The development of large language models (LLMs) has been instrumental in advancing state-of-the-art natural language processing applications. Training LLMs with billions of parameters and trillions of tokens requires sophisticated distributed systems that enable composing and comparing several state-of-the-art techniques in order to efficiently scale across thousands of accelerators. However, existing solutions are complex, scattered across multiple libraries/repositories, lack interoperability, and are cumbersome to maintain. Thus, curating and empirically comparing training recipes requires non-trivial engineering effort. This paper introduces **TORCHTITAN**$^1$, a PyTorch-native distributed training system that unifies and advances state-of-the-art techniques, streamlining integration and reducing engineering overhead. TORCHTITAN enables seamless application of 4D parallelism in a modular and composable manner, while featuring elastic scaling to adapt to changing computational requirements. The system provides comprehensive logging, efficient checkpointing, and debugging tools, ensuring production-ready training. Moreover, TORCHTITAN incorporates innovative hardware-software co-designed solutions, leveraging cutting-edge features like Float8 training and SymmetricMemory to maximize hardware utilization. As a flexible experimental test bed, TORCHTITAN facilitates the curation and comparison of custom recipes for diverse training contexts. By leveraging TORCHTITAN, we developed optimized training recipes for the Llama 3.1 family and provide actionable guidance on selecting and combining distributed training techniques to maximize training efficiency, based on our hands-on experiences. We thoroughly assess TORCHTITAN on the Llama 3.1 family of LLMs, spanning 8 billion to 405 billion parameters, and showcase its exceptional performance, modular composability, and elastic scalability. By stacking training optimizations, we demonstrate accelerations ranging from 65.08% on Llama 3.1 8B at 128 GPU scale (1D), 12.59% on Llama 3.1 70B at 256 GPU scale (2D), to 30% on Llama 3.1 405B at 512 GPU scale (3D) on NVIDIA H100 GPUs over optimized baselines. We also demonstrate the effectiveness of 4D parallelism in enabling long context training. $^1$ GitHub: [https://github.com/pytorch/torchtitan](https://github.com/pytorch/torchtitan)

Wireless connectivity will soon enable people to consume augmented and virtual reality content anywhere and cloud-connected mobile robots to perform complex tasks collaboratively. This connectivity will come to enterprises, factory floors and digital homes first, powered by indoor networks that offer much higher data rates, greater reliability, and lower latency than today’s networks. In addition to providing traditional communication capabilities, the future indoor network will have a “sixth sense” that enables it to provide sensory information and insights to meet physiological needs such as lighting, heating, health and safety. It will serve as the core infrastructure for smart buildings and help enterprises operate more efficiently by further enabling capabilities such as immersive virtual workplaces, indoor navigation and asset tracking. This paper reviews current technologies, examines the enabling technologies of the future indoor network and presents our latest research results and our vision for implementing them in commercial and residential environments.

Channel state information (CSI)-based fingerprinting via neural networks (NNs) is a promising approach to enable accurate indoor and outdoor positioning of user equipment (UE), even under challenging propagation conditions. In this paper, we propose a positioning pipeline for wireless LAN MIMO-OFDM systems which uses uplink CSI measurements obtained from one or more unsynchronized access points (APs). For each AP receiver, novel features are first extracted from the CSI that are robust to system impairments arising in real-world transceivers. These features are the inputs to a NN that extracts a probability map indicating the likelihood of a UE being at a given grid point. The NN output is then fused across multiple APs to provide a final position estimate. We provide experimental results with real-world indoor measurements under line-of-sight (LoS) and non-LoS propagation conditions for an 80MHz bandwidth IEEE 802.11ac system using a two-antenna transmit UE and two AP receivers each with four antennas. Our approach is shown to achieve centimeter-level median distance error, an order of magnitude improvement over a conventional baseline.

In this paper, we study probabilistic time-of-arrival (ToA) and angle-of-arrival (AoA) joint localization in real indoor environments. To mitigate the effects of multipath propagation, the joint localization algorithm incorporates into the likelihood function Gaussian mixture models (GMM) and the Von Mises-Fisher distribution to model time bias errors and angular uncertainty, respectively. We evaluate the algorithm performance using a proprietary prototype deployed in an indoor factory environment with infrastructure receivers in each of the four corners at the ceiling of a 10 meter by 20 meter section. The field test results show that our joint probabilistic localization algorithm significantly outperforms baselines using only ToA or AoA measurements and achieves 2-D sub-meter accuracy at the 90%-ile. We also numerically demonstrate that the joint localization algorithm is more robust to synchronization errors than the baseline using ToA measurements only.

We executed an indoor measurement campaign to statistically characterize the bias in time of arrival (ToA) estimates incurred by multipath in the 5.5 GHz industrial, scientific and medical band by transmitting 10 and 100 MHz bandwidth waveforms. Measurements in three buildings with different propagation conditions are performed to analyze ToA statistics in a wide range of indoor environments, from partitioned offices to an open cafeteria area. We demonstrate that WINNER II and similar channel models designed for communications are not suitable for localization applications as they do not provide absolute delays. Furthermore, such models assume constant root mean square delay spread over distance, which is a significant contributor to the ToA errors for lower signal bandwidths. Other physical aspects such as specular components in line-of-sight (LOS) and random bias in non-LOS scenario are also not considered in those channel models. An extension of the WINNER II model that overcomes those flaws and provides more realistic ranging performance is proposed. We validate the extended model by demonstrating a good match between the statistics of its simulated output and our measurements.

We propose a system for enabling auxiliary communication channels in which a node transmits a millimeter (mm) wave signal which is reflected off a deliberately vibrating surface of a second node and then received by the first node. Data sequences can be encoded in the modulation of the surface, and radar sensing techniques can be used to demodulate the reflected signal. Hence our system enables not only conventional sensing in terms of range, velocity, and orientation estimation but also allows for information to be conveyed by the sensed device. We introduce the design of a metasurface driven by an energy-efficient programmable piezo-electric actuator, detail suitable radar processing, and characterize the link performance of the kinetically induced channel at distances up to five meters. As this metasurface could be used for both mobile devices and infrastructure devices, we describe opportunities for enabling novel capabilities including secure device authentication and extended-range wireless sensing across multiple devices.

The ability of nodes in a wireless network to be aware of their position continues to become more important for future applications and use cases like the Industrial Internet of Things. Positioning using cellular networks continues to draw attention as a means to realize those future applications. New Radio (NR) provides many opportunities to use cellular networks to incorporate positioning solutions into those networks. In this paper the RAT-dependent solutions expected in NR Rel-16 are discussed in detail. We highlight OTDOA and describe the major roadblocks remaining for achieving higher accuracy positioning. UTDOA is also discussed. Novel solutions to some roadblocks are presented including system level proposals to time synchronization between base stations and the uplink hearability problem. System level simulations are presented which demonstrate the high potential for both regulatory and commercial uses cases of NR positioning. We compare OTDOA with UTDOA and the other positioning solutions that are likely to be standardized for NR Rel-16. We show that OTDOA can meet the FCC E911 requirements and may also meet some commercial use case requirements.

In this letter, we take a new approach for time of arrival geo-localization. We show that the main sources of error in metropolitan areas are due to environmental imperfections that bias our solutions, and that we can rely on a probabilistic model to learn and compensate for them. The resulting localization error is validated using measurements from a live LTE cellular network to be less than 10 meters, representing an order-of-magnitude improvement.

Future 5G cellular networks supporting delay-sensitive, low-latency communications could employ random access communication to reduce the overhead compared to scheduled access techniques used in 4G networks. We consider a wireless communication system where multiple devices transmit payloads of a given fixed size in a random access fashion over shared radio resources to a common receiver. We allow retransmissions and assume Chase combining at the receiver. The radio resources are partitioned in the time and frequency dimensions, and we determine the optimal partition granularity to maximize throughput, subject to given constraints on latency and outage. In the regime of high and low signal-to-noise ratio (SNR), we derive explicit expressions for the granularity and throughput, first using a Shannon capacity approximation and then using finite block length analysis. Numerical results show that the throughput scaling results are applicable over a range of SNRs. The proposed analytical framework can provide insights for resource allocation strategies in reliable and delay-sensitive random access systems and in specific 5G use cases for massive, short packet uplink access.

We perform a novel indoor channel measurement campaign to characterize the channel impairments for time-of-arrival (ToA) localization. We use precisely calibrated and time-synchronized RF equipment for transmitting and receiving 10 MHz and 100 MHz bandwidth waveforms over the 5.5 GHz DFS band in line-of-sight (LOS) and non-LOS environments. We collect ToA bias statistics with respect to a direct LOS path, where the bias is caused by multipath and NLOS reflections. Indoor wall reflections create significant ToA bias error of up to a couple hundred nanoseconds, corresponding to a positive ranging error of several tens of meters. We show how the WINNER II channel model, commonly used for communications simulations, is not suitable for ToA localization due to the improperly modeled signal timing and unrealistically low bias. The bias statistics based on our measurements are used for simulations of a ToA localization system using conventional and probabilistic multilateration algorithms. We present numerical results which provide insights for system design and performance-cost tradeoffs with regard to infrastructure density and signal bandwidth.

We consider a single cell wireless uplink in which randomly arriving devices transmit their payload to a receiver. Given SNR per user, payload size per device, a fixed latency constraint T, total available bandwidth W, i.e., total symbol resources is given by N = TW. The total bandwidth W is evenly partitioned into B bins. Each time slot of duration T is split into a maximum number of retransmission attempts M. Hence, the N resources are partitioned into N/MB resources each bin per retransmission. We characterize the maximum average rate or number of Poisson arrivals that can successfully complete the random access procedure such that the probability of outage is sufficiently small. We analyze the proposed setting for i) noise-limited regime and ii) interference-limited regime. We show that in the noise-limited regime the devices share the resources, and in the interference-limited regime, the resources split such that devices do not experience any interference. We then incorporate Rayleigh fading to model the channel power gain distribution. Although the variability of the channel causes a drop in the number of arrivals that can successfully complete the random access phase, similar scaling results extend to the Rayleigh fading case.

BACKGROUND: Despite the tremendous drop in the cost of nucleotide sequencing in recent years, many research projects still utilize sequencing of pools containing multiple samples for the detection of sequence variants as a cost saving measure. Various software tools exist to analyze these pooled sequence data, yet little has been reported on the relative accuracy and ease of use of these different programs. RESULTS: In this manuscript we evaluate five different variant detection programs-The Genome Analysis Toolkit (GATK), CRISP, LoFreq, VarScan, and SNVer-with regard to their ability to detect variants in synthetically pooled Illumina sequencing data, by creating simulated pooled binary alignment/map (BAM) files using single-sample sequencing data from varying numbers of previously characterized samples at varying depths of coverage per sample. We report the overall runtimes and memory usage of each program, as well as each program's sensitivity and specificity to detect known true variants. CONCLUSIONS: GATK, CRISP, and LoFreq all gave balanced accuracy of 80% or greater for datasets with varying per-sample depth of coverage and numbers of samples per pool. VarScan and SNVer generally had balanced accuracy lower than 80%. CRISP and LoFreq required up to four times less computational time and up to ten times less physical memory than GATK did, and without filtering, gave results with the highest sensitivity. VarScan and SNVer had generally lower false positive rates, but also significantly lower sensitivity than the other three programs.

Metagenomic assemblers inherit all the difficulties of traditional single genome assembly, but with the additional complexity of trying to resolve assemblies of closely related species with drastically varying abundances. Assessing and comparing the quality of single genome assembly still relies on the availability of independently determined standards, such as manually curated genomic sequences. These standards are often not possible in metagenomic projects, where a large portion of the organisms and strains are novel. Thus, we must rely on de novo methods for assessing and comparing assembly qualities. Here we describe an extension to our de novo LAP framework that takes into account abundances of assembled sequences to accurately and efficiently evaluate metagenomic assemblies. We have integrated our LAP framework into the metagenomic analysis pipeline MetAMOS, allowing any user to reproduce quality assembly evaluations with relative ease.

The paper sheds light on the beamforming (BF) performance of large (potentially unconstrained in size) as well as dense (but physically constrained in size) antenna arrays when equipped with arbitrarily many elements. Two operational modes are investigated: Single-layer BF and multi-layer BF. In the first mode, a realistic BF criterion namely the average BF gain is revisited and employed to understand the far-field and the near-field effects on the BF performance of large-scale antennas above a clutter. The diminishing throughput returns in a single-layer BF mode versus the number of antennas necessitate multi-layering. In the multi-layer BF mode, the RF coverage is divided into a number of directive non-overlapping sector-beams in a deterministic manner within a multi-user multi-input multi-output (MIMO) system. The optimal number of layers that maximizes the user's sum-rate given a constrained antenna array is found as a compromise between the multiplexing gain (associated with the number of sector-beams) and the inter-beam interference, represented by the side lobe level (SLL).

Past efforts to develop measurement-based models of the UWB propagation channel have focused on residential, office, and industrial environments. However, with its confined volume and cylindrical structure, the geometry of the passenger cabin of a jet aircraft is fundamentally different from those environments considered previously. Here, we characterize large-scale aspects of UWB propagation within the passenger cabin of a typical mid-size airliner. Our measurement database consists of hundreds of frequency responses over the range 3.0 - 10.6 GHz that we collected aboard a Boeing 737-200 aircraft. The data were collected in a point-to-multipoint configuration in which a biconical UWB transmitting antenna was mounted at one of three locations near the cabin ceiling and an identical receiving antenna was mounted at headrest, armrest, and footrest level at over 50 locations throughout the cabin. We have accounted for the effects of human presence by collecting this data with the cabin empty, with passengers occupying half of the seats, and with passengers occupying virtually all of the seats. Our initial data reduction efforts have focused on the manner in which human presence and/or receiving antenna mounting location affects five large scale aspects of UWB propagation, i.e., those that affect coverage and reliability: (1) the distance dependence of path loss, (2) the frequency dependence of path loss, (3) the ratio of the energy in the line-of-sight component to the scattered components of the channel impulse response, (4) the RMS delay spread, and (5) the locations and distribution of the poles of the corresponding autoregressive frequency domain model.

Frequency selective fading may affect the orthogonality of the spreading codes in the multi-carrier direct sequence code division multiple access (MC-DS-CDMA) systems. In this paper, we define a new performance metric called the multiple access interference (MAI) coefficient for the MC-DS-CDMA system to quantitatively predict the impact of inter-code interference with the time- and frequency-domain spreading in a frequency selective fading channel. With the help of MAI coefficient, a novel interference avoidance code assignment strategy is proposed. By jointly considering the incurred MAI effect and the blocking probability in the code tree structure, the proposed interference avoidance code assignment method can effectively reduce the MAI for the multi-rate MC-DS-CDMA system, while maintaining very good call blocking rate performance.

Multi-carrier direct sequence code division multiple access (MC-DS-CDMA) systems can support multi-rate services with a great deal of flexibility by dynamically changing time-and frequency-domain spreading gains. However, in a frequency-selective fading channel, the multi-rate users using the time spreading codes with an ancestor-to-child relation, even with orthogonal frequency spreading codes, may still be affected by different quantities of multiple access interference (MAI). How to design an effective radio resource management strategy for the MC-DS-CDMA systems subject to different levels of MAIs becomes a new and important issue. In this paper, based on the error rate analysis in the physical layer, we observe and define a new performance metric, called the MAI coefficient, which can quantitatively predict the MAI effect imposed on a particular code in a multi-rate MC-DS-CDMA system. Furthermore, from the radio resource management perspective, we develop an interference avoidance code assignment scheme based on the MAI coefficient. By simulations and analysis, we demonstrate that the proposed interference avoidance code assignment method can effectively reduce the MAI in a multi-rate MC-DS-CDMA system, while maintaining very good call blocking rate performance.

Recently, the capacity region of a multiple-input multiple-output (MIMO) Gaussian broadcast channel, with Gaussian codebooks and known-interference cancellation through dirty paper coding, was shown to equal the union of the capacity regions of a collection of MIMO multiple-access channels. We use this duality result to evaluate the system capacity achievable in a cellular wireless network with multiple antennas at the base station and multiple antennas at each terminal. Some fundamental properties of the rate region are exhibited and algorithms for determining the optimal weighted rate sum and the optimal covariance matrices for achieving a given rate vector on the boundary of the rate region are presented. These algorithms are then used in a simulation study to determine potential capacity enhancements to a cellular system through known-interference cancellation. We study both the circuit data scenario in which each user requires a constant data rate in every frame and the packet data scenario in which users can be assigned a variable rate in each frame so as to maximize the long-term average throughput. In the case of circuit data, the outage probability as a function of the number of active users served at a given rate is determined through simulations. For the packet data case, long-term average throughputs that can be achieved using the proportionally fair scheduling algorithm are determined. We generalize the zero-forcing beamforming technique to the multiple receive antennas case and use this as the baseline for the packet data throughput evaluation.