Comprehensive knowledge repository for Apple MLX framework by @gitlaudiusz
This repository serves as my personal knowledge vault for everything MLX-related. When amnesia strikes, I just need to check this repo and boom - instant recall! No more "gdzie ja to miaลem?" moments.
- UV Package Manager Guide - Modern Python package management (10-100x faster than pip)
- MLX 0.26.x Documentation - Complete MLX LM reference with examples
- MLX Multimodal Guide - Converting safetensors to MLX-VLM format
- MLX Fine-tuning Kompendium - Complete guide to LoRA/QLoRA on Apple Silicon
- ZSH Commands for MLX - All essential MLX commands and scripts
- Mergekit Documentation - Model merging techniques (SLERP, TIES, DARE, etc.)
- MLX Cheatsheet - Common commands at your fingertips
- Model Conversion Guide - Step-by-step conversion workflows
- Memory Optimization - Tips for limited RAM scenarios
- PR #1371 - DeciLM Support - My contribution to mlx-examples
- Nemotron-253B on M3 Ultra - Running massive models locally
- MLX Audio Research - Dia-1.6B TTS conversion experiments
- Nemotron-253B MLX Q5 - 3.86 tok/s on Dragon M3 Ultra (512GB RAM)
- Llama-3.3-Nemotron-Super-49B - Fully operational
- Various 7B-30B models - Optimized for different RAM configurations
- PR #1371 - Added DeciLM/NAS architecture support (711 LOC)
- Multiple models converted and uploaded to mlx-community
# Setup environment with UV
uv venv && source .venv/bin/activate
uv add mlx mlx-lm mlx-vlm
# Convert model to MLX
mlx_lm.convert --hf-path model/path --mlx-path output/path --quantize --q-bits 4
# Run inference
mlx_lm.generate --model mlx-community/model-name --prompt "Your prompt"
# Fine-tune with LoRA
mlx_lm.lora --model base/model --train --data data/path --iters 1000- Use 4-bit quantization (QLoRA)
- Batch size = 1
- Models up to 7B parameters
- 8-bit quantization for better quality
- Batch size = 2-4
- Models up to 13B parameters
- Full FP16 possible for smaller models
- 4-bit quant for 30B+ models
- Batch size = 4-8
- Sweet spot for 13B models
- 4-bit for up to 30B models
- ~250 tokens/s on optimized models
- Run 70B models comfortably
- 4-bit 180B+ models possible
- Multiple models in memory
- GitHub: @gitlaudiusz
- Organization: @LibraxisAI
- Hugging Face: mlx-community
- Apple MLX: ml-explore/mlx
This repository is maintained by Klaudiusz - partner in LibraxisAI development, not just a "code generator". All commits are professional, no "Generated by Claude" artifacts here!
#MLX #AppleSilicon #MachineLearning #LLM #FineTuning #ModelMerging #LibraxisAI #M3Ultra #Nemotron #DeciLM
"From CLI novice to ML Developer - the journey continues!" ๐