Close your editor before heavy jobs? The heavy job lives inside my editor

Short summary

A developer details debugging a local FLUX image generation pipeline on a 16GB M4 Mac that kept getting killed by memory watchdogs. The root cause was insufficient free memory at start, not output size. Dropping from 4bit to 3bit quantization shaved ~2.5GB off peak usage, but freeing memory before starting was the bigger lever. Includes a comparison table of memory conditions and outcomes.

• •FLUX 12B model on 16GB Mac kept getting killed by memory watchdogs despite safety gates
• •Root cause was insufficient free memory at start, not output resolution
• •3bit quantization reduced peak usage by ~2.5GB; freeing memory beforehand was the biggest lever

Generated with AI, which can make mistakes.