---
title: "Automagic v2 Cuts Diffusion Model Training VRAM by 5x on Consumer GPUs"
subtitle: "Automagic v2 brings a fused backward pass and per-layer adaptive learning rates to AI Toolkit, slashing VRAM from 13.65GB to 2.75GB versus Prodigy on a 500M-param model."
url: "https://alphasignal.ai/news/automagic-v2-cuts-diffusion-model-training-vram-by-5x-on-consumer-gpus"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-06-23T16:43:48.000Z"
updated: "2026-06-23T19:13:05.345Z"
type: "Repo"
topics:
  - "Post Training"
  - "Open Source"
subtopics:
  - "Fine Tuning"
  - "Lora"
  - "Distillation"
upvotes: 140
source_url: "https://github.com/ostris/ai-toolkit"
comments: 0
truncated: true
---

# Automagic v2 Cuts Diffusion Model Training VRAM by 5x on Consumer GPUs

> Automagic v2 brings a fused backward pass and per-layer adaptive learning rates to AI Toolkit, slashing VRAM from 13.65GB to 2.75GB versus Prodigy on a 500M-param model.

[AI Toolkit](https://github.com/ostris/ai-toolkit), the open-source diffusion model finetuning suite from ostris, just shipped a significant update to its homegrown optimizer: Automagic v2\. The new version introduces a fused backward pass and per-layer adaptive learning rates, and it is now the most memory-efficient optimizer available in the toolkit. For anyone training LoRAs on consumer hardware, the numbers are hard to ignore.

## The VRAM problem that won't go away

Training large diffusion models like FLUX.1 on consumer GPUs has always been a balancing act between model quality and memory. For models like FLUX.1 that require 24GB+ VRAM, 8-bit optimizers are essential for reducing the footprint of optimizer states. The standard go-to for adaptive, learning-rate-free training has been Prodigy, but it comes at a steep memory cost. In a test training a 40-block, 500M parameter transformer, peak VRAM for Automagic v2 was 2.75GB, versus Prodigy which was 13.65GB. That is a nearly 5x reduction.

The root cause of high optimizer memory is straightforward: standard optimizers like Adam store momentum and variance buffers for every trainable parameter, typically in 32-bit float. 8-bit optimizers reduce memory consumption by storing optimizer state in 8-bit precision rather than the standard 32-bit floating-point format, which can reduce optimizer state memory by approximately 75%, enabling training of larger models or higher batch sizes on memory-constrained hardware.

## What Automagic v2 actually does

Automagic v2 automatically adjusts the learning rate per layer and has a fused backward pass. It is extremely memory efficient, and uses less VRAM than any other optimizer AI Toolkit supports. Those two features are the core of the update, and they work together to achieve the memory savings.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [ostris/ai-toolkit GitHub Repository](https://github.com/ostris/ai-toolkit)
- [Ostris on X: Automagic v2 announcement](https://x.com/ostrisai/status/2051698964299399602)
- [Ostris on X: VRAM benchmark results](https://x.com/ostrisai/status/2051698965712937240)
- [8-bit Optimizers | ostris/ai-toolkit | DeepWiki](https://deepwiki.com/ostris/ai-toolkit/18.2-8-bit-optimizers)
- [Automagic v2 not working at all · Issue #826 · ostris/ai-toolkit](https://github.com/ostris/ai-toolkit/issues/826)