arXiv subjects

Avinash Vem

Publications and source records attributed to Avinash Vem.

6 recordsLinked to original sources

arXiv2026

Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

We introduce Nemotron 3 Ultra, a 550 billion total and 55 billion active parameter Mixture-of-Experts Hybrid Mamba-Attention language model. We pre-trained Nemotron 3 Ultra on 20 trillion text tokens, then extended the context length to 1M tokens, and post-trained using Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD). Nemotron 3 Ultra is our most capable model yet, employing multiple key technologies - LatentMoE, Multi Token Prediction (MTP), NVFP4 pre-training, multi-environment RLVR, MOPD, and reasoning budget control. Nemotron 3 Ultra achieves up to ~6x higher inference throughput as compared to state-of-the-art publicly available LLMs while attaining on-par accuracy. The state-of-the-art accuracy, high inference throughput, and 1M token context length make Nemotron 3 Ultra ideal for long-running autonomous agentic tasks. We open-source the base, post-trained, and quantized checkpoints, along with the training data and recipe on HuggingFace.

cs.CL↗

arXiv2023

Sionna: An Open-Source Library for Next-Generation Physical Layer Research

Sionna is a GPU-accelerated open-source library for link-level simulations based on TensorFlow. It enables the rapid prototyping of complex communication system architectures and provides native support for the integration of neural networks. Sionna implements a wide breadth of carefully tested state-of-the-art algorithms that can be used for benchmarking and end-to-end performance evaluation. This allows researchers to focus on their research, making it more impactful and reproducible, while saving time implementing components outside their area of expertise. This white paper provides a brief introduction to Sionna, explains its design principles and features, as well as future extensions, such as integrated ray tracing and custom CUDA kernels. We believe that Sionna is a valuable tool for research on next-generation communication systems, such as 6G, and we welcome contributions from our community.

cs.IT↗

arXiv2020

Sparse IDMA: A Joint Graph-Based Coding Scheme for Unsourced Random Access

This article introduces a novel communication paradigm for the unsourced, uncoordinated Gaussian multiple access problem. The major components of the envisioned framework are as follows. The encoded bits of every message are partitioned into two groups. The first portion is transmitted using a compressive sensing scheme, whereas the second set of bits is conveyed using a multi-user coding scheme. The compressive sensing portion is key in sidestepping some of the challenges posed by the unsourced aspect of the problem. The information afforded by the compressive sensing is employed to create a sparse random multi-access graph conducive to joint decoding. This construction leverages the lessons learned from traditional IDMA into creating low-complexity schemes for the unsourced setting and its inherent randomness. Under joint message-passing decoding, the proposed scheme offers comparable performance to existing low-complexity alternatives. Findings are supported by numerical simulations.

cs.IT↗

arXiv2018

A Coupled Compressive Sensing Scheme for Unsourced Multiple Access

This article introduces a novel paradigm for the unsourced multiple-access communication problem. This divide-and-conquer approach leverages recent advances in compressive sensing and forward error correction to produce a computationally efficient algorithm. Within the proposed framework, every active device first partitions its data into several sub-blocks, and subsequently adds redundancy using a systematic linear block code. Compressive sensing techniques are then employed to recover sub-blocks, and the original messages are obtained by connecting pieces together using a low-complexity tree-based algorithm. Numerical results suggest that the proposed scheme outperforms other existing practical coding schemes. Measured performance lies approximately $4.3$~dB away from the Polyanskiy achievability limit, which is obtained in the absence of complexity constraints.

cs.IT↗

arXiv2017

Sub-string/Pattern Matching in Sub-linear Time Using a Sparse Fourier Transform Approach

We consider the problem of querying a string (or, a database) of length $N$ bits to determine all the locations where a substring (query) of length $M$ appears either exactly or is within a Hamming distance of $K$ from the query. We assume that sketches of the original signal can be computed off line and stored. Using the sparse Fourier transform computation based approach introduced by Pawar and Ramchandran, we show that all such matches can be determined with high probability in sub-linear time. Specifically, if the query length $M = O(N^μ)$ and the number of matches $L=O(N^λ)$, we show that for $λ< 1-μ$ all the matching positions can be determined with a probability that approaches 1 as $N \rightarrow \infty$ for $K \leq \frac{1}{6}M$. More importantly our scheme has a worst-case computational complexity that is only $O\left(\max\{N^{1-μ}\log^2 N, N^{μ+λ}\log N \}\right)$, which means we can recover all the matching positions in {\it sub-linear} time for $λ<1-μ$. This is a substantial improvement over the best known computational complexity of $O\left(N^{1-0.359 μ} \right)$ for recovering one matching position by Andoni {\em et al.} \cite{andoni2013shift}. Further, the number of Fourier transform coefficients that need to be computed, stored and accessed, i.e., the sketching complexity of this algorithm is only $O\left(N^{1-μ}\log N\right)$. Several extensions of the main theme are also discussed.

cs.IT↗

arXiv2017

Group Testing using left-and-right-regular sparse-graph codes

We consider the problem of non-adaptive group testing of $N$ items out of which $K$ or less items are known to be defective. We propose a testing scheme based on left-and-right-regular sparse-graph codes and a simple iterative decoder. We show that for any arbitrarily small $ε>0$ our scheme requires only $m=c_εK\log \frac{c_1N}{K}$ tests to recover $(1-ε)$ fraction of the defective items with high probability (w.h.p) i.e., with probability approaching $1$ asymptotically in $N$ and $K$, where the value of constants $c_ε$ and $\ell$ are a function of the desired error floor $ε$ and constant $c_1=\frac{\ell}{c_ε}$ (observed to be approximately equal to 1 for various values of $ε$). More importantly the iterative decoding algorithm has a sub-linear computational complexity of $\mathcal{O}(K\log \frac{N}{K})$ which is known to be optimal. Also for $m=c_2 K\log K\log \frac{N}{K}$ tests our scheme recovers the \textit{whole} set of defective items w.h.p. These results are valid for both noiseless and noisy versions of the problem as long as the number of defective items scale sub-linearly with the total number of items, i.e., $K=o(N)$. The simulation results validate the theoretical results by showing a substantial improvement in the number of tests required when compared to the testing scheme based on left-regular sparse-graphs.

cs.IT↗