src/leaderboard/read_evals.py

import glob
import json
import logging
import os
import traceback
from dataclasses import dataclass
from typing import Optional

import dateutil
import numpy as np

from huggingface_hub import ModelCard

from src.display.formatting import make_clickable_model
from src.display.utils import auto_eval_cols, ModelType, Tasks, Precision, WeightType, QuantType, WeightDtype, ComputeDtype 

logger = logging.getLogger(__name__)


@dataclass
class EvalResult:
    # Also see src.display.utils.AutoEvalColumn for what will be displayed.
    eval_name: str # org_model_precision (uid)
    full_model: str # org/model (path on hub)
    org: str 
    model: str
    revision: str # commit hash, "" if main
    results: dict
    quant_type: QuantType = QuantType.Unknown
    precision: Precision = Precision.Unknown
    weight_dtype: WeightDtype = WeightDtype.Unknown
    compute_dtype: ComputeDtype = ComputeDtype.Unknown
    double_quant: bool = False 
    model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
    weight_type: WeightType = WeightType.Original # Original or Adapter
    architecture: str = "Unknown" # From config file
    license: str = "?"
    likes: int = 0
    num_params: int = 0
    model_size: int = 0
    group_size: int = -1
    date: str = "" # submission date of request file
    still_on_hub: bool = True
    is_merge: bool = False
    flagged: bool = False
    status: str = "Finished"
    tags: Optional[list] = None
    result_file: str = ""

    @classmethod
    def init_from_json_file(cls, json_filepath):
        """Inits the result from the specific model result file"""

        result_file = "/".join(json_filepath.split("/")[2:])
        with open(json_filepath) as fp:
            data = json.load(fp)

        # We manage the legacy config format
        config = data.get("config_general")
        if not isinstance(config, dict):
            raise ValueError(f"Missing or invalid config_general in {json_filepath}")

        # Precision
        precision = Precision.from_str(config.get("precision", "4bit"))
        quant_type = QuantType.from_str(str(config.get("quant_type", "GPTQ")))
        weight_dtype = WeightDtype.from_str(data["task_info"].get("weight_dtype", "int4"))
        compute_dtype = ComputeDtype.from_str(data["task_info"].get("compute_dtype", "bfloat16"))
        
        model_params = round(float(config["model_params"]), 2)
        model_size = round(float(config["model_size"]), 2)
        
        if data.get("quantization_config", None):
            double_quant = data["quantization_config"].get("bnb_4bit_use_double_quant", False)
            group_size = data["quantization_config"].get("group_size", -1)
        else:
            double_quant = False
            group_size = -1

        local = config.get("local", False)
        if not local:
            local = data["task_info"].get("local", False)

        # Get model and org
        org_and_model = config.get("model_name")
        org_and_model = org_and_model.split("/", 1)

        if local and org_and_model[0] != "Intel":
            org_and_model = config.get("model_name").split("/")
            org_and_model = ["local", org_and_model[-1]]
            quant_type = QuantType.autoround

        if len(org_and_model) == 1:
            org = None
            model = org_and_model[0]
            result_key = f"{model}_{precision.value.name}"
        else:
            org = org_and_model[0]
            model = org_and_model[1]
            result_key = f"{org}_{model}_{precision.value.name}"
        full_model = "/".join(org_and_model)

        # Extract results
        results = {}
        for task in Tasks:
            task = task.value
            if task.benchmark == "mmlu":
                mmlu_data = data["results"].get("harness|mmlu|0")
                if mmlu_data is None:
                    continue
                accs = np.array([mmlu_data.get(task.metric)])
            else:
                accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark in k])
            if accs.size == 0 or any([acc is None for acc in accs]):
                continue

            mean_acc = np.mean(accs) * 100.0
            mean_acc = round(mean_acc, 2)
            results[task.benchmark] = mean_acc

        return cls(
            eval_name=result_key,
            full_model=full_model,
            org=org,
            model=model,
            results=results,
            precision=precision,
            quant_type=quant_type,
            weight_dtype=weight_dtype,
            compute_dtype=compute_dtype,
            double_quant=double_quant,
            revision=config.get("model_sha", "main"),
            num_params=model_params,
            model_size=model_size,
            group_size=group_size,
            result_file=result_file
        )

    def update_with_request_file(self, requests_path):
        """Finds the relevant request file for the current model and updates info with it"""
        request_file = get_request_file_for_model(requests_path, self.full_model,
                self.quant_type.value.name, self.precision.value.name,
                self.weight_dtype.value.name, self.compute_dtype.value.name)

        try:
            with open(request_file, "r") as f:
                request = json.load(f)
            self.date = request.get("submitted_time", "")
            self.architecture = request.get("architectures", "Unknown")
            self.status = request.get("status", "Failed")
        except Exception as e:
            self.status = "Failed"
            logger.warning("Could not find request file for %s/%s: %s", self.org, self.model, e)

    def update_with_dynamic_file_dict(self, file_dict):
        self.license = file_dict.get("license", "?")
        self.likes = file_dict.get("likes", 0)
        self.still_on_hub = file_dict["still_on_hub"]
        self.tags = file_dict.get("tags", [])
        self.flagged = any("flagged" in tag for tag in self.tags)
        

    def to_dict(self):
        """Converts the Eval Result to a dict compatible with our dataframe display"""
        
        valid_results = [v for v in self.results.values() if v is not None]
        average = sum(valid_results) / len(valid_results) if valid_results else 0
        
        data_dict = {
            "eval_name": self.eval_name,
            "date": self.date, 
            auto_eval_cols.precision.name: self.precision.value.name,
            auto_eval_cols.quant_type.name: self.quant_type.value.name,
            auto_eval_cols.model_type_symbol.name: self.quant_type.value.symbol,
            auto_eval_cols.weight_dtype.name: self.weight_dtype.value.name,
            auto_eval_cols.compute_dtype.name: self.compute_dtype.value.name,
            auto_eval_cols.model.name: make_clickable_model(self.full_model, self.result_file),
            auto_eval_cols.revision.name: self.revision,
            auto_eval_cols.average.name: average,
            auto_eval_cols.model_size.name: self.model_size,
            auto_eval_cols.dummy.name: self.full_model,
        }

        data_dict[auto_eval_cols.still_on_hub.name] = self.still_on_hub
        data_dict[auto_eval_cols.flagged.name] = self.flagged
        
        if hasattr(auto_eval_cols, "double_quant"):
            data_dict[auto_eval_cols.double_quant.name] = self.double_quant
        if hasattr(auto_eval_cols, "architecture"):
            data_dict[auto_eval_cols.architecture.name] = self.architecture
        if hasattr(auto_eval_cols, "params"):
            data_dict[auto_eval_cols.params.name] = self.num_params
        if hasattr(auto_eval_cols, "license"):
            data_dict[auto_eval_cols.license.name] = self.license
        if hasattr(auto_eval_cols, "likes"):
            data_dict[auto_eval_cols.likes.name] = self.likes
        if hasattr(auto_eval_cols, "group_size"):
            data_dict[auto_eval_cols.group_size.name] = self.group_size
        
        if hasattr(auto_eval_cols, "merged"):
            data_dict[auto_eval_cols.merged.name] = "merge" in (self.tags if self.tags else [])
        if hasattr(auto_eval_cols, "moe"):
            data_dict[auto_eval_cols.moe.name] = ("moe" in (self.tags if self.tags else [])) or "moe" in self.full_model.lower()

        for task in Tasks:
            data_dict[task.value.col_name] = self.results.get(task.value.benchmark, 0)

        return data_dict



def get_request_file_for_model(requests_path, model_name,
        quant_type, precision, weight_dtype, compute_dtype):
    """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED"""
    request_files = os.path.join(
        requests_path,
        f"{model_name}_eval_request_*.json",
    )
    request_files = glob.glob(request_files)

    request_file = ""
    request_files = sorted(request_files, reverse=True)
    for tmp_request_file in request_files:
        with open(tmp_request_file, "r") as f:
            req_content = json.load(f)
            if (
                req_content["status"] in ["Finished"]
                and req_content["precision"] == precision.split(".")[-1]
                and str(req_content["quant_type"]) == quant_type
                and req_content["weight_dtype"] == weight_dtype.split(".")[-1]
                and req_content["compute_dtype"] == compute_dtype.split(".")[-1]
            ):
                request_file = tmp_request_file
            elif (
                req_content["status"] in ["Finished"]
                and req_content["precision"] == precision.split(".")[-1]
                and quant_type == "AutoRound"
                and req_content["weight_dtype"] == weight_dtype.split(".")[-1]
                and req_content["compute_dtype"] == compute_dtype.split(".")[-1]
            ):
                request_file = tmp_request_file
    return request_file


def get_raw_eval_results(results_path: str, requests_path: str, dynamic_path: str) -> list[EvalResult]:
    """From the path of the results folder root, extract all needed info for results"""
    model_result_filepaths = []

    for root, _, files in os.walk(results_path):
        result_files = [f for f in files if f.startswith("results_") and f.endswith(".json")]
        if len(result_files) == 0:
            continue

        try:
            result_files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
        except Exception:
            result_files = [result_files[-1]]

        for file in result_files:
            model_result_filepaths.append(os.path.join(root, file))

    dynamic_data = {}
    if os.path.exists(dynamic_path):
        with open(dynamic_path) as f:
            dynamic_data = json.load(f)

    eval_results = {}
    for model_result_filepath in model_result_filepaths:
        try:
            eval_result = EvalResult.init_from_json_file(model_result_filepath)
        except Exception as e:
            logger.warning("Skipping malformed eval result %s: %s", model_result_filepath, e)
            continue
        
        eval_result.update_with_request_file(requests_path)
        
        if eval_result.full_model in dynamic_data:
            eval_result.update_with_dynamic_file_dict(dynamic_data[eval_result.full_model])
            if "meta-llama" in eval_result.full_model: 
                eval_result.still_on_hub = True

        eval_name = eval_result.eval_name
        if eval_name in eval_results:
            eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})
        else:
            eval_results[eval_name] = eval_result

    results = []
    for v in eval_results.values():
        try:
            if v.status == "Finished":
                v.to_dict()  # Validate that to_dict() doesn't raise
                results.append(v)
        except Exception as e:
            logger.warning("Error processing %s: %s", v.eval_name, e)
            continue

    return results