# [OPEN SOURCE] Fast IdeogramV4 + Qwen 3.X class model trainer. Train more in less time. — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/b4ab0e096f8e616b

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-03 00:45 ingested 2026-11-04 02:39

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xmet7t/open_source_fast_ideogramv4_qwen_3x_class_model/)

## retained text



```
https://preview.redd.it/lp5addtjlazh1.png?width=1672&format=png&auto=webp&s=f1438e4e2defa833c4d39792a0fbaa355d7eae80

As part of the effort I did on training AI models that are able to create great art, we've had to design two components of what makes a design good:

1. Text to image models to transform a caption into a image.
2. Customer request to caption / brief to then send to the AI model (which in turn generates it).

For both trains, I've found pre-existing trainers are quite slow in terms of performance and leave a lot of MFU underutilized (which leads to less tokens / images trained a second).

I've optimized this for BH200 (Datacenter GPUs), but this should work for H100 and likely also works well on A100, and can be adapted to work on RTX series GPUs as many things translate over to customer GPUs (although unfortunately I don't have any to test and develop this).

https://preview.redd.it/4ackh039mazh1.png?width=2688&format=png&auto=webp&s=cefdf9a703ba00c96c6642098a0329fdcd246ea9

The results are very good - on one GPU, our solution is 58% better than ms-swift (a optimized training kit), 76% better than unsloth, and 178% better than transformers + PEFT.

For example, for training a designer (takes a concept / idea, then refines it into a very detailed caption for the model to then render / generate) via Qwen3.5-9B, it does 12,523 tokens/second, quite a bit (against 989 TFLOPS, what the GPU can do).

It also has a 43.7% best-config number from our micro-batch sweep, so play around with it and sweep to maximize perf.

However, the Arrhenius supercomputer we've tested this on does power-capping, so the true MFU is 53.9% clock-adjusted.

Same data and setup, other trainers:

* ms-swift ≈ 24% (1.58× slower);
* **Unsloth ≈ 22% (1.76× slower);**
* transformers + PEFT ≈ 14% (2.78× slower).

To give some comparison numbers, i've made some graphs.

[Comparison graph versus other solutions](https://preview.redd.it/hh362qkgnazh1.png?width=2932&format=png&auto=webp&s=ef4878427b38e74e1cc4e407713f734c9d399b38)

It virtually has loss parity (<6e-4 loss difference, which is negligible), and this, in short, allows training more for less.

https://preview.redd.it/juifzpkjnazh1.png?width=2596&format=png&auto=webp&s=05bc3414ee866ee7646c4dc8c3c5fefdca8e177c

For the Ideogram V4 trainer , below is the performance analysis:

https://preview.redd.it/87ki938nnazh1.png?width=2800&format=png&auto=webp&s=5e6829554d57e735623b80f9d5b08a7808094d8f

Also, there are some ablations in terms of what increases or decreases step time.

https://preview.redd.it/sqdl4musnazh1.png?width=2660&format=png&auto=webp&s=b775c7cad6437424118d3445766c1a689324b8e8

Code here: [https://github.com/logolabs/perf\_kit](https://github.com/logolabs/perf_kit)

If you want to support us in developing and releasing more free tools, you can **star the repository**.

**Hope it is useful to you to help you improve your trains efficiency.**
```



document fetched 2026-11-04 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xmet7t/open_source_fast_ideogramv4_qwen_3x_class_model
