arXiv subjects

Sameer Kumar

Publications and source records attributed to Sameer Kumar.

At least 19 recordsLinked to original sources

arXiv2026

Effect of Weak Non-Conservative Dynamics on Pattern Formation in Scalar Active Matter

Biological systems such as bacteria and cells undergo growth or degradation, resulting in weak violations of mass conservation. We investigate how such weak non-conservative dynamics affect phase separation in scalar active matter by incorporating a reaction term into a minimal continuum model. Through numerical simulations and linear stability analysis, we show that even weak non-conservative reactions arrest coarsening and stabilize nonequilibrium microphase-separated states. With increasing activity, the system undergoes a morphological transition from interconnected labyrinthine patterns to worm-like structures and eventually to isolated droplets. Quantitative analysis of the correlation function and static structure factor reveals a well-defined steady-state characteristic length. Qualitative analysis of the resulting phases shows that the non-conservative reaction primarily promotes microphase separation and enhances local hexagonal ordering, while activity predominantly controls the domain morphology. Our results demonstrate that weak violations of mass conservation fundamentally alter the nonlinear coarsening dynamics of active phase separation and provide a minimal framework for understanding pattern formation in related systems.

cond-mat.soft↗

arXiv2026

HCCL: Collective Communication for Meta Training and Inference Accelerators

We present HCCL, a collective communication library co-designed with Meta's MTIA 300 accelerator, the first Meta chip to integrate backend networking directly on chip package. MTIA 300 includes dedicated message engines (MEs) with near-memory compute (NMC) that fully offload collective execution from the compute grid, enabling large overlap between computation and communication. HCCL uses a compiled communication model in which the host generates a complete description of each collective including dependencies. We describe the control and data path architecture, topology-aware algorithm selection across MTIA 300's asymmetric scale-up and scale-out network, and optimizations for both training and inference workloads. For training, HCCL achieves up to 940 GB/s on intra-rack collectives while introducing less than 0.5% degradation to concurrent compute throughput. For inference, we leverage one-sided communication primitives that bypass the scheduling path to minimize collective latency and describe collective designs that improve compute-communication pipelining for latency-sensitive workloads.

cs.NI↗

arXiv2026

Defect Localization by Vanishing Deviatoric Stress in Active Nematics

Collective stress generation in cellular monolayers is a key phenomenological process governing coordinated migration and emergent multicellular dynamics. We employ a generic active nematics model to investigate stress generation and its associated properties. By analyzing the maximal principal stress and its correlation with the nematic director across different activity strengths, we find that the principal stress aligns perpendicular (parallel) to the nematic director for extensile (contractile) activity. In the turbulent regime, we identify a rotation-invariant scalar measure of the in-plane deviatoric stress whose zero-level contour coincides with the locations of all $\pm 1/2$ topological defects (both nematic and principal stress defects) are localized. This feature is robust and remains unchanged with variations in both the magnitude and nature (extensile or contractile) of activity. Our findings thus open up a new route to probe the spatial alignment from the mechanical and rheological properties of confluent cell layers, where stress measurements are more accessible than detailed cell shape or size characterisation.

cond-mat.soft↗

arXiv2026

Sensitivity of External Magnetic Field on the Change in Cross-section of a Toroidal Current

Due to any toroidal current column, the magnetic field is found to be sensitive as well as insensitive to its cross-sectional area depending on location of subject point, as predicted by numerical approaches [S. Aich, J. Thakkar, and J. Ghosh, Plasma Fusion Res. 17, 2403055 (2022)], and hence the presence of an angle of invariance is found to be present for any toroidal geometry. Present study aims to validate those numerical observations using the measured magnetic field due to Aditya Upgrade tokamak plasma.

physics.plasm-ph↗

arXiv2025

Hydrodynamic bend instability of motile particles on a substrate

The emergence of hydrodynamic bend instabilities in ordered suspensions of active particles is widely observed across diverse living and synthetic systems, and is considered to be governed by dipolar active stresses generated by the self-propelled particles. Here, using linear stability analyses and numerical simulations, we show that a hydrodynamic bend instability can emerge in the absence of any dipolar active stress and solely due to the self-propulsion force acting on polar active units suspended in an incompressible fluid confined to a substrate. Specifically, we show analytically, and confirm in simulations, that a uniformly ordered state develops bend instability above a critical self-propulsion force. Numerical simulations show that a further increase in the self-propulsion strength leads the system towards a disorderly flow state. The results offer a new route for development of hydrodynamic instabilities in two-dimensional self-propelled materials that are in contact with a substrate, with wide implications in layers of orientationally ordered cells and synthetic active particles.

cond-mat.soft↗

arXiv2023

Proof of the bounded conformal conjecture

Given any asymptotically flat 3-manifold $(M,g)$ with smooth, non-empty, compact boundary $\Sigma$, the conformal conjecture states that for every $\delta>0$, there exists a metric $g' = u^4 g$, with $u$ a harmonic function, such that the area of outermost minimal area enclosure $\tilde{\Sigma}_{g'}$ of $\Sigma$ with respect to $g'$ is less than $\delta$. Recently, the conjecture was used to prove the Riemannian Penrose inequality for black holes with zero horizon area, and was proven to be true under the assumption of existence of only a finite number of minimal area enclosures of boundary $\Sigma$, and boundedness of harmonic function $u$. We prove the conjecture assuming only the boundedness of $u$.

math.DG↗

arXiv2023

On 10 dimensional Exceptional Drinfel'd Algebras

Based on Mubarakzyanov's classification of four-dimensional real Lie algebras, we classify ten-dimensional Exceptional Drinfeld algebras (EDA). The classification is restricted to EDA's whose maximal isotropic (geometric) subalgebras cannot be represented as a product of a 3D Lie algebra and a 1D abelian factor. We collect the obtained algebras into families depending on the dualities found between them. Despite algebras related by a generalized Yang-Baxter deformation we find two algebras related by a different Nambu-Lie U-duality transformation. We show that this duality relates two Type IIA backgrounds.

hep-th↗

arXiv2022

ST-MoE: Designing Stable and Transferable Sparse Expert Models

Scale has opened new frontiers in natural language processing -- but at a high cost. In response, Mixture-of-Experts (MoE) and Switch Transformers have been proposed as an energy efficient path to even larger and more capable language models. But advancing the state-of-the-art across a broad set of natural language tasks has been hindered by training instabilities and uncertain quality during fine-tuning. Our work focuses on these issues and acts as a design guide. We conclude by scaling a sparse model to 269B parameters, with a computational cost comparable to a 32B dense encoder-decoder Transformer (Stable and Transferable Mixture-of-Experts or ST-MoE-32B). For the first time, a sparse model achieves state-of-the-art performance in transfer learning, across a diverse set of tasks including reasoning (SuperGLUE, ARC Easy, ARC Challenge), summarization (XSum, CNN-DM), closed book question answering (WebQA, Natural Questions), and adversarially constructed tasks (Winogrande, ANLI R3).

cs.CL↗

arXiv2022

Active nematic gel with quenched disorder

With quenched disorder, we introduce two-dimensional active nematics suspended in an incompressible fluid. We write the coarse-grained hydrodynamic equations of motion for slow variables, viz. density, orientation and flow fields. The quenched disorder is introduced such that it interacts with the local orientation at every point with some strength. Disorder strength is tuned from zero to large values. We numerically study the defect dynamics and system's kinetics and find that the finite disorder slows the ordering. The presence of fluid induces large fluctuation in the orientation field, further disturbing the ordering. The large fluctuation in the orientation field due to the fluid is so dominant that it reduces the effect of the quenched disorder. We have also found that the disorder's effect is almost the same for both the contractile and extensile nature of active stresses in the system. This study can help {\color{black} to} understand the impact of quenched {\color{black} disorder} on the ordering kinetics of active gels with nematic interaction among the constituent objects.

cond-mat.soft↗

arXiv2021

Bond disorder enhances the information transfer in polar flock

Collection of self-propelled particles (SPPs) exhibit coherent motion and show true long-range order in two-dimensions. Inhomogeneity, in general destroys the usual long-range order of the polar SPPs. We model a system of polar self-propelled particles with inhomogeneous interaction strength or bond disorder. The system is studied near the order-to-disorder transition for different strengths of the disorder. The nature of phase transition changes from discontinuous to continuous type by tuning the strength of the disorder. The bond disorder also enhances the ordering near the transition due to the formation of a homogeneous flock state for the large disorder. It leads to faster information transfer in the system and enhances the system information entropy. Our study gives a new understanding of the effect of intrinsic inhomogeneity in the self-propelled particle system.

cond-mat.soft↗

arXiv2021

Effect of polydispersity on the dynamics of active Brownian particles

We numerically study the dynamics and the phases of self-propelled disk-shaped particles of different sizes with soft repulsive potential in two dimensions. Size diversity is introduced by the polydispersity index (PDI) $\epsilon$, which is the width of the uniform distribution of the particle's radius. The self-propulsion speed of the particles controls the activity $v$. We observe enhanced dynamics for large size diversity among the particles. We calculate the effective diffusion coefficient $D_{eff}$ in the steady-state. The system exhibits four distinct phases, jammed phase with small $D_{eff}$ for small activity and liquid phase with enhanced $D_{eff}$ for large activity. The number fluctuation is larger and smaller than the equilibrium limit in the liquid and jammed phase, respectively. Further, the jammed phase is of two types: solid-jammed and liquid jammed for small and large PDI. Whereas the liquid phase is called motility induced phase separation (MIPS)-liquid for small PDI and for large PDI, we find enhanced diffusivity and call it the {\em pure liquid} phase. The system is studied for three packing densities $\phi$, and the response of the system for polydispersity is the same for all $\phi$'s. Our study can help understand the behavior of cells of various sizes in a tissue, artificial self-driven granular particles, or living organisms of different sizes in a dense environment.

cond-mat.soft↗

arXiv2020

Exploring the limits of Concurrency in ML Training on Google TPUs

Recent results in language understanding using neural networks have required training hardware of unprecedentedscale, with thousands of chips cooperating on a single training run. This paper presents techniques to scaleML models on the Google TPU Multipod, a mesh with 4096 TPU-v3 chips. We discuss model parallelism toovercome scaling limitations from the fixed batch size in data parallelism, communication/collective optimizations,distributed evaluation of training metrics, and host input processing scaling optimizations. These techniques aredemonstrated in both the TensorFlow and JAX programming frameworks. We also present performance resultsfrom the recent Google submission to the MLPerf-v0.7 benchmark contest, achieving record training times from16 to 28 seconds in four MLPerf models on the Google TPU-v3 Multipod machine.

cs.LG↗

arXiv2020

Highly Available Data Parallel ML training on Mesh Networks

Data parallel ML models can take several days or weeks to train on several accelerators. The long duration of training relies on the cluster of resources to be available for the job to keep running for the entire duration. On a mesh network this is challenging because failures will create holes in the mesh. Packets must be routed around the failed chips for full connectivity. In this paper, we present techniques to route gradient summation allreduce traffic around failed chips on 2-D meshes. We evaluate performance of our fault tolerant allreduce techniques via the MLPerf-v0.7 ResNet-50 and BERT benchmarks. Performance results show minimal impact to training throughput on 512 and 1024 TPU-v3 chips.

cs.LG↗

arXiv2020

Training EfficientNets at Supercomputer Scale: 83% ImageNet Top-1 Accuracy in One Hour

EfficientNets are a family of state-of-the-art image classification models based on efficiently scaled convolutional neural networks. Currently, EfficientNets can take on the order of days to train; for example, training an EfficientNet-B0 model takes 23 hours on a Cloud TPU v2-8 node. In this paper, we explore techniques to scale up the training of EfficientNets on TPU-v3 Pods with 2048 cores, motivated by speedups that can be achieved when training at such scales. We discuss optimizations required to scale training to a batch size of 65536 on 1024 TPU-v3 cores, such as selecting large batch optimizers and learning rate schedules as well as utilizing distributed evaluation and batch normalization techniques. Additionally, we present timing and performance benchmarks for EfficientNet models trained on the ImageNet dataset in order to analyze the behavior of EfficientNets at scale. With our optimizations, we are able to train EfficientNet on ImageNet to an accuracy of 83% in 1 hour and 4 minutes.

cs.LG↗

arXiv2020

Active nematics with quenched disorder

We introduce a two-dimensional active nematic with quenched disorder. We write the coarse-grained hydrodynamic equations of motion for slow variables, viz. density, and orientation. Disorder strength is tuned from zero to large values. Results from the numerical solution of equations of motion as well as the calculation of two-point orientation correlation function using linear approximation show that the ordered steady-state follows a disorder-dependent crossover from quasi long-range order (QLRO) to short-range order (SRO). Such crossover is due to the pinning of +1/2 and -1/2 topological defects in the presence of finite disorder, which breaks the system in uncorrelated domains. Finite disorder slows the dynamics of +1/2 defect, and it leads to slower growth dynamics. The two-point correlation functions for the density and orientation fields show good dynamic scaling but no static scaling for the different disorder strengths. Our findings can motivate experimentalists to verify the results and find applications in living and artificial apolar systems in the presence of a quenched disorder.

cond-mat.soft↗

arXiv2020

Dynamics of a particle moving in a two dimensional Lorentz lattice gas

We study the dynamics of a particle moving in a square two-dimensional Lorentz lattice-gas. The underlying lattice-gas is occupied by two kinds of rotators, "right-rotator (R)" and "left-rotator (L)" and some of the sites are empty {\it{viz.}} vacancy "V".The density of $R$ and $L$ are the same and density of $V$ is one of the key parameters of our model. The rotators deterministically rotate the direction of a particle's velocity to the right or left and vacancies leave it unchanged. We characterise the dynamics of particle motion for different densities of vacancies. Since the system is deterministic, the particle forms a closed trajectory asymptotically. The probability of the particle being in a closed or open trajectory at time $t$ is a function of the density of vacancies. \textcolor{black}{The motion of the particle is {\it{uniform}} throughout in a fully occupied lattice. However, it is divided in two distinct phases in partially vacant lattices}: The first phase of the motion, which is the focus of this study, is characterised by anomalous diffusion and a power-law decay of the probability of being in an open trajectory. The second phase of the motion is characterised by subdiffusive motion and an exponential decay of the probability of being in an open trajectory. For lattices with a non-zero density of vacancies, the first phase of motion lasts for a longer period of time as the density of vacancies increases.

cond-mat.soft↗

arXiv2019

Scale MLPerf-0.6 models on Google TPU-v3 Pods

The recent submission of Google TPU-v3 Pods to the industry wide MLPerf v0.6 training benchmark demonstrates the scalability of a suite of industry relevant ML models. MLPerf defines a suite of models, datasets and rules to follow when benchmarking to ensure results are comparable across hardware, frameworks and companies. Using this suite of models, we discuss the optimizations and techniques including choice of optimizer, spatial partitioning and weight update sharding necessary to scale to 1024 TPU chips. Furthermore, we identify properties of models that make scaling them challenging, such as limited data parallelism and unscaled weights. These optimizations contribute to record performance in transformer, Resnet-50 and SSD in the Google MLPerf-0.6 submission.

cs.LG↗

arXiv2019

Dynamics of particle moving in one dimensional Lorentz lattice gas

We study the dynamics of a particle moving in one-dimensional Lorentz lattice-gas where particle performs mainly three different kinds of motion {\it viz} ballistic motion, diffusion and confinement. There are two different types of scatterers, {\it viz} reflector and transmitters, randomly placed in the lattice. Reflectors are such that they reverse the particle's velocity direction and transmitters let it pass through. Scatterers also change their character with flipping probability $1-\alpha$, once the particle interacts with a scatterer. Hence the system is defined by two sets of parameters, $r$, which is the initial density of reflector/transmitter and $\alpha$. For $\alpha=0$ and $\alpha=1$ dynamics of the particle is purely deterministic else it is probabilistic. In the pure deterministic case dynamics of the particle is either propagation in one direction or confined between two near-by reflectors present. For the probabilistic case $\alpha \ne 1$ and $\ne 0$, although the dynamics of particle shows anomalous diffusion where dynamics is faster, slower and comparable to normal diffusion on the variation of system parameters $(\alpha, r)$, but the asymptotic behaviour of the particle is normal diffusion. We plot the phase diagram for the asymptotic behaviour, in the plane of $\alpha$ and $r$.

cond-mat.soft↗