": pd.Interval(-1, 0, closed="right"),
    "~1.5": pd.Interval(0, 2, closed="right"),
    "~3": pd.Interval(2, 4, closed="right"),
    "~7": pd.Interval(4, 9, closed="right"),
    "~13": pd.Interval(9, 20, closed="right"),
    "~35": pd.Interval(20, 45, closed="right"),
    "~60": pd.Interval(45, 70, closed="right"),
    "70+": pd.Interval(70, 10000, closed="right"),
}