🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

🚀 Exciting news! We’ve just released a new LLM: Llama-3.1-Nemotron-51B = LLaMa-70B-Instruct + Block Distillation + NAS + Logics Distillation; Powered by a single H100 GPU with nearly the same accuracy! ⚡ This gives a 2.2x inference speed-up with MT Bench 8.99 ➡️ 8.94.

MaskLLM Learnable Semi-Structured Sparsity for Large Language Models discuss: huggingface.co/papers/2409.17… Large Language Models (LLMs) are distinguished by their massive parameter counts, which typically result in significant redundancy. This work introduces MaskLLM, a learnable

🚀 NeurIPS Conference Spotlight! 🥳 Imagine fine-tuning an LLM with just a sparsity mask! In our latest work, we freeze the LLM and use 2:4 structured sparsity to learn binary masks for each linear layer. Thanks to NVIDIA Ampere’s 2:4 sparsity, we can achieve up to 2x compute

Tomorrow we will be hosting the first Efficient DL for Foundation Model workshop tomorrow at #ECCV2024! The place will be Brown 3, at Milan time 2pm-6pm, with keynotes and posters. Co-organized by NVIDIA, Microsoft Research, MIT, and UCSD. Come and join us!

Explore VILA-U: multi-modal token in, multi-modal token out, a single autoregressive next-token prediction model for both image/video generation and understanding. VILA-U is open sourced: github.com/mit-han-lab/vi…

Very proud to announce VILA-HD, an ultracheap method for VLMs to crack high resolution tasks! 20x cheaper than current tiling based methods. Surpass GPT-4o, Gemini-1.5, and Qwen2 for high resolution benchmarks. Scale to 8Kx8K resolution. #CVPR 2025. Check below for repository.

Thanks AK for the nice intro!

🔥 Vision encoder upgrade: RADIOv2.5 = DFN_CLIP + DINOv2 + SAM + SigLIP + ToMe + multi-res training + teacher loss balancing + smart augmentations, CVPR2025. Current foundation models have too many limitations: i) tailored for a single task, ii) not flexible on resolution (like

#RSS2025 NaVILA constitutes a successful attempt for VILA to drive real world robotic dogs and humanoid! Fully deployable. Money saving. Fast inference. Check out our project page: navila-bot.github.io Many more amazing things to come!

New efficient Hybrid LLMs from @NVIDIA: Nemotron-H! Introducing a family of models combining Mamba-2, Self-Attention & FFNs for 8B, 47B and 56B sizes. • 3x faster and 1.5x smaller 47B model is on par with Qwen-72B and Llama-70B • 1.8x faster Hybrid 8B than transformers

Nvidia just dropped CLIMB on Hugging Face CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training

Sadly I will not be attending ICLR in Singapore. We have been researching on VILA across the entire system, model, and application stack. Cost-saving. Agile. Capable, yet deployable. Talk to our colleagues this week at ICLR, and the upcoming CVPR, RSS, MLSys!

Check out the great competition from NVIDIA! Great prizes ahead!