Artefact2

Which GGUF is right for me? (Opinionated)

Good question! I am collecting human data on how quantization affects outputs. See here for more information: ggerganov/llama.cpp#5962

In the meantime, use the largest that fully fits in your GPU. If you can comfortably fit Q4_K_S, try using a model with more parameters.

llama.cpp feature matrix

See the wiki upstream: https://github.com/ggerganov/llama.cpp/wiki/Feature-matrix

Artefact2

#!/usr/bin/env sh

URI=$1
BASE=$(basename $1)

[ -f ../models/$BASE-f16.gguf ] && exit 0

(. ../huggingface-cli/bin/activate && HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download $URI --local-dir ../models/$BASE --cache-dir ../models/$BASE/.hf-cache --exclude 'pytorch_model*' --exclude 'consolidated*' --resume-download) || exit 1

DTYPE=$(jq -r '.torch_dtype' < ../models/$BASE/config.json)

Artefact2

#!/usr/bin/env zsh

if test $# -ne 1; then
	echo "Tests against fake capacity flash storage. Script should print two identical checksums."
	echo "Usage: $0 /dev/sdX" >&2
	exit 1
fi

parallel --version >/dev/null 2>&1 || exit 1
hexdump --version >/dev/null 2>&1 || exit 1

Artefact2

#!/usr/bin/env php
<?php
$p = popen('btrfs-search-metadata file '.escapeshellarg($argv[1]), 'r');
$logsz = $physz = null;
$nbins = intval(shell_exec('tput cols')) - 3;
$bins = array_fill(0, $nbins, 0);
while(($line = fgets($p)) !== false) {
    $fields = explode(' ', substr($line, 0, -1));
    if($fields[0] === 'inode' && $fields[1] === 'objectid') {
        $logsz = intval($fields[8]);

Artefact2

-- Override mpv's default volume keybinds with decibel controls
-- Put this file (dbvol.lua) in ~/.config/mpv/scripts/
-- WARNING: positive gain values may cause clipping

-- Author: Romain "Artefact2" Dal Maso <romain.dalmaso@artefact2.com>
-- Released under the WTFPLv2

function round(n)
    -- https://stackoverflow.com/a/58411671
    return n + (2^52 + 2^51) - (2^52 + 2^51)

Artefact2

Artefact2

Considerations

• Tests were run in single-user mode, one at a time.
• Data was read from and written to tmpfs.
• borg-1.1rc2 seems about 10% slower than borg-1.0.11 (many possible reasons, borg-1.1rc2 binary comes from GitHub but 1.0.11 from the Arch repos)
• Still making sense of the auto,X,Y compression modes

Artefact2

<?php
/* Author: Romain Dal Maso <artefact2@gmail.com>
 *
 * This program is free software. It comes without any warranty, to the
 * extent permitted by applicable law. You can redistribute it and/or
 * modify it under the terms of the Do What The Fuck You Want To Public
 * License, Version 2, as published by Sam Hocevar. See
 * http://sam.zoy.org/wtfpl/COPYING for more details. */

if($argc !== 2) {

Artefact2

CHROMASTART=0
CHROMAEND=0

MAINSTART=0
MAINEND=0
NJOBS=8

BATCH=$$(( ($(END) - $(START)) / $(NJOBS) ))
LAST=$$(( $(NJOBS)-1 ))
BLENDFILES=$(wildcard *.blend)