arXiv subjects

Maryam Mehri Dehnavi

Publications and source records attributed to Maryam Mehri Dehnavi.

At least 19 recordsLinked to original sources

arXiv2026

PATCH: Learnable Tile-level Hybrid Sparsity for LLMs

Large language models (LLMs) deliver impressive performance but incur prohibitive memory and compute costs at deployment. Model pruning is an effective way to reduce these overheads, yet existing approaches face challenges: unstructured sparsity, where nonzeros can appear anywhere, preserves accuracy but yields irregular access patterns that prevent GPU acceleration, while semi-structured 2:4 sparsity is hardware-friendly but enforces a rigid 50% pattern that degrades model quality. To bridge this gap, we introduce PATCH, a hybrid sparsity framework that enables a continuous sparsity ratio between 0% and 50%. PATCH partitions weight matrices into tiles, assigning each tile to be either dense or 2:4 sparse via a learnable mask selection mechanism. This design provides fine-grained control over accuracy-acceleration tradeoffs and supports non-uniform sparsity across layers, leading to superior overall quality. Across models from 0.5B to 13B parameters, PATCH consistently narrows the gap to dense accuracy while delivering practical speedups. For instance, on LLaMA-2 7B with an A6000 GPU, PATCH achieves 1.18x-1.38x end-to-end speedup over dense baselines while improving accuracy by 0.37%-2.96% compared to the state-of-the-art 2:4 pruning method, MaskLLM.

cs.LG↗

arXiv2026

Fast Sparse Matrix Permutation for Mesh-Based Direct Solvers

We present a fast sparse matrix permutation algorithm tailored to linear systems arising from triangle meshes. Our approach produces nested-dissection-style permutations while significantly reducing permutation runtime overhead. Rather than enforcing strict balance and separator optimality, the algorithm deliberately relaxes these design decisions to favor fast partitioning and efficient elimination-tree construction. Our method decomposes permutation into patch-level local orderings and a compact quotient-graph ordering of separators, preserving the essential structure required by sparse Cholesky factorization while avoiding its most expensive components. We integrate our algorithm into vendor-maintained sparse Cholesky solvers on both CPUs and GPUs. Across a range of graphics applications, including single factorizations and repeated factorizations, our method reduces permutation time and improves the sparse Cholesky solve performance by up to 6.27x. Our code is available at https://github.com/BehroozZare/fast-permute.

cs.GR↗

arXiv2025

OPTIMA: Optimal One-shot Pruning for LLMs via Quadratic Programming Reconstruction

Post-training model pruning is a promising solution, yet it faces a trade-off: simple heuristics that zero weights are fast but degrade accuracy, while principled joint optimization methods recover accuracy but are computationally infeasible at modern scale. One-shot methods such as SparseGPT offer a practical trade-off in optimality by applying efficient, approximate heuristic weight updates. To close this gap, we introduce OPTIMA, a practical one-shot post-training pruning method that balances accuracy and scalability. OPTIMA casts layer-wise weight reconstruction after mask selection as independent, row-wise Quadratic Programs (QPs) that share a common layer Hessian. Solving these QPs yields the per-row globally optimal update with respect to the reconstruction objective given the estimated Hessian. The shared-Hessian structure makes the problem highly amenable to batching on accelerators. We implement an accelerator-friendly QP solver that accumulates one Hessian per layer and solves many small QPs in parallel, enabling one-shot post-training pruning at scale on a single accelerator without fine-tuning. OPTIMA integrates with existing mask selectors and consistently improves zero-shot performance across multiple LLM families and sparsity regimes, yielding up to 3.97% absolute accuracy improvement. On an NVIDIA H100, OPTIMA prunes a 8B-parameter transformer end-to-end in 40 hours with 60GB peak memory. Together, these results set a new state-of-the-art accuracy-efficiency trade-offs for one-shot post-training pruning.

cs.LG↗

arXiv2025

Optimal Software Pipelining and Warp Specialization for Tensor Core GPUs

GPU architectures have continued to grow in complexity, with recent incarnations introducing increasingly powerful fixed-function units for matrix multiplication and data movement to accompany highly parallel general-purpose cores. To fully leverage these machines, software must use sophisticated schedules that maximally utilize all hardware resources. Since realizing such schedules is complex, both programmers and compilers routinely employ program transformations, such as software pipelining (SWP) and warp specialization (WS), to do so in practice. However, determining how best to use SWP and WS in combination is a challenging problem that is currently handled through a mix of brittle compilation heuristics and fallible human intuition, with little insight into the space of solutions. To remedy this situation, we introduce a novel formulation of SWP and WS as a joint optimization problem that can be solved holistically by off-the-shelf constraint solvers. We reify our approach in Twill, the first system that automatically derives optimal SWP and WS schedules for a large class of iterative programs. Twill is heuristic-free, easily extensible to new GPU architectures, and guaranteed to produce optimal schedules. We show that Twill can rediscover, and thereby prove optimal, the SWP and WS schedules manually developed by experts for Flash Attention on both the NVIDIA Hopper and Blackwell GPU architectures.

cs.PL↗

arXiv2025

SLiM: One-shot Quantization and Sparsity with Low-rank Approximation for LLM Weight Compression

Conventional model compression techniques for LLMs address high memory consumption and slow inference challenges but typically require computationally expensive retraining to preserve accuracy. In contrast, one-shot compression methods eliminate retraining cost, but struggle to achieve accuracy comparable to dense models. This paper presents SLIM, a new one-shot compression framework that holistically integrates hardware-friendly quantization, sparsity, and low-rank approximation into a unified process. First, we formulate the quantization process using a probabilistic approach (SLIM-Quant) that enables us to apply uniform quantization. Then, we use an existing one-shot pruning method to apply semi-structured sparsity on top of the quantized weights. Finally, to compensate for the introduced aggregated quantization and sparsity error, we use a novel saliency function with unique invertible and additive features that enables us to mathematically compute the value of low-rank adapters. SLIM improves model accuracy by up to 5.66% (LLaMA-2-7B) for 2:4 sparsity with 4-bit weight quantization, outperforming prior methods. Models compressed with SLIM achieve up to 4.3x and 3.8x on Nvidia RTX3060 and A100 GPUs, respectively. Additionally, they achieve up to 0.23x end-to-end memory reduction in comparison to their dense counterparts. We also propose an optional PEFT recipe that further improves accuracy by up to 1.66% (LLaMA-2-13B) compared to SLIM without fine-tuning.

cs.LG↗

arXiv2025

Adaptive Algebraic Reuse of Reordering in Cholesky Factorization with Dynamic Sparsity Pattern

Cholesky linear solvers are a critical bottleneck in challenging applications within computer graphics and scientific computing. These applications include but are not limited to elastodynamic barrier methods such as Incremental Potential Contact (IPC), and geometric operations such as remeshing and morphology. In these contexts, the sparsity patterns of the linear systems frequently change across successive calls to the Cholesky solver, necessitating repeated symbolic analyses that dominate the overall solver runtime. To address this bottleneck, we evaluate our method on over 150,000 linear systems generated from diverse nonlinear problems with dynamic sparsity changes in Incremental Potential Contact (IPC) and patch remeshing on a wide range of triangular meshes of various sizes. Our analysis using three leading sparse Cholesky libraries, Intel MKL Pardiso, SuiteSparse CHOLMOD, and Apple Accelerate, reveals that the primary performance constraint lies in the symbolic re-ordering phase of the solver. Recognizing this, we introduce Parth, an innovative re-ordering method designed to update ordering vectors only where local connectivity changes occur adaptively. Parth employs a novel hierarchical graph decomposition algorithm to break down the dual graph of the input matrix into fine-grained subgraphs, facilitating the selective reuse of fill-reducing orderings when sparsity patterns exhibit temporal coherence. Our extensive evaluation demonstrates that Parth achieves up to a 255x and 13x speedup in fill-reducing ordering for our IPC and remeshing benchmark and a 6.85x and 10.7x acceleration in symbolic analysis. These enhancements translate to up to 2.95x and 5.89x reduction in overall solver runtime. Additionally, Parth's integration requires only three lines of code, resulting in significant computational savings without the requirement of changes to the computational stack.

math.NA↗

arXiv2025

SLoPe: Double-Pruned Sparse Plus Lazy Low-Rank Adapter Pretraining of LLMs

We propose SLoPe, a Double-Pruned Sparse Plus Lazy Low-rank Adapter Pretraining method for LLMs that improves the accuracy of sparse LLMs while accelerating their pretraining and inference and reducing their memory footprint. Sparse pretraining of LLMs reduces the accuracy of the model, to overcome this, prior work uses dense models during fine-tuning. SLoPe improves the accuracy of sparsely pretrained models by adding low-rank adapters in the final 1% iterations of pretraining without adding significant overheads to the model pretraining and inference. In addition, SLoPe uses a double-pruned backward pass formulation that prunes the transposed weight matrix using N:M sparsity structures to enable an accelerated sparse backward pass. SLoPe accelerates the training and inference of models with billions of parameters up to $1.25\times$ and $1.54\times$ respectively (OPT-33B and OPT-66B) while reducing their memory usage by up to $0.63\times$ and $0.61\times$ for training and inference respectively.

cs.LG↗

arXiv2024

A Framework for Fine-Grained Synchronization of Dependent GPU Kernels

Machine Learning (ML) models execute several parallel computations including Generalized Matrix Multiplication, Convolution, Dropout, etc. These computations are commonly executed on Graphics Processing Units (GPUs), by dividing the computation into independent processing blocks, known as tiles. Since the number of tiles are usually higher than the execution units of a GPU, tiles are executed on all execution units in one or more waves. However, the number of tiles is not always a multiple of the number of execution units. Thus, tiles executed in the final wave can under-utilize the GPU. To address this issue, we present cuSync, a framework for synchronizing dependent kernels using a user-defined fine-grained synchronization policy to improve the GPU utilization. cuSync synchronizes tiles instead of kernels, which allows executing independent tiles of dependent kernels concurrently. We also present a compiler to generate diverse fine-grained synchronization policies based on dependencies between kernels. Our experiments found that synchronizing CUDA kernels using cuSync reduces the inference times of four popular ML models: MegatronLM GPT-3 by up to 15%, LLaMA by up to 14%, ResNet-38 by up to 22%, and VGG-19 by up to 16% over several batch sizes.

cs.DC↗

arXiv2024

MKOR: Momentum-Enabled Kronecker-Factor-Based Optimizer Using Rank-1 Updates

This work proposes a Momentum-Enabled Kronecker-Factor-Based Optimizer Using Rank-1 updates, called MKOR, that improves the training time and convergence properties of deep neural networks (DNNs). Second-order techniques, while enjoying higher convergence rates vs first-order counterparts, have cubic complexity with respect to either the model size and/or the training batch size. Hence they exhibit poor scalability and performance in transformer models, e.g. large language models (LLMs), because the batch sizes in these models scale by the attention mechanism sequence length, leading to large model size and batch sizes. MKOR's complexity is quadratic with respect to the model size, alleviating the computation bottlenecks in second-order methods. Because of their high computation complexity, state-of-the-art implementations of second-order methods can only afford to update the second order information infrequently, and thus do not fully exploit the promise of better convergence from these updates. By reducing the communication complexity of the second-order updates as well as achieving a linear communication complexity, MKOR increases the frequency of second order updates. We also propose a hybrid version of MKOR (called MKOR-H) that mid-training falls backs to a first order optimizer if the second order updates no longer accelerate convergence. Our experiments show that MKOR outperforms state -of-the-art first order methods, e.g. the LAMB optimizer, and best implementations of second-order methods, i.e. KAISA/KFAC, up to 2.57x and 1.85x respectively on BERT-Large-Uncased on 64 GPUs.

cs.LG↗

arXiv2022

TENGraD: Time-Efficient Natural Gradient Descent with Exact Fisher-Block Inversion

This work proposes a time-efficient Natural Gradient Descent method, called TENGraD, with linear convergence guarantees. Computing the inverse of the neural network's Fisher information matrix is expensive in NGD because the Fisher matrix is large. Approximate NGD methods such as KFAC attempt to improve NGD's running time and practical application by reducing the Fisher matrix inversion cost with approximation. However, the approximations do not reduce the overall time significantly and lead to less accurate parameter updates and loss of curvature information. TENGraD improves the time efficiency of NGD by computing Fisher block inverses with a computationally efficient covariance factorization and reuse method. It computes the inverse of each block exactly using the Woodbury matrix identity to preserve curvature information while admitting (linear) fast convergence rates. Our experiments on image classification tasks for state-of-the-art deep neural architecture on CIFAR-10, CIFAR-100, and Fashion-MNIST show that TENGraD significantly outperforms state-of-the-art NGD methods and often stochastic gradient descent in wall-clock time.

cs.LG↗

arXiv2021

Vectorizing Sparse Matrix Codes with Dependency Driven Trace Analysis

Sparse computations frequently appear in scientific simulations and the performance of these simulations rely heavily on the optimization of the sparse codes. The compact data structures and irregular computation patterns in sparse matrix computations introduce challenges to vectorizing these codes. Available approaches primarily vectorize regular regions of computations in the sparse code. They also reorganize data and computations, at a cost, to increase the number of regular regions. In this work, we propose a novel polyhedral model, called the partially strided codelets (PSC), that enables the vectorization of computation regions with irregular data access patterns. PSCs also improve data locality in sparse computation. Our DDF inspector-executor framework efficiently mines the memory accesses in the sparse computation, using an access function differentiation approach, to find PSC codelets. It generates vectorized code for the sparse matrix multiplication kernel (SpMV), a kernel with parallel outer loops, and for kernels with carried dependence, specifically the sparse triangular solver (SpTRSV). We demonstrate the performance of the DDF-generated code on a set of 60 large and small matrices (0.05-130M nonzeros). DDF outperforms the highly specialized library MKL with an average speedup of 1.93 and 4.5X for SpMV and SpTRSV, respectively. For the same matrices, DDF outperforms the state-of-the-art inspector-executor framework Sympiler [1] for the SpTRSV kernel by up to 11X and the work by Augustine et. al [2] for the SpMV kernel by up to 12X.

cs.PL↗

arXiv2021

Composing Loop-carried Dependence with Other Loops

Sparse fusion is a compile-time loop transformation and runtime scheduling implemented as a domain-specific code generator. Sparse fusion generates efficient parallel code for the combination of two sparse matrix kernels where at least one of the kernels has loop-carried dependencies. Available implementations optimize individual sparse kernels. When optimized separately, the irregular dependence patterns of sparse kernels create synchronization overheads and load imbalance, and their irregular memory access patterns result in inefficient cache usage, which reduces parallel efficiency. Sparse fusion uses a novel inspection strategy with code transformations to generate parallel fused code for sparse kernel combinations that is optimized for data locality and load balance. Code generated by Sparse fusion outperforms the existing implementations ParSy and MKL on average 1.6X and 5.1X respectively and outperforms the LBC and DAGP coarsening strategies applied to a fused data dependence graph on average 5.1X and 7.2X respectively for various kernel combinations.

cs.PL↗

arXiv2021

Randomized Gossiping with Effective Resistance Weights: Performance Guarantees and Applications

The effective resistance between a pair of nodes in a weighted undirected graph is defined as the potential difference induced when a unit current is injected at one node and extracted from the other, treating edge weights as the conductance values of edges. The effective resistance is a key quantity of interest in many applications, e.g., solving linear systems, Markov Chains, and continuous-time averaging networks. We consider effective resistances (ER) in the context of designing randomized gossiping methods for the consensus problem, where the aim is to compute the average of node values in a distributed manner through iteratively computing weighted averages among randomly chosen neighbors. We show that employing ER weights improves the averaging time corresponding to the traditional choice of uniform weights -the amount of improvement depends on the network structure. We illustrate these results through numerical experiments. We also present an application of the ER gossiping to distributed optimization: we numerically verified that using ER gossiping within EXTRA and DPGA-W methods improves their practical performance in terms of communication efficiency.

math.OC↗

arXiv2021

L-DQN: An Asynchronous Limited-Memory Distributed Quasi-Newton Method

This work proposes a distributed algorithm for solving empirical risk minimization problems, called L-DQN, under the master/worker communication model. L-DQN is a distributed limited-memory quasi-Newton method that supports asynchronous computations among the worker nodes. Our method is efficient both in terms of storage and communication costs, i.e., in every iteration the master node and workers communicate vectors of size $O(d)$, where $d$ is the dimension of the decision variable, and the amount of memory required on each node is $O(md)$, where $m$ is an adjustable parameter. To our knowledge, this is the first distributed quasi-Newton method with provable global linear convergence guarantees in the asynchronous setting where delays between nodes are present. Numerical experiments are provided to illustrate the theory and the practical performance of our method.

math.OC↗

arXiv2020

ASYNC: A Cloud Engine with Asynchrony and History for Distributed Machine Learning

ASYNC is a framework that supports the implementation of asynchrony and history for optimization methods on distributed computing platforms. The popularity of asynchronous optimization methods has increased in distributed machine learning. However, their applicability and practical experimentation on distributed systems are limited because current bulk-processing cloud engines do not provide a robust support for asynchrony and history. With introducing three main modules and bookkeeping system-specific and application parameters, ASYNC provides practitioners with a framework to implement asynchronous machine learning methods. To demonstrate ease-of-implementation in ASYNC, the synchronous and asynchronous variants of two well-known optimization methods, stochastic gradient descent and SAGA, are demonstrated in ASYNC.

cs.DC↗

arXiv2019

MatRox: Modular approach for improving data locality in Hierarchical (Mat)rix App(Rox)imation

Hierarchical matrix approximations have gained significant traction in the machine learning and scientific community as they exploit available low-rank structures in kernel methods to compress the kernel matrix. The resulting compressed matrix, HMatrix, is used to reduce the computational complexity of operations such as HMatrix-matrix multiplications with tuneable accuracy in an evaluation phase. Existing implementations of HMatrix evaluations do not preserve locality and often lead to unbalanced parallel execution with high synchronization. Also, current solutions require the compression phase to re-execute if the kernel method or the required accuracy change. In this work, we describe MatRox, a framework that uses novel structure analysis strategies, blocking and coarsen, with code specialization and a storage format to improve locality and create load-balanced parallel tasks for HMatrix-matrix multiplications. Modularization of the matrix compression phase enables the reuse of computations when there are changes to the input accuracy and the kernel function. The MatRox-generated code for matrix-matrix multiplication is 2.98x, 1.60x, and 5.98x faster than library implementations available in GOFMM, SMASH, and STRUMPACK respectively. Additionally, the ability to reuse portions of the compression computation for changes to the accuracy leads to up to 2.64x improvement with MatRox over five changes to accuracy using GOFMM.

cs.DC↗

arXiv2019

DAve-QN: A Distributed Averaged Quasi-Newton Method with Local Superlinear Convergence Rate

In this paper, we consider distributed algorithms for solving the empirical risk minimization problem under the master/worker communication model. We develop a distributed asynchronous quasi-Newton algorithm that can achieve superlinear convergence. To our knowledge, this is the first distributed asynchronous algorithm with superlinear convergence guarantees. Our algorithm is communication-efficient in the sense that at every iteration the master node and workers communicate vectors of size $O(p)$, where $p$ is the dimension of the decision variable. The proposed method is based on a distributed asynchronous averaging scheme of decision vectors and gradients in a way to effectively capture the local Hessian information of the objective function. Our convergence theory supports asynchronous computations subject to both bounded delays and unbounded delays with a bounded time-average. Unlike in the majority of asynchronous optimization literature, we do not require choosing smaller stepsize when delays are huge. We provide numerical experiments that match our theoretical results and showcase significant improvement comparing to state-of-the-art distributed algorithms.

math.OC↗

arXiv2018

Sparse Matrix Code Dependence Analysis Simplification at Compile Time

Analyzing array-based computations to determine data dependences is useful for many applications including automatic parallelization, race detection, computation and communication overlap, verification, and shape analysis. For sparse matrix codes, array data dependence analysis is made more difficult by the use of index arrays that make it possible to store only the nonzero entries of the matrix (e.g., in A[B[i]], B is an index array). Here, dependence analysis is often stymied by such indirect array accesses due to the values of the index array not being available at compile time. Consequently, many dependences cannot be proven unsatisfiable or determined until runtime. Nonetheless, index arrays in sparse matrix codes often have properties such as monotonicity of index array elements that can be exploited to reduce the amount of runtime analysis needed. In this paper, we contribute a formulation of array data dependence analysis that includes encoding index array properties as universally quantified constraints. This makes it possible to leverage existing SMT solvers to determine whether such dependences are unsatisfiable and significantly reduces the number of dependences that require runtime analysis in a set of eight sparse matrix kernels. Another contribution is an algorithm for simplifying the remaining satisfiable data dependences by discovering equalities and/or subset relationships. These simplifications are essential to make a runtime-inspection-based approach feasible.

cs.PL↗