RT @devoto_alessio: 🏆 Our @nvidia KV Cache Compression Leaderboard is now live! . Compare state-of-the-art compression methods side-by-sid….

RT @igtmn: We've released a series of OpenReasoning-Nemotron models (1.5B, 7B, 14B and 32B) that set new SOTA on a wide range of reasoning….

More than one point is needed to draw a Pareto frontier 😅

RT @gonedarragh: AIMO-2 Winning Solution: Building State-of-the-Art Mathematical Reasoning Models with OpenMathReasoning dataset.abs: https….

* 1 min on M3 Pro without any parallelization on the 2^19 states.

Common 🇫🇷 variant where four of a kind is 40points:

Implementation is based on the excellent paper "An Optimal Strategy for Yahtzee" by @DrGlennNFA (2006): (2/2).

🎲 Did you know Yahtzee can be solved optimally in less than 100 lines of Python and under 5 minutes with 2 vCPU? I built a @Gradio app so you can try it yourself: (1/2).

RT @a_erdem4: logits-processor-zoo: New version (0.1.3) is out 🎉. You can now tune the thinking duration of R1 models and force them to sta….

RT @NVIDIAAIDev: The growing context windows of LLMs unlock new possibilities but pose significant memory challenges with linearly scaling….

RT @giffmana: This actually reproduces as of today. In 5 out of 8 generations, DeepSeekV3 claims to be ChatGPT (v4), while claiming to be D….

RT @pcuenq: Great post by nvidia (published in Hugging Face ❤️) about KVPress, a library to compress transformers KV Cache and thus reduce….

-> Repo: (⭐️ me) -> Blog post: (🔺 me).-> Space: (❤️ me). (2/2).

Thanks a lot @ariG23498 for your contributions in this release !.

Fresh news from kvpress, our open source library for KV cache compression 🔥. 1. We published a blog post with @huggingface.2. We published a Space for you to try it.3. Following feedback from the research community, we added a bunch of presses and benchmarks. Links👇(1/2)

RT @ariG23498: ⚠️1 million tokens on Llama 3-70B running in bfloat16 precision would take ~ 470 GB of memory. The KV cache alone accounts….

RT @ariG23498: Controlling text generation has always been a problem. What if I wanted my LLM to always end with a particular phrase, or wa….

New version of the notebook using the sinkhorn operator for a fully differentiable relaxation:

How do you find the permutation of words that minimize their perplexity as measured by an LLM ? In this year @kaggle Santa, I shared an approach to move to a continuous space where you can use gradient-descent using REINFORCE:

Thanks @MichaelHassid for your contribution to kvpress, we hope other researchers will follow your footsteps 🔥.