arXiv ScienceSearch

arXiv subjects

Uriya Pumerantz

Publications and source records attributed to Uriya Pumerantz.

2 recordsLinked to original sources

Jamba-1.5: Hybrid Transformer-Mamba Models at Scale

We present Jamba-1.5, new instruction-tuned large language models based on our Jamba architecture. Jamba is a hybrid Transformer-Mamba mixture of experts architecture, providing high throughput and low memory usage across context lengths, while retaining the same or better quality as Transformer models. We release two model sizes: Jamba-1.5-Large, with 94B active parameters, and Jamba-1.5-Mini, with 12B active parameters. Both models are fine-tuned for a variety of conversational and instruction-following capabilties, and have an effective context length of 256K tokens, the largest amongst open-weight models. To support cost-effective inference, we introduce ExpertsInt8, a novel quantization technique that allows fitting Jamba-1.5-Large on a machine with 8 80GB GPUs when processing 256K-token contexts without loss of quality. When evaluated on a battery of academic and chatbot benchmarks, Jamba-1.5 models achieve excellent results while providing high throughput and outperforming other open-weight models on long-context benchmarks. The model weights for both sizes are publicly available under the Jamba Open Model License and we release ExpertsInt8 as open source.

cs.CL

Random walks, word metric and orbits distribution on the plane

Given a countably infinite group $G$ acting on some space $X$, an increasing family of finite subsets $G_n$ and $x\in X$, a natural question to ask is what asymptotical distribution the sets $G_nx$ form. More formally, we define for a function $f$ over $X$ the sums $S_n(f,x)=\sum_{g\in G_n}f(gx)$ and ask whether exists a function $\Psi(n):\mathbb{N}\to\mathbb{R}$ such that the sequence $\Psi(n)S_n(f,x)$ converges. This is a delicate problem that was studied under various settings. We first show a full solution when elements are chosen using a carefully chosen word metric from a specific lattice in $SL(2,\mathbb{Z})$ acting on the circle. In addition, it is proven that the resulting measure is stationary with respect to a certain random walk and has a tight connection to a well studied function from the field of Diophantine approximations. We then proceed to study the asymptotic distribution problem when elements are chosen using a random walk over $SL(2,\mathbb{R})$ acting on $\mathbb{R}^2$. We offer a variant of our initial problem which yields some surprising and interesting results.

math.DS