Endurance-100B-v1-exl2-5bpw

Quantization Notes:

bpw: 5 hb: 6 calibration_length: 8192 measurement_length: 8192

Quantization Code:

Posting this here for convenience in case anyone is interested or finds it useful. I run this code using a conda 3.12 python env in WSL 2 Ubuntu. Steps to run include creating conda env and installing / upgrading exllamav2, logging into huggingface using the "huggingface-cli login" terminal command, configuring the config.yaml file, then running the python script.

base_model_name: "Endurance-100B-v1"
input_model: "~/models/TheDrummer_Endurance-100B-v1"
download_output_dir: "~/models"
output_base_path: "~/models/exl2-converted"
hf_username: "NobodySpecial"
default_hb: 6  # Default head bits value
exllama_path: "~/exllamav2"

quantizations:
  - bpw: 5
    calibration_length: 8192  # Optional: specify calibration length in tokens
    measurement_length: 8192  # Optional: specify measurement length in tokens
  - bpw: 6
    hb: 8 # Optional
    calibration_length: 8192  # Optional: specify calibration length in tokens
    measurement_length: 8192  # Optional: specify measurement length in tokens

import yaml
import os
import sys
import subprocess
import logging
import re
from tqdm import tqdm
from pathlib import Path
from huggingface_hub import HfApi, create_repo, login, hf_hub_download

# Set up logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')

def run_command(command_list, timeout=300):
    try:
        process = subprocess.Popen(
            command_list,
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            text=True,
            bufsize=1,
            universal_newlines=True
        )

        while True:
            output = process.stdout.readline()
            if output == '' and process.poll() is not None:
                break
            if output:
                logging.info(output.strip())

        rc = process.poll()
        if rc != 0:
            error_output = process.stderr.read()
            logging.error(f"Error executing command: {' '.join(command_list)}")
            logging.error(f"Error output: {error_output}")
            return False

        logging.info(f"Command executed successfully: {' '.join(command_list)}")
        return True
    except subprocess.TimeoutExpired:
        logging.error(f"Command timed out: {' '.join(command_list)}")
        process.kill()
        return False
    except Exception as e:
        logging.error(f"Unexpected error executing command: {' '.join(command_list)}")
        logging.error(f"Error: {str(e)}")
        return False

def validate_config(config):
    required_keys = [
        'exllama_path',
        'base_model_name',
        'input_model',
        'output_base_path',
        'hf_username',
        'quantizations'
    ]
    missing_keys = [key for key in required_keys if key not in config]
    if missing_keys:
        logging.error(f"Missing required configuration keys: {', '.join(missing_keys)}")
        return False

    # Validate exllama_path
    if not os.path.isdir(os.path.expanduser(config['exllama_path'])):
        logging.error(f"exllama_path does not exist or is not a directory: {config['exllama_path']}")
        return False

    # Validate output_base_path
    output_base_path = os.path.expanduser(config['output_base_path'])
    if not os.path.isdir(output_base_path):
        try:
            os.makedirs(output_base_path, exist_ok=True)
            logging.info(f"Created output_base_path directory: {output_base_path}")
        except OSError as e:
            logging.error(f"Failed to create output_base_path directory: {str(e)}")
            return False

    return True

def authenticate_hf():
    try:
        # Read the token from the local cache file
        token_path = os.path.expanduser("~/.cache/huggingface/token")
        with open(token_path, "r") as token_file:
            hf_token = token_file.read().strip()

        # Use the token to login
        login(token=hf_token)
        logging.info("Authenticated with Hugging Face successfully.")
    except Exception as e:
        logging.error(f"Failed to authenticate with Hugging Face: {str(e)}")
        return False
    return True

def sanitize_model_and_branch_names(model, branch):
    # Remove trailing slash if present
    model = model.rstrip('/')

    # Remove base URL if present
    if model.startswith("https://huggingface.co/"):
        model = model[len("https://huggingface.co/"):]

    # Split model and branch if provided in model name
    model_parts = model.split(":")
    model = model_parts[0]
    branch = model_parts[1] if len(model_parts) > 1 else branch

    # Use 'main' as default branch if not specified
    if branch is None:
        branch = "main"

    # Validate branch name
    if not re.match(r"^[a-zA-Z0-9._-]+$", branch):
        raise ValueError("Invalid branch name. Only alphanumeric characters, period, underscore and dash are allowed.")

    return model, branch

def download_model(model_name, branch_name, output_dir):
    # Sanitize model and branch names
    model_name, branch_name = sanitize_model_and_branch_names(model_name, branch_name)

    # Expand user directory if needed
    output_dir = os.path.expanduser(output_dir)

    # Initialize Hugging Face API
    api = HfApi()

    # Create output directory
    output_folder = Path(output_dir) / f"{'_'.join(model_name.split('/')[-2:])}"
    if branch_name != "main":
        output_folder = output_folder.with_name(f"{output_folder.name}_{branch_name}")
    output_folder.mkdir(parents=True, exist_ok=True)

    # Get file list
    try:
        files = api.list_repo_files(model_name, revision=branch_name)
    except Exception as e:
        logging.error(f"Error accessing repository: {e}")
        return None

    # Download files
    for file in tqdm(files, desc="Downloading files"):
        try:
            hf_hub_download(
                repo_id=model_name,
                filename=file,
                revision=branch_name,
                local_dir=output_folder,
                local_dir_use_symlinks=False
            )
        except Exception as e:
            logging.error(f"Error downloading {file}: {e}")

    logging.info(f"Model downloaded to {output_folder}")
    return output_folder

def resolve_input_model(config):
    input_model = os.path.expanduser(config['input_model'])
    if os.path.isdir(input_model):
        # Input model is a local directory
        logging.info(f"Using local model directory: {input_model}")
        return input_model
    else:
        # Input model is a Hugging Face repository
        logging.info(f"Input model is a Hugging Face model: {input_model}")
        download_output_dir = os.path.expanduser(config.get('download_output_dir', './models'))
        if not os.path.isdir(download_output_dir):
            try:
                os.makedirs(download_output_dir, exist_ok=True)
                logging.info(f"Created download_output_dir directory: {download_output_dir}")
            except OSError as e:
                logging.error(f"Failed to create download_output_dir directory: {str(e)}")
                sys.exit(1)
        model_name, branch_name = sanitize_model_and_branch_names(input_model, branch=None)
        output_folder = download_model(model_name, branch_name, download_output_dir)
        if output_folder is None:
            logging.error("Failed to download the model.")
            sys.exit(1)
        return str(output_folder)

def quantize_and_upload(config, input_model_path):
    exllama_path = os.path.expanduser(config['exllama_path'])
    base_model_name = config['base_model_name']
    output_base_path = os.path.expanduser(config['output_base_path'])
    hf_username = config['hf_username']
    default_hb = config.get('default_hb', 8)

    for quant_config in config['quantizations']:
        if 'bpw' not in quant_config:
            logging.warning("Skipping quantization config without 'bpw'.")
            continue

        bpw = quant_config['bpw']
        hb = quant_config.get('hb', default_hb)
        calibration_length = quant_config.get('calibration_length', 2048)
        measurement_length = quant_config.get('measurement_length', calibration_length)

        if not isinstance(calibration_length, int) or not isinstance(measurement_length, int):
            logging.error(f"Invalid calibration_length or measurement_length values. Expected integers.")
            continue

        if calibration_length <= 0 or measurement_length <= 0:
            logging.error(f"Invalid calibration_length or measurement_length values. Must be positive integers.")
            continue

        quant_name = f"{base_model_name}-exl2-{bpw}bpw"
        work_dir = os.path.join(output_base_path, base_model_name, f"{quant_name}-work")
        output_dir = os.path.join(output_base_path, base_model_name, quant_name)

        try:
            os.makedirs(work_dir, exist_ok=True)
            os.makedirs(output_dir, exist_ok=True)
            logging.info(f"Directories created or already exist: {work_dir}, {output_dir}")
        except OSError as e:
            logging.error(f"Failed to create directories for {quant_name}: {str(e)}")
            continue

        # Run quantization
        command_list = [
            "python", os.path.join(exllama_path, "convert.py"),
            "-i", input_model_path,
            "-o", work_dir,
            "-cf", output_dir,
            "-b", str(bpw),
            "-hb", str(hb),
            "-l", str(calibration_length),
            "-ml", str(measurement_length)
        ]
        if not run_command(command_list):
            logging.error(f"Quantization failed for {quant_name}. Skipping upload.")
            continue

        logging.info(f"Quantization completed for {quant_name}")

        # Try to upload to Hugging Face
        repo_name = f"{hf_username}/{quant_name}"
        try:
            create_repo(repo_name, repo_type="model", exist_ok=True)
            logging.info(f"Repository '{repo_name}' is ready.")
            api = HfApi()
            api.upload_folder(
                folder_path=output_dir,
                repo_id=repo_name,
                repo_type="model"
            )
            logging.info(f"Successfully uploaded {quant_name} to Hugging Face")
        except Exception as e:
            logging.error(f"Failed to upload {quant_name} to Hugging Face: {str(e)}")
            logging.info(f"Quantized model is still available locally at {output_dir}")

        logging.info(f"Completed processing for {quant_name}")

if __name__ == "__main__":
    config_path = "config.yaml"
    try:
        with open(config_path, "r") as f:
            config = yaml.safe_load(f)
            logging.info(f"Configuration loaded from {config_path}")
    except yaml.YAMLError as e:
        logging.error(f"Error parsing {config_path}: {str(e)}")
        sys.exit(1)
    except FileNotFoundError:
        logging.error(f"{config_path} not found. Please create a config file.")
        sys.exit(1)

    if not validate_config(config):
        logging.error("Configuration validation failed. Exiting.")
        sys.exit(1)

    if not authenticate_hf():
        logging.error("Hugging Face authentication failed. Exiting.")
        sys.exit(1)

    input_model_path = resolve_input_model(config)
    if not input_model_path:
        logging.error("Failed to resolve input model path. Exiting.")
        sys.exit(1)

    quantize_and_upload(config, input_model_path)
    logging.info("Script execution completed.")

---

base_model:

• TheDrummer/Lazarus-2407-100B

---

Join our Discord! https://discord.gg/Nbv9pQ88Xb

2500+ members strong 💪 Now with more channels! A hub for users and makers alike!

---

Endurance 100B v1 🎡

A finetune of Lazarus 2407 100B, a pruned Mistral Large 2407 123B!

Do not go gentle into that good night. Rage, rage against the dying of the light!

---

Links

• Original: https://huggingface.co/TheDrummer/Endurance-100B-v1
• GGUF: https://huggingface.co/TheDrummer/Endurance-100B-v1-GGUF
• iMatrix: https://huggingface.co/bartowski/Endurance-100B-v1-GGUF (recommended for smaller quants)

Arsenal (Supported Chat Templates)

• Metharme (Pygmalion in ST)
  • Creative, unhinged, unique

Favorite RP Format

*action* Dialogue *thoughts* Dialogue *narration* in 1st person PoV

Favorite Card

Audrey by thecooler

---

Technical Details

Refer to Lazarus 2407 100B for pruning details.

Endurance used the same hyperparameters as Behemoth. Training loss indicates that they are exactly the same albeit with lower confidence.

Notes on Lazarus 100B (base model for Endurance):

• Testers have noted that 100B seemed nearly identical to 123B.
• One tester said that only one minor mistake was made by the model, requiring a rewrite for failing to pick up on the nuance.
• Another tester went through a satisfying 32K playthrough without any issues.

Endurance 100B v1.0 has gone through additional RP & Instruct training.