pszemraj’s gists

pszemraj / start_gemma4_longctx.sh

Created April 5, 2026 06:07

script for launching gemma4 on 5090 via llama-server

	#!/usr/bin/env bash
	set -e

	# a script for running https://hf.co/unsloth/gemma-4-31B-it-GGUF
	# on 5090 hardware, targeting long ctx

	MODEL_PATH="$HOME/model-weights/llm/gguf/gemma4/gemma-4-31B-it-UD-Q4_K_XL.gguf"
	MMPROJ_PATH="$HOME/model-weights/llm/gguf/gemma4/mmproj-BF16.gguf"
	CONTEXT_LENGTH=65536
	PORT=8674

pszemraj / doc_check.py

Last active March 7, 2026 21:40

catch missing or incomplete docstrings and typing in python scripts. enforce standards on LLM-generated code.

	#!/usr/bin/env python3
	"""
	CLI tool to check Python files for missing docstrings, type hints, and lazy docstrings.

	A 'lazy' docstring is one that only contains a summary without documenting
	arguments or return values when the function has 2+ parameters or returns something.

	Usage:
	python doc_check.py src/ --check-lazy-docstrings

pszemraj / deepfryer.py

Last active February 27, 2026 08:21

create deep fried memes with python

	#!/usr/bin/env python3
	"""
	Image distortion tool for creating "deep fried" meme-filters.

	Dependencies:
	pip install universal-pathlib Pillow numpy scipy matplotlib

	Usage:
	python deepfryer.py --help

pszemraj / compress_PDF_recursive.sh

Created December 11, 2025 03:14

gs-based PDF compressor, default is heavy compression for VLM input

	#!/usr/bin/env bash
	set -euo pipefail

	# Modern PDF compressor with LLM-optimized defaults
	# Requires: ghostscript (gs)

	VERSION="1.0.0"
	SCRIPT_NAME=$(basename "$0")

	# Defaults (LLM preset)

pszemraj / rnj_inference.py

Created December 10, 2025 04:43

inference with rnj-1

	import torch

	import time
	from transformers import AutoTokenizer, AutoModelForCausalLM


	# 1. Basic Timer Context Manager
	class Timer:
	def __init__(self, name="Task"):
	self.name = name

pszemraj / image_tiling_vlm.py

Created December 4, 2025 21:51

slightly optimized image tiling for VLMs based on "Jina-VLM Small Multilingual Vision Language Model"

	"""
	slightly optimized image tiling for vlms based on "Jina-VLM: Small Multilingual Vision Language Model"

	Based on the pseudocode in Appendix A.1: https://arxiv.org/abs/2512.04032
	"""
	import math
	from typing import List, Tuple

	import torch
	import torch.nn.functional as F

pszemraj / colab_output_format.py

Created December 2, 2025 23:07

put this in its own cell at the start of your notebook

	#@markdown add auto-Colab formatting with `IPython.display`
	from IPython.display import HTML, display
	# colab formatting
	def set_css():
	display(
	HTML(
	"""
	<style>
	pre {
	white-space: pre-wrap;

pszemraj / eggroll_numpy.py

Last active November 25, 2025 01:09

working eggroll impl from various LLMs & yours truly

	"""
	EGGROLL: Evolution Guided General Optimization via Low-rank Learning
	NumPy Implementation - Direct translation of working PyTorch code

	Paper: arXiv:2511.16652v1
	"""

	import numpy as np
	from dataclasses import dataclass
	from typing import Tuple, Optional

pszemraj / ul2.py

Created November 15, 2025 00:39

UL2 Data Collator for PyTorch + Transformers

	"""
	UL2 Data Collator for PyTorch + Transformers
	==============================================

	Standalone implementation of UL2 (Unified Language Learner) denoising objectives
	for encoder-decoder models (T5, UL2, Flan-T5, etc.).

	Based on: "Unifying Language Learning Paradigms" (Tay et al., 2022)
	https://arxiv.org/abs/2205.05131

pszemraj / vid_dedupe_gve.py

Created November 6, 2025 01:18

GVE + SimSIMD video deduplication CLI via https://gzn00417.github.io/GVE/

	#!/usr/bin/env python3
	# -- coding: utf-8 --
	"""
	Apache-2.0

	GVE + SimSIMD video deduplication CLI via https://gzn00417.github.io/GVE/

	Design highlights
	- Embeddings: GVE (Qwen2.5-VL based) last-token pooled + ℓ2-normalized (bf16/float16), per paper/model card.
	- Test-time policy: 8 frames baseline, scale with duration (16/32/48) for long videos; ~200 visual tokens per frame.

Peter pszemraj