Pinned
@Dropbox is acquiring our IP, and the Mobius core tech team is joining them. A huge milestone for us!
Mobius Labs
420 posts
Mobius Labs
@Mobius_Labs
Multimodal AI for the world's scale.
Proponents of Open Source and Open Intelligence.
mobiusml.github.io/blog/ for some of our recent work.
- Super thrilled to release our work on extreme quantization (1-bit and 2-bit)! We're starting with the Llama2-7b since it's a well-understood model. Check out our detailed blog post: mobiusml.github.io/1bit_blog/
GIF - Dropping soon: A 1-bit version of Llama7b with the new iteration of HQQ(+), along with a detailed blog post.
- AI powered computer vision technology that gives you deep visual analysis and unparalleled flexibility with a convenient interface that anyone can use. 🚀
- Another quantization method dropped in @huggingface transformers library ! Half Quadratic Quantization 🔥 HQQ implements on-the-fly quantization via fast robust optimization. It doesn’t require calibration data and can be used to quantize any model, up to 1-bit precision !
- Prioritize your visual content based on client needs to increase sales, identify additional revenue streams and reduce operational costs.🚀mobiuslabs.com/media-business
- Early preview of the new backend (torchao int4) for HQQ with transformers. Llama2-7B running at 150 tokens/sec on an RTX 4090 now. More details and code coming soon this week!
00:00 - Enrich your content with visual intelligence that understands clients and serves them with the most relevant visuals. Leverage our next-gen AI solution to organize, tag and search for visual content with total ease and privacy.👇
- Replying to @Mobius_LabsYou can find the models here: huggingface.co/collections/mo… and a Colab notebook to run the 1-bit version here: colab.research.google.com/drive/15A6sVvd…
- Our solution runs in your system with no internet required, giving you absolute privacy, speed and security. 🚀👇 mobiuslabs.com/media-business
- 🚀 Introducing two new kernels to HQQ! One is based on TorchAO and the other on Marlin. Currently supports only 4-bit models, achieving speeds up to 200 tokens/sec on 4090 and 60 tokens/sec on L4. 🔗 Get it here: github.com/mobiusml/hqq 📊 Colab: colab.research.google.com/drive/1uomMVKC…
- Our #imagerecognition solution works out of the box to deliver market-leading precision & speed. Curious? Try our Superhuman Vision ™ demo today. ct.mobiuslabs.com/register
- Replying to @Mobius_LabsA 2-bit model performs quite well. Specifically, the base Llama2-7B 2-bit model with HQQ+ outperforms the full-precision model on Wikitext. The chat model exceeds its full-precision counterpart on GSM8K with adequate math and reasoning data.
- Thrilled to contribute to the amazing work by Answer.AI on efficiently training large models (70B) with consumer GPUs. Putting our hopes on the next large models of consequence, trained & shared by the GPU-poor, now that per GPU memory is no longer a prerequisite.Today, with @Tim_Dettmers, @huggingface, & @Mobius_Labs, we're releasing FSDP/QLoRA, a new project that lets you efficiently train very large (70b) models on a home computer with consumer gaming GPUs. 1/🧵 answer.ai/posts/2024-03-…
