Giuseppe Franco, Pablo Monteagudo-Lago, Ian Colbert, Nicholas Fraser, Michaela Blott: Improving Quantization with Post-Training Model Expansion https://arxiv.org/abs/2503.17513 https://arxiv.org/pdf/2503.17513 https://arxiv.org/html/2503.17513

Some interesting news on the Brevitas front!

We recently published a small paper analyzing how post-training model expansion can add more flexibility during quantization time.

It is possible to improve the accuracy of your quantized model by slightly increasing your model size, without retraining.

Cecilia Sala, a prominent Italian journalist, was arrested in Iran and jailed in the country’s infamous Evin prison after spending days reporting in Tehran, Italian officials said. The reason for the arrest has not been made public.

Thanks everyone for this work and the insights about the most used transformers models!

It helps to put things in perspective and step away from the sometimes unjustified LLM hype.

If you're at #NeurIPS, Ian Colbert will give a talk about Brevitas and how #AMD is preparing for the next generation of data types!

If you're interested in accumulation aware quantization or you want to see some interesting results on 4bit LLM quantization with Integer/minifloat/MX, check it out!

The Lichess database of games, puzzles, and engine evaluations is now on @hf.co - https://huggingface.co/Lichess. Billions of chess data points to download, query, and stream and we're excited to see what you'll build with it! ♟️ 🤗

Good, published, benchmarks of machine learning / data science is crucial.

But so hard.
Well-cited "SOTA" methods typically crash often. They tend to be very computational expensive. Both make a systematic study impossible.

Finally, reviewers always ask for more methods, and more "SOTA".

FYI, here's the entire code to create a dataset of every single bsky message in real time:

```
from atproto import *
def f(m): print(m.header, parse_subscribe_repos_message())
FirehoseSubscribeReposClient().start(f)
```

While we prepare the next release of Brevitas, check out the 0.11.0 release here: github.com/Xilinx/brevi...

Lots of new things, including:
- QAT/PTQ for minifloat
- QAT/PTQ for MX Int and MX Float
- Compatibility with huggingface's accelerate
- New PTQ techniques (HQQ, Channel splitting)

“a protest against the broader commodification of creative expertise in AI development…strategically framed to highlight OpenAI's alleged disregard for the economic value of artistic labor, echoing sentiments of discontent already prevalent in the AI ethics discourse”
www.forbes.com/sites/moinro...

Hello!

I'm Giuseppe, currently working at AMD on open-source ML quantization.

If that's something you're interested in, check out: github.com/Xilinx/brevi...

We support all the fanciest SOTA quantization algorithm, with a focus on modularity and composability.