Daily Papers

As large language models continue to scale, fine-grained block-scaled low-precision formats such as NVFP4 are increasingly adopted for their substantial throughput and memory benefits. However, a single FP4 micro-format often mismatches heterogeneous block-level tensor statistics. To address this without changing the standard block-scaled MMA/GEMM execution path, we propose MixFP4, a mixed micro-format extension to NVFP4 that selects between two stored FP4 micro-formats (E2M1 and E1M2) per block. MixFP4 reuses NVFP4's scale hierarchy and encodes the format choice with zero additional metadata by repurposing the sign bit of the FP8 E4M3 block scale. By decoding both micro-formats into a unified internal E2M2 compute representation, MixFP4 avoids datapath duplication. Across representative LLM families, MixFP4 improves FP4 quantization robustness and accuracy over NVFP4 baselines with modest tensor-core overhead (3.1\% area, 1.5\% power).

Quantization has become one of the most effective methodologies to compress LLMs into smaller size. However, the existing quantization solutions still show limitations of either non-negligible accuracy drop or system inefficiency. In this paper, we make a comprehensive analysis of the general quantization principles on their effect to the triangle of accuracy, memory consumption and system efficiency. We propose MixLLM that explores the new optimization space of mixed-precision quantization between output features based on the insight that different output features matter differently in the model. MixLLM identifies the output features with high salience in the global view rather than within each single layer, effectively assigning the larger bit-width to output features that need it most to achieve good accuracy with low memory consumption. We present the sweet spot of quantization configuration of algorithm-system co-design that leads to high accuracy and system efficiency. To address the system challenge, we design the two-step dequantization to make use of the int8 Tensor Core easily and fast data type conversion to reduce dequantization overhead significantly, and present the software pipeline to overlap the memory access, dequantization and the MatMul to the best. Extensive experiments show that with only 10% more bits, the PPL increasement can be reduced from about 0.5 in SOTA to within 0.2 for Llama 3.1 70B, while on average MMLU-Pro improves by 0.93 over the SOTA of three popular models. In addition to its superior accuracy, MixLLM also achieves state-of-the-art system efficiency.

Tensor Cores have been an important unit to accelerate Fused Matrix Multiplication Accumulation (MMA) in all NVIDIA GPUs since Volta Architecture. To program Tensor Cores, users have to use either legacy wmma APIs or current mma APIs. Legacy wmma APIs are more easy-to-use but can only exploit limited features and power of Tensor Cores. Specifically, wmma APIs support fewer operand shapes and can not leverage the new sparse matrix multiplication feature of the newest Ampere Tensor Cores. However, the performance of current programming interface has not been well explored. Furthermore, the computation numeric behaviors of low-precision floating points (TF32, BF16, and FP16) supported by the newest Ampere Tensor Cores are also mysterious. In this paper, we explore the throughput and latency of current programming APIs. We also intuitively study the numeric behaviors of Tensor Cores MMA and profile the intermediate operations including multiplication, addition of inner product, and accumulation. All codes used in this work can be found in https://github.com/sunlex0717/DissectingTensorCores.

We present MiniTensor, an open source tensor operations library that focuses on minimalism, correctness, and performance. MiniTensor exposes a familiar PyTorch-like Python API while it executes performance critical code in a Rust engine. The core supports dense n dimensional tensors, broadcasting, reductions, matrix multiplication, reverse mode automatic differentiation, a compact set of neural network layers, and standard optimizers. In this paper, we describe the design of MiniTensor's architecture, including its efficient memory management, dynamic computation graph for gradients, and integration with Python via PyO3. We also compare the install footprint with PyTorch and TensorFlow to demonstrate that MiniTensor achieves a package size of only a few megabytes, several orders of magnitude smaller than mainstream frameworks, while preserving the essentials needed for research and development on CPUs. The repository can be found at https://github.com/neuralsorcerer/minitensor

The deployment of Large Language Models (LLMs) on edge devices is fundamentally constrained by the "Memory Wall" the bottleneck where data movement latency outstrips arithmetic throughput. Standard inference runtimes often incur significant overhead through high-level abstractions, dynamic dispatch, and unaligned memory access patterns. In this work, we present a novel "Virtual Tensor Core" architecture implemented in software, optimized specifically for ARM64 microarchitectures (Apple Silicon). By bypassing standard library containers in favor of direct memory mapping (mmap) and implementing hand-tuned NEON SIMD kernels, we achieve a form of "Software-Defined Direct Memory Access (DMA)." Our proposed Tensor Virtualization Layout (TVL) guarantees 100% cache line utilization for weight matrices, while our zero-copy loader eliminates initialization latency. Experimental results on a 110M parameter model demonstrate a stable throughput of >60 tokens/second on M2 hardware. While proprietary hardware accelerators (e.g., Apple AMX) can achieve higher peak throughput, our architecture provides a fully open, portable, and deterministic reference implementation for studying the memory bottleneck on general-purpose ARM silicon, meeting the 200ms psycholinguistic latency threshold without opaque dependencies.

Tensor computations, with matrix multiplication being the primary operation, serve as the fundamental basis for data analysis, physics, machine learning, and deep learning. As the scale and complexity of data continue to grow rapidly, the demand for tensor computations has also increased significantly. To meet this demand, several research institutions have started developing dedicated hardware for tensor computations. To further improve the computational performance of tensor process units, we have reexamined the issue of computation reuse that was previously overlooked in existing architectures. As a result, we propose a novel EN-T architecture that can reduce chip area and power consumption. Furthermore, our method is compatible with existing tensor processing units. We evaluated our method on prevalent microarchitectures, the results demonstrate an average improvement in area efficiency of 8.7\%, 12.2\%, and 11.0\% for tensor computing units at computational scales of 256 GOPS, 1 TOPS, and 4 TOPS, respectively. Similarly, there were energy efficiency enhancements of 13.0\%, 17.5\%, and 15.5\%.

W4A4 quantization promises full utilization of INT4 Tensor Cores, yet group dequantization overhead on CUDA Cores has driven existing systems to mixed-precision fallbacks. We present the first systematic study of how intra-SM compute balance governs this bottleneck. Through controlled benchmarks across four GPUs from Ampere and Ada architectures, we identify the Tensor Cores to CUDA Cores throughput ratio (ρ) as the primary hardware indicator: the W4A4-g128 kernel yields 2.0--2.5times speedup on RTX~3090 (ρ=16) yet degrades to 0.43--0.47times on A100 (ρ=64) in compute-bond scenarios, establishing W4A4 viability as platform-dependent rather than universally infeasible. Guided by this finding, we build APEX4, which co-designs pure INT4 GEMM kernels with ρ-aware granularity adaptation to mitigate the CUDA Cores dequantization bottleneck. APEX4 achieves perplexity within 0.63 of FP16 on LLaMA-2-70B and outperforms W4Ax Atom-g128 by 4.0\%--4.4\% in zero-shot accuracy. Deployed as a drop-in replacement in unmodified vLLM, it delivers up to 1.66times end-to-end speedup on L40S (ρ=8), and 1.78times on RTX~3090 (ρ=16), 2.09times on A40 (ρ=16), while recovering A100 (ρ=64) to 1.20--1.40times via the mixed-granularity mode. Our code is available at https://github.com/APEX4-W4A4/APEX4-W4A4.

Deploying deep learning models on various devices has become an important topic. The wave of hardware specialization brings a diverse set of acceleration primitives for multi-dimensional tensor computations. These new acceleration primitives, along with the emerging machine learning models, bring tremendous engineering challenges. In this paper, we present TensorIR, a compiler abstraction for optimizing programs with these tensor computation primitives. TensorIR generalizes the loop nest representation used in existing machine learning compilers to bring tensor computation as the first-class citizen. Finally, we build an end-to-end framework on top of our abstraction to automatically optimize deep learning models for given tensor computation primitives. Experimental results show that TensorIR compilation automatically uses the tensor computation primitives for given hardware backends and delivers performance that is competitive to state-of-art hand-optimized systems across platforms.

Modern LLM infrastructure increasingly manages tensors not only as computation data, but also as persistent states shared across distributed components. Existing systems optimize individual tensor management tasks, such as model weight loading, KV cache management, and checkpoint synchronization, by deeply integrating task-specific mechanisms with execution engines, networks, or storage backends. However, this specialization creates isolated silos that hinder the reuse and composition of tensor management strategies across evolving LLM workloads. In this paper, we identify tensor lifecycle management as a missing abstraction layer in LLM infrastructure and propose Tensor-as-a-Service (TaaS), which decouples tensor state management from computation logic. We design and build TensorCast, a distributed tensor management layer that provides first-class tensor abstractions, programmable lifecycle primitives, and a runtime that separates tensor management policies from execution mechanisms. This enables developers to write tensor management programs using TensorCast APIs while transparently leveraging distributed execution and data movement. We integrate TensorCast with vLLM and SGLang and evaluate it across diverse tensor lifecycle workloads, including model weight materialization, weight synchronization, KV cache management, and programmable request routing. Our results show that TensorCast achieves competitive performance with specialized tensor management systems while enabling new cross-component optimization policies. A programmable policy implemented with TensorCast improves median TTFT by up to 93.2% under highly concurrent multi-turn agent workloads.

Tensor accelerators have gained popularity because they provide a cheap and efficient solution for speeding up computational-expensive tasks in Deep Learning and, more recently, in other Scientific Computing applications. However, since their features are specifically designed for tensor algebra (typically dense matrix-product), it is commonly assumed that they are not suitable for applications with sparse data. To challenge this viewpoint, we discuss methods and present solutions for accelerating sparse matrix multiplication on such architectures. In particular, we present a 1-dimensional blocking algorithm with theoretical guarantees on the density, which builds dense blocks from arbitrary sparse matrices. Experimental results show that, even for unstructured and highly-sparse matrices, our block-based solution which exploits Nvidia Tensor Cores is faster than its sparse counterpart. We observed significant speed-ups of up to two orders of magnitude on real-world sparse matrices.

E(3)-equivariant neural networks have demonstrated success across a wide range of 3D modelling tasks. A fundamental operation in these networks is the tensor product, which interacts two geometric features in an equivariant manner to create new features. Due to the high computational complexity of the tensor product, significant effort has been invested to optimize the runtime of this operation. For example, Luo et al. (2024) recently proposed the Gaunt tensor product (GTP) which promises a significant speedup. In this work, we provide a careful, systematic analysis of a number of tensor product operations. In particular, we emphasize that different tensor products are not performing the same operation. The reported speedups typically come at the cost of expressivity. We introduce measures of expressivity and interactability to characterize these differences. In addition, we realized the original implementation of GTP can be greatly simplified by directly using a spherical grid at no cost in asymptotic runtime. This spherical grid approach is faster on our benchmarks and in actual training of the MACE interatomic potential by 30%. Finally, we provide the first systematic microbenchmarks of the various tensor product operations. We find that the theoretical runtime guarantees can differ wildly from empirical performance, demonstrating the need for careful application-specific benchmarking. Code is available at https://github.com/atomicarchitects/PriceofFreedom.

The scaling of Large Language Models (LLMs) drives interest in matrix-based optimizers (e.g., Shampoo, Muon, SOAP) for their convergence efficiency; yet their requirement for holistic updates conflicts with the tensor fragmentation in distributed frameworks like Megatron. Existing solutions are suboptimal: synchronous approaches suffer from computational redundancy, while layer-wise partitioning fails to reconcile this conflict without violating the geometric constraints of efficient communication primitives. To bridge this gap, we propose Canzona, a Unified, Asynchronous, and Load-Balanced framework that decouples logical optimizer assignment from physical parameter distribution. For Data Parallelism, we introduce an alpha-Balanced Static Partitioning strategy that respects atomicity while neutralizing the load imbalance. For Tensor Parallelism, we design an Asynchronous Compute pipeline utilizing Micro-Group Scheduling to batch fragmented updates and hide reconstruction overhead. Extensive evaluations on the Qwen3 model family (up to 32B parameters) on 256 GPUs demonstrate that our approach preserves the efficiency of established parallel architectures, achieving a 1.57x speedup in end-to-end iteration time and reducing optimizer step latency by 5.8x compared to the baseline.

Large language models have led to state-of-the-art accuracies across a range of tasks. However, training these models efficiently is challenging for two reasons: a) GPU memory capacity is limited, making it impossible to fit large models on even a multi-GPU server, and b) the number of compute operations required to train these models can result in unrealistically long training times. Consequently, new methods of model parallelism such as tensor and pipeline parallelism have been proposed. Unfortunately, naive usage of these methods leads to fundamental scaling issues at thousands of GPUs, e.g., due to expensive cross-node communication or devices spending significant time waiting on other devices to make progress. In this paper, we show how different types of parallelism methods (tensor, pipeline, and data parallelism) can be composed to scale to thousands of GPUs and models with trillions of parameters. We survey techniques for pipeline parallelism and propose a novel interleaved pipeline parallelism schedule that can improve throughput by 10+% with memory footprint comparable to existing approaches. We quantitatively study the trade-offs between tensor, pipeline, and data parallelism, and provide intuition as to how to configure distributed training of a large model. Our approach allows us to perform training iterations on a model with 1 trillion parameters at 502 petaFLOP/s on 3072 GPUs with achieved per-GPU throughput of 52% of theoretical peak. Our code is open sourced at https://github.com/nvidia/megatron-lm.

NVIDIA's 2:4 Sparse Tensor Cores deliver 2x throughput but demand strict 50% pruning -- a ratio that collapses LLM reasoning accuracy (Qwen3: 54% to 15%). Milder (2N-2):2N patterns (e.g., 6:8, 25% pruning) preserve accuracy yet receive no hardware support, falling back to dense execution without any benefit from sparsity. We present SlideSparse, the first system to unlock Sparse Tensor Core acceleration for the (2N-2):2N model family on commodity GPUs. Our Sliding Window Decomposition reconstructs any (2N-2):2N weight block into N-1 overlapping 2:4-compliant windows without any accuracy loss; Activation Lifting fuses the corresponding activation rearrangement into per-token quantization at near-zero cost. Integrated into vLLM, SlideSparse is evaluated across various GPUs (A100, H100, B200, RTX 4090, RTX 5080, DGX-spark), precisions (FP4, INT8, FP8, BF16, FP16), and model families (Llama, Qwen, BitNet). On compute-bound workloads, the measured speedup ratio (1.33x) approaches the theoretical upper-bound N/(N-1)=4/3 at 6:8 weight sparsity in Qwen2.5-7B, establishing (2N-2):2N as a practical path to accuracy-preserving LLM acceleration. Code available at https://github.com/bcacdwk/vllmbench.

Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate, and over 90% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90%, requiring only 3.1 GB of memory for 70B-scale models.

Dense linear layers are the dominant computational bottleneck in foundation models. Identifying more efficient alternatives to dense matrices has enormous potential for building more compute-efficient models, as exemplified by the success of convolutional networks in the image domain. In this work, we systematically explore structured matrices as replacements for dense matrices. We show that different structures often require drastically different initialization scales and learning rates, which are crucial to performance, especially as models scale. Using insights from the Maximal Update Parameterization, we determine the optimal scaling for initialization and learning rates of these unconventional layers. Finally, we measure the scaling laws of different structures to compare how quickly their performance improves with compute. We propose a novel matrix family containing Monarch matrices, the Block Tensor-Train (BTT), which we show performs better than dense matrices for the same compute on multiple tasks. On CIFAR-10/100 with augmentation, BTT achieves exponentially lower training loss than dense when training MLPs and ViTs. BTT matches dense ViT-S/32 performance on ImageNet-1k with 3.8 times less compute and is more efficient than dense for training small GPT-2 language models.

Flagship Mixture-of-Experts (MoE) models are growing fast along two axes at once: total parameter count and the number of experts. In elastic deployment scenarios, many GPUs across many nodes must become serving-ready quickly, and this growth makes weight loading a noticeable part of the latency budget. Even on NVIDIA's GB300 NVL72, today's state-of-the-art loaders leave most of that bandwidth unused. The losses are structural: (C1) weight memory is fragmented into tens of thousands of per-tensor objects, so transfers run far below link bandwidth; (C2) cross-node replication is gated by NCCL communicator setup, which costs 10-110 s before a single weight byte moves; and (C3) the existing cross-node GPU->GPU clone path is serial and scales poorly to concurrent multi-node bring-up. We present FlashBoot, a hardware-friendly, framework-workflow co-designed weight-loading subsystem built on SGLang. At its core is FabricArena, a contiguous, exportable and inter-node addressable tensor memory layout. On top of it, FlashLoad loads from CPU as a single bulk, zero-copy transfer, and FlashClone replicates a resident model from a remote GPU via a remote-mapping mechanism that removes NCCL setup. In experiments on NVL72 with DeepSeek-V4-Pro and DeepSeek-V4-Flash, FlashClone maps remote weight memory in ~10 ms (versus 10-110 s for NCCL) and sustains >=700 GB/s per clone. Against the state of the art, FlashBoot accelerates single-node weight loading by up to 50x (from 20.1 s to 0.4 s) and concurrent rack-level weight loading by >270x (from 87 s to 0.32 s). Our code will be made publicly available.

Quantization is a widely-used compression technology to reduce the overhead of serving large language models (LLMs) on terminal devices and in cloud data centers. However, prevalent quantization methods, such as 8-bit weight-activation or 4-bit weight-only quantization, achieve limited performance improvements due to poor support for low-precision (e.g., 4-bit) activation. This work, for the first time, realizes practical W4A4KV4 serving for LLMs, fully utilizing the INT4 tensor cores on modern GPUs and reducing the memory bottleneck caused by the KV cache. Specifically, we propose a novel fine-grained mixed-precision quantization algorithm (FMPQ) that compresses most activations into 4-bit with negligible accuracy loss. To support mixed-precision matrix multiplication for W4A4 and W4A8, we develop a highly optimized W4Ax kernel. Our approach introduces a novel mixed-precision data layout to facilitate access and fast dequantization for activation and weight tensors, utilizing the GPU's software pipeline to hide the overhead of data loading and conversion. Additionally, we propose fine-grained streaming multiprocessor (SM) scheduling to achieve load balance across different SMs. We integrate the optimized W4Ax kernel into our inference framework, COMET, and provide efficient management to support popular LLMs such as LLaMA-3-70B. Extensive evaluations demonstrate that, when running LLaMA family models on a single A100-80G-SMX4, COMET achieves a kernel-level speedup of 2.88times over cuBLAS and a 2.02 times throughput improvement compared to TensorRT-LLM from an end-to-end framework perspective.

General matrix-matrix multiplication (GEMM) is a cornerstone of AI computations, making tensor processing engines (TPEs) increasingly critical in GPUs and domain-specific architectures. Existing architectures primarily optimize dataflow or operand reuse strategies. However, considering the interaction between matrix multiplication and multiply-accumulators (MACs) offers greater optimization potential. This work introduces a novel hardware perspective on matrix multiplication, focusing on the bit-weight dimension of MACs. We propose a finer-grained TPE notation using matrix triple loops as an example, introducing new methods for designing and optimizing PE microarchitectures. Based on this notation and its transformations, we propose four optimization techniques that improve timing, area, and power consumption. Implementing our design in RTL using the SMIC-28nm process, we evaluate its effectiveness across four classic TPE architectures: systolic array, 3D-Cube, multiplier-adder tree, and 2D-Matrix. Our techniques achieve area efficiency improvements of 1.27x, 1.28x, 1.56x, and 1.44x, and energy efficiency gains of 1.04x, 1.56x, 1.49x, and 1.20x, respectively. Applied to a bit-slice architecture, our approach achieves a 12.10x improvement in energy efficiency and 2.85x in area efficiency compared to Laconic. Our Verilog HDL code, along with timing, area, and power reports, is available at https://github.com/wqzustc/High-Performance-Tensor-Processing-Engines

Training large-scale Mixture-of-Experts (MoE) models is bottlenecked by activation memory and expert-parallel communication, yet FP4 training remains impractical on Hopper-class GPUs without native MXFP4 or NVFP4 support. In this work, we present a training recipe that enables MXFP4 efficiency for MoE models on Hopper architectures without native 4-bit computation support. A central challenge is to integrate FP4 into an existing BF16/FP8 hybrid training pipeline without incurring costly precision round-trips (e.g., FP4 leftrightarrow BF16 leftrightarrow FP8). We address this challenge by introducing direct FP8-to-FP4 quantization and de-quantization, together with scaling-aware FP4 row-wise to column-wise conversion, enabling FP4 activations and expert-parallel communication with minimal overhead. Core MoE computations are executed in FP8, while activations and expert-parallel communication are compressed using MXFP4, achieving substantial memory and bandwidth savings without degrading convergence. At the 671B parameter scale, our method achieves end-to-end training performance comparable to strong FP8 baselines, while reducing peak activation memory by 14.8\% (11.8 GB) and improving training throughput by 12.5\%, from 1157 to 1302 tokens per GPU per second. These results show that FP4 efficiency can be practically realized for large-scale MoE training through careful software-hardware co-design, even without native FP4 Tensor Core support.

Large Language Models (LLMs) demonstrate exceptional performance but entail significant memory and computational costs, restricting their practical deployment. While existing INT4/INT8 quantization reduces these costs, they often degrade accuracy or lack optimal efficiency. INT6 quantization offers a superior trade-off between model accuracy and inference efficiency, but lacks hardware support in modern GPUs, forcing emulation via higher-precision arithmetic units that limit acceleration. In this paper, we propose FlexQ, a novel post-training INT6 quantization framework combining algorithmic innovation with system-level optimizations. FlexQ employs uniform 6-bit weight quantization across all layers, with adaptive retention of 8-bit activations in layers identified through layer-wise sensitivity analysis. To maximize hardware efficiency, we develop a specialized high-performance GPU kernel supporting matrix multiplication for W6A6 and W6A8 representations via Binary Tensor Core (BTC) equivalents, effectively bypassing the lack of native INT6 tensor cores. Evaluations on LLaMA models show FlexQ maintains near-FP16 accuracy, with perplexity increases of no more than 0.05. The proposed kernel achieves an average 1.39times speedup over ABQ-LLM on LLaMA-2-70B linear layers. End-to-end, FlexQ delivers 1.33times inference acceleration and 1.21times memory savings over SmoothQuant. Code is released at https://github.com/FlyFoxPlayer/FlexQ.

Efficient parallelism is necessary for achieving low-latency, high-throughput inference with large language models (LLMs). Tensor parallelism (TP) is the state-of-the-art method for reducing LLM response latency, however GPU communications reduces combined token throughput. On the other hand, data parallelism (DP) obtains a higher throughput yet is slow in response latency. Best of both worlds does not exist, and it is not possible to combine TP and DP because of the KV cache variance across the parallelisms. We notice Sequence Parallelism (SP - Ulysses in training) has similar properties as DP but with KV cache invariance. We adapt SP to inference, and combine it with TP to get the best of both worlds. Our solution: Shift Parallelism. Shift Parallelism dynamically switches across TP and SP, and minimizes latency in low traffic without losing throughput in high traffic. The efficient GPU communications of Shift Parallelism yields up to i) 1.51x faster response in interactive workloads and ii) 50% higher throughput in batch workloads, compared to a TP-only solution. We evaluate Shift Parallelism with real-world production traces with dynamic traffic patterns as well as synthetic benchmarking patterns across models, context sizes, and arrival rates. All results affirm the same: Shift Parallelism has a better the latency vs. throughput tradeoff than TP or DP, and hence obtains low latency without degrading throughput in dynamic workloads.

Autoregressive Transformer KV caches grow linearly with context length; sliding-window caching bounds memory but discards evicted tokens entirely, so relevant evidence outside the window becomes inaccessible. We introduce Tensor Cache, a two-level cache that pairs sliding-window softmax attention as a first-level cache (L1) with a fixed-size outer-product fast-weight memory as a second-level cache (L2) fed by KV pairs evicted from the window. Recent tokens remain in exact local attention; evicted pairs are compressed into a per-layer matrix A and read by future queries through a single matrix multiplication, exploiting the linear-attention identity q_t(k_i otimes v_i)=langle q_t,k_irangle v_i. A learned scalar gate fuses the L1 and L2 outputs, and per-head decay and write-rate parameters are trained end-to-end. The outer-product memory and the read identity are well-known; our contribution is their use as an L2 cache fed exclusively by sliding-window evictions, plus identifying that the common chunked-mean training shortcut A!leftarrow!λA!+!η(bar k!otimes!bar v) silently introduces C^2{-}C spurious cross-token outer products per chunk, and closing the gap with a parallel weighted-sum scan equivalent to per-token writes within float32 epsilon. Across systems scaling, controlled associative recall, long-context language modeling, and memory-capacity diagnostics, Tensor Cache improves the memory--quality frontier over bounded-state baselines.

Distributed inference of large language models (LLMs) can introduce overheads of up to 20% even over GPUs connected via high-speed interconnects such as NVLINK. Multiple techniques have been proposed to mitigate these overheads by decomposing computations into finer-grained tasks and overlapping communication with sub-tasks as they complete. However, fine-grained decomposition of a large computation into many smaller computations on GPUs results in overheads. Further, the communication itself uses many streaming multiprocessors (SMs), adding to the overhead. We present TokenWeave to address these challenges. TokenWeave proposes a Token-Splitting technique that divides the tokens in the inference batch into two approximately equal subsets in a wave-aware manner. The computation of one subset is then overlapped with the communication of the other. In addition, TokenWeave optimizes the order of the layer normalization computation with respect to communication operations and implements a novel fused AllReduce-RMSNorm kernel carefully leveraging Multimem instruction support available on NVIDIA Hopper GPUs. These optimizations allow TokenWeave to perform communication and RMSNorm using only 2-8 SMs. Moreover, our kernel enables the memory bound RMSNorm to be overlapped with the other batch's computation, providing additional gains. Our evaluations demonstrate up to 29% latency gains and up to 26% throughput gains across multiple models and workloads. In several settings, TokenWeave results in better performance compared to an equivalent model with all communication removed.

Optimizing Large Language Model (LLM) inference in production systems is increasingly difficult due to dynamic workloads, stringent latency/throughput targets, and a rapidly expanding configuration space. This complexity spans not only distributed parallelism strategies (tensor/pipeline/expert) but also intricate framework-specific runtime parameters such as those concerning the enablement of CUDA graphs, available KV-cache memory fractions, and maximum token capacity, which drastically impact performance. The diversity of modern inference frameworks (e.g., TRT-LLM, vLLM, SGLang), each employing distinct kernels and execution policies, makes manual tuning both framework-specific and computationally prohibitive. We present AIConfigurator, a unified performance-modeling system that enables rapid, framework-agnostic inference configuration search without requiring GPU-based profiling. AIConfigurator combines (1) a methodology that decomposes inference into analytically modelable primitives - GEMM, attention, communication, and memory operations while capturing framework-specific scheduling dynamics; (2) a calibrated kernel-level performance database for these primitives across a wide range of hardware platforms and popular open-weights models (GPT-OSS, Qwen, DeepSeek, LLama, Mistral); and (3) an abstraction layer that automatically resolves optimal launch parameters for the target backend, seamlessly integrating into production-grade orchestration systems. Evaluation on production LLM serving workloads demonstrates that AIConfigurator identifies superior serving configurations that improve performance by up to 40% for dense models (e.g., Qwen3-32B) and 50% for MoE architectures (e.g., DeepSeek-V3), while completing searches within 30 seconds on average. Enabling the rapid exploration of vast design spaces - from cluster topology down to engine specific flags.

As LLMs scale to multi-million-token KV histories, real-time autoregressive decoding under tight Token-to-Token Latency (TTL) constraints faces growing pressure. Two core bottlenecks dominate: accessing Feed-Forward Network (FFN) weights and reading long KV caches. While Tensor Parallelism (TP) helps mitigate the cost of FFN weight reads, it does not scale well for attention. When TP width exceeds the number of KV heads, it leads to inefficient KV duplication, limits parallelism, and constrains batch size. Simultaneously, DRAM reads for long KV histories scale linearly with batch size, further capping efficiency. We introduce Helix Parallelism, a hybrid execution strategy that applies KV parallelism during attention to shard KV caches across GPUs, then reuses the same GPUs for TP in dense LLMs or TPxExpert Parallel (EP) in MoEs during FFN computation. To preserve exact attention behavior, Helix includes a lightweight communication step. To minimize the exposed communication cost, we introduce Helix HOP-B. Helix HOP-B effectively minimizes communication overhead through batchwise overlap, preserving low TTL while improving GPU efficiency. Compared to conventional parallelism approaches, Helix reduces TTL by up to 1.5x at fixed batch sizes and supports up to 32x larger batches under the same latency budget for DeepSeek-R1, pushing forward the throughput-latency Pareto on Blackwell and making real-time inference with ultra-long-sequence practical.

Six-bit quantization (FP6) can effectively reduce the size of large language models (LLMs) and preserve the model quality consistently across varied applications. However, existing systems do not provide Tensor Core support for FP6 quantization and struggle to achieve practical performance improvements during LLM inference. It is challenging to support FP6 quantization on GPUs due to (1) unfriendly memory access of model weights with irregular bit-width and (2) high runtime overhead of weight de-quantization. To address these problems, we propose TC-FPx, the first full-stack GPU kernel design scheme with unified Tensor Core support of float-point weights for various quantization bit-width. We integrate TC-FPx kernel into an existing inference system, providing new end-to-end support (called FP6-LLM) for quantized LLM inference, where better trade-offs between inference cost and model quality are achieved. Experiments show that FP6-LLM enables the inference of LLaMA-70b using only a single GPU, achieving 1.69x-2.65x higher normalized inference throughput than the FP16 baseline. The source code will be publicly available soon.

Attention, as a core layer of the ubiquitous Transformer architecture, is the bottleneck for large language models and long-context applications. While FlashAttention-3 optimized attention for Hopper GPUs through asynchronous execution and warp specialization, it primarily targets the H100 architecture. The AI industry has rapidly transitioned to deploying Blackwell-based systems such as the B200 and GB200, which exhibit fundamentally different performance characteristics due to asymmetric hardware scaling: tensor core throughput doubles while other functional units (shared memory bandwidth, exponential units) scale more slowly or remain unchanged. We develop several techniques to address these shifting bottlenecks on Blackwell GPUs: (1) redesigned pipelines that exploit fully asynchronous MMA operations and larger tile sizes, (2) software-emulated exponential and conditional softmax rescaling that reduces non-matmul operations, and (3) leveraging tensor memory and the 2-CTA MMA mode to reduce shared memory traffic and atomic adds in the backward pass. We demonstrate that our method, FlashAttention-4, achieves up to 1.3times speedup over cuDNN 9.13 and 2.7times over Triton on B200 GPUs with BF16, reaching up to 1613 TFLOPs/s (71% utilization). Beyond algorithmic innovations, we implement FlashAttention-4 entirely in CuTe-DSL embedded in Python, achieving 20-30times faster compile times compared to traditional C++ template-based approaches while maintaining full expressivity.

With the fast growth of parameter size, it becomes increasingly challenging to deploy large generative models as they typically require large GPU memory consumption and massive computation. Unstructured model pruning has been a common approach to reduce both GPU memory footprint and the overall computation while retaining good model accuracy. However, the existing solutions do not provide a highly-efficient support for handling unstructured sparsity on modern GPUs, especially on the highly-structured Tensor Core hardware. Therefore, we propose Flash-LLM for enabling low-cost and highly-efficient large generative model inference with the sophisticated support of unstructured sparsity on high-performance but highly restrictive Tensor Cores. Based on our key observation that the main bottleneck of generative model inference is the several skinny matrix multiplications for which Tensor Cores would be significantly under-utilized due to low computational intensity, we propose a general Load-as-Sparse and Compute-as-Dense methodology for unstructured sparse matrix multiplication. The basic insight is to address the significant memory bandwidth bottleneck while tolerating redundant computations that are not critical for end-to-end performance on Tensor Cores. Based on this, we design an effective software framework for Tensor Core based unstructured SpMM, leveraging on-chip resources for efficient sparse data extraction and computation/memory-access overlapping. At SpMM kernel level, Flash-LLM significantly outperforms the state-of-the-art library, i.e., Sputnik and SparTA by an average of 2.9x and 1.5x, respectively. At end-to-end framework level on OPT-30B/66B/175B models, for tokens per GPU-second, Flash-LLM achieves up to 3.8x and 3.6x improvement over DeepSpeed and FasterTransformer, respectively, with significantly lower inference cost.

Multi-Head Latent Attention (MLA), introduced in DeepSeek-V2, compresses key-value states into a low-rank latent vector, caching only this vector to reduce memory. In tensor parallelism (TP), however, attention heads are computed across multiple devices, and each device must load the full cache, eroding the advantage of MLA over Grouped Query Attention (GQA). We propose Tensor-Parallel Latent Attention (TPLA): a scheme that partitions both the latent representation and each head's input dimension across devices, performs attention independently per shard, and then combines results with an all-reduce. TPLA preserves the benefits of a compressed KV cache while unlocking TP efficiency. Unlike Grouped Latent Attention (GLA), every head in TPLA still leverages the full latent representation, maintaining stronger representational capacity. TPLA is drop-in compatible with models pre-trained using MLA: it supports MLA-style prefilling and enables efficient tensor-parallel decoding without retraining. Applying simple orthogonal transforms -- e.g., the Hadamard transform or PCA -- before TP slicing further mitigates cross-shard interference, yielding minimal accuracy degradation. By reducing the per-device KV cache for DeepSeek-V3 and Kimi-K2, we achieve 1.79x and 1.93x speedups, respectively, at a 32K-token context length while maintaining performance on commonsense and LongBench benchmarks. TPLA can be implemented with FlashAttention-3, enabling practical end-to-end acceleration.

The key-value (KV) cache has become the dominant memory cost of transformer inference: it grows with batch size, context length, and depth, and at long context it, rather than the model weights, sets the throughput ceiling. Existing reductions fall into two families. Low-rank methods factor two-dimensional slices of the cache, either per-head matrices or cross-layer feature blocks, and quantization methods lower the bit-width of every entry. Neither exploits the fact that the cache at a layer is naturally a third-order tensor whose three axes, the heads, the tokens, and the features, carry very different amounts of redundancy. We take this tensor view directly. Our method, JoLT (Joint Lagrangian Tucker), applies a partial Tucker decomposition that compresses only the token and feature axes while leaving the head and layer axes intact, then restores the energy that truncation discards with a rotated low-bit residual: a random orthogonal rotation followed by low-bit quantization. A single Lagrangian dual allocates the Tucker ranks and the residual bit-widths together, per layer group and separately for keys and values, under one byte budget. The result is a near-lossless 2-3x compression. Perplexity stays near-lossless on both a grouped-query-attention model (Mistral-7B-v0.3) and a multi-head-attention model (LLaMA-2-13B), and GSM8K accuracy and needle-in-a-haystack retrieval hold at the uncompressed baseline at 2x on both architectures and through 3x on the GQA model. At 2x, JoLT reconstructs the cache to relative Frobenius error 0.009 (K) and 0.006 (V) on both architectures. A randomized-SVD variant, FlashJoLT, delivers a 5-13x compression-time speedup at 1024-token context and matched quality.

Orthogonal finetuning (OFT) offers highly parameter-efficient adaptation while preventing catastrophic forgetting, but its high runtime and memory demands limit practical deployment. We identify the core computational bottleneck in OFT as its weight-centric implementation, which relies on costly matrix-matrix multiplications with cubic complexity. To overcome this, we propose OFTv2, an input-centric reformulation that instead uses matrix-vector multiplications (i.e., matrix-free computation), reducing the computational cost to quadratic. We further introduce the Cayley-Neumann parameterization, an efficient orthogonal parameterization that approximates the matrix inversion in Cayley transform via a truncated Neumann series. These modifications allow OFTv2 to achieve up to 10x faster training and 3x lower GPU memory usage without compromising performance. In addition, we extend OFTv2 to support finetuning quantized foundation models and show that it outperforms the popular QLoRA in training stability, efficiency, and memory usage.

As large language model (LLM) inference demands ever-greater resources, there is a rapid growing trend of using low-bit weights to shrink memory usage and boost inference efficiency. However, these low-bit LLMs introduce the need for mixed-precision matrix multiplication (mpGEMM), which is a crucial yet under-explored operation that involves multiplying lower-precision weights with higher-precision activations. Unfortunately, current hardware does not natively support mpGEMM, resulting in indirect and inefficient dequantization-based implementations. To address the mpGEMM requirements in low-bit LLMs, we explored the lookup table (LUT)-based approach for mpGEMM. However, a conventional LUT implementation falls short of its potential. To fully harness the power of LUT-based mpGEMM, we introduce LUT Tensor Core, a software-hardware co-design optimized for low-bit LLM inference. Specifically, we introduce software-based operator fusion and table symmetrization techniques to optimize table precompute and table storage, respectively. Then, LUT Tensor Core proposes the hardware design featuring an elongated tiling shape design to enhance table reuse and a bit-serial design to support various precision combinations in mpGEMM. Moreover, we design an end-to-end compilation stack with new instructions for LUT-based mpGEMM, enabling efficient LLM compilation and optimizations. The evaluation on low-bit LLMs (e.g., BitNet, LLAMA) shows that LUT Tensor Core achieves more than a magnitude of improvements on both compute density and energy efficiency.

Efficiently processing long sequences with Transformer models usually requires splitting the computations across accelerators via context parallelism. The dominant approaches in this family of methods, such as Ring Attention or DeepSpeed Ulysses, enable scaling over the context dimension but do not focus on memory efficiency, which limits the sequence lengths they can support. More advanced techniques, such as Fully Pipelined Distributed Transformer or activation offloading, can further extend the possible context length at the cost of training throughput. In this paper, we present UPipe, a simple yet effective context parallelism technique that performs fine-grained chunking at the attention head level. This technique significantly reduces the activation memory usage of self-attention, breaking the activation memory barrier and unlocking much longer context lengths. Our approach reduces intermediate tensor memory usage in the attention layer by as much as 87.5% for 32B Transformers, while matching previous context parallelism techniques in terms of training speed. UPipe can support the context length of 5M tokens when training Llama3-8B on a single 8timesH100 node, improving upon prior methods by over 25%.

This report evaluates the performance impact of enabling Trusted Execution Environments (TEE) on NVIDIA Hopper GPUs for large language model (LLM) inference tasks. We benchmark the overhead introduced by TEE mode across various LLMs and token lengths, with a particular focus on the bottleneck caused by CPU-GPU data transfers via PCIe. Our results indicate that while there is minimal computational overhead within the GPU, the overall performance penalty is primarily attributable to data transfer. For the majority of typical LLM queries, the overhead remains below 7%, with larger models and longer sequences experiencing nearly zero overhead.

The development of efficient machine learning models for molecular systems representation is becoming crucial in scientific research. We introduce TensorNet, an innovative O(3)-equivariant message-passing neural network architecture that leverages Cartesian tensor representations. By using Cartesian tensor atomic embeddings, feature mixing is simplified through matrix product operations. Furthermore, the cost-effective decomposition of these tensors into rotation group irreducible representations allows for the separate processing of scalars, vectors, and tensors when necessary. Compared to higher-rank spherical tensor models, TensorNet demonstrates state-of-the-art performance with significantly fewer parameters. For small molecule potential energies, this can be achieved even with a single interaction layer. As a result of all these properties, the model's computational cost is substantially decreased. Moreover, the accurate prediction of vector and tensor molecular quantities on top of potential energies and forces is possible. In summary, TensorNet's framework opens up a new space for the design of state-of-the-art equivariant models.

LLMs have traditionally scaled along dense dimensions, where performance is coupled with near-linear increases in computational cost. While MoE decouples capacity from compute, it introduces large memory overhead and hardware efficiency challenges. To overcome these, we propose token-indexed parameters as a novel, orthogonal scaling axis that decouple model capacity from FLOPs. Specifically, we introduce Joint-Token (JTok) and Mixture of Joint-Token (JTok-M), which augment Transformer layers with modulation vectors retrieved from auxiliary embedding tables. These vectors modulate the backbone via lightweight, element-wise operations, incurring negligible FLOPs overhead. Extensive experiments on both dense and MoE backbones, spanning from 650M (190M + 460M embedding) to 61B (17B + 44B embedding) total parameters, demonstrate that our approach consistently reduces validation loss and significantly improves downstream task performance (e.g., +4.1 on MMLU, +8.3 on ARC, +8.9 on CEval). Rigorous isoFLOPs analysis further confirms that JTok-M fundamentally shifts the quality-compute Pareto frontier, achieving comparable model quality with 35% less compute relative to vanilla MoE architectures, and we validate that token-indexed parameters exhibit a predictable power-law scaling behavior. Moreover, our efficient implementation ensures that the overhead introduced by JTok and JTok-M remains marginal.

Large Language Models increasingly rely on distributed techniques for their training and inference. These techniques require communication across devices which can reduce scaling efficiency as the number of devices increases. While some distributed techniques can overlap, and thus, hide this communication with independent computations, techniques such as Tensor Parallelism (TP) inherently serialize communication with model execution. One approach to hide this serialized communication is to interleave it with the producer operation (of the communicated data) in a fine-grained manner. However, this fine-grained interleaving of communication and computation in software can be difficult. Furthermore, as with any concurrent execution, it requires compute and memory resources to be shared between computation and communication, causing resource contention that reduces overlapping efficacy. To overcome these challenges, we propose T3 which applies hardware-software co-design to transparently overlap serialized communication while minimizing resource contention with compute. T3 transparently fuses producer operations with the subsequent communication via a simple configuration of the producer's output address space and requires minor software changes. At the hardware level, T3 adds a lightweight track and trigger mechanism to orchestrate the producer's compute, and communication. It further uses compute-enhanced memories for communication's attendant compute. As a result, T3 reduces resource contention, and efficiently overlaps serialized communication with computation. For important Transformer models like T-NLG, T3 speeds up communication-heavy sublayers by 30% geomean (max 47%) and reduces data movement by 22% geomean (max 36%). Furthermore, T3's benefits persist as models scale: geomean 29% for sublayers in sim500-billion parameter models, PALM and MT-NLG.

Lower numerical precision reduces storage and memory traffic but raises the perturbation floor. We study rank compensation: reinvesting saved memory in a larger approximation rank. For matrices, the singular-value error identity yields a directly testable sufficient condition requiring the additional singular component to offset the perturbation from storing the rank-augmented approximation in lower precision. On ten SuiteSparse matrices, all 100 truncation-dominated configurations (50 FP32 and 50 FP16) are certified non-increases and strict accuracy wins, with mean error ratio 0.963 and storage ratios 58.8% and 29.4% relative to the FP64 baseline. FP16 failures occur only in tail-rank stress tests near the perturbation floor. At the largest resident matrix-application batch, compensated FP32 and FP16 achieve geometric-mean A100 speedups of 1.28times and 2.12times; neither accelerates the smallest batch. For Tensor-Train (TT) approximation, we give a conditional a posteriori extension based on the measured truncation gain and rounded-core perturbation. Across three-way and six-way synthetic tests, FP32 and FP16 achieve combined accuracy-memory wins in 10 of 20 and 14 of 20 trials. On public hyperspectral tensors and FROSTT top-active subtensors, the corresponding counts are 44 of 60 and 54 of 60; four FP16 Salinas-A tail-stress cases fail. No certified TT case exceeds the FP64 error beyond numerical tolerance. Reconstruction of six public tensors yields geometric-mean compensated speedups of 1.38times (FP32) and 1.94times (FP16). Timings cover resident downstream kernels, not factorization, transfers, or end-to-end acceleration.

The efficiency of attention is important due to its quadratic time complexity. We enhance the efficiency of attention through two key contributions: First, we leverage the new FP4 Tensor Cores in Blackwell GPUs to accelerate attention computation. Our implementation achieves 1038 TOPS on RTX5090, which is a 5x speedup over the fastest FlashAttention on RTX5090. Experiments show that our FP4 attention can accelerate inference of various models in a plug-and-play way. Second, we pioneer low-bit attention to training tasks. Existing low-bit attention works like FlashAttention3 and SageAttention focus only on inference. However, the efficiency of training large models is also important. To explore whether low-bit attention can be effectively applied to training tasks, we design an accurate and efficient 8-bit attention for both forward and backward propagation. Experiments indicate that 8-bit attention achieves lossless performance in fine-tuning tasks but exhibits slower convergence in pretraining tasks. The code will be available at https://github.com/thu-ml/SageAttention.

With the rapid development of deep learning models and hardware support for dense computing, the deep learning workload characteristics changed significantly from a few hot spots on compute-intensive operations to a broad range of operations scattered across the models. Accelerating a few compute-intensive operations using the expert-tuned implementation of primitives does not fully exploit the performance potential of AI hardware. Various efforts have been made to compile a full deep neural network (DNN) graph. One of the biggest challenges is to achieve high-performance tensor compilation by generating expert level performance code for the dense compute-intensive operations and applying compilation optimization at the scope of DNN computation graph across multiple compute-intensive operations. We present oneDNN Graph Compiler, a tensor compiler that employs a hybrid approach of using techniques from both compiler optimization and expert-tuned kernels for high performance code generation of the deep neural network graph. oneDNN Graph Compiler addresses unique optimization challenges in the deep learning domain, such as low-precision computation, aggressive fusion of graph operations, optimization for static tensor shapes and memory layout, constant weight optimization, and memory buffer reuse. Experimental results demonstrate significant performance gains over existing tensor compiler and primitives library for performance-critical DNN computation graphs and end-to-end models on Intel Xeon Scalable Processors.

The increasing computational demands of foundation models have spurred research into low-precision training, with 4-bit floating-point (FP4) formats emerging as a frontier for maximizing hardware throughput. While numerous techniques have been proposed to stabilize FP4 training, they often present isolated solutions with varying, and not always clear, computational overheads. This paper aims to provide a unified view of the design space of FP4 training. We introduce a comprehensive, quantisation gradient-based framework for microscaling quantization that allows for a theoretical analysis of the computational costs associated with different stabilization methods on both the forward and backward passes. Using a simulator built on this framework, we conduct an extensive empirical study across a wide range of machine learning tasks, including regression, image classification, diffusion models, and language models. By systematically evaluating thousands of combinations of techniques, such as novel gradient approximations, rounding strategies, and scaling methods, we identify which configurations offer the most favourable performance-to-overhead trade-off. We find that the techniques enabling the best trade-off involve carefully combining Hadamard transformations, tensor scaling and stochastic rounding. We further find that using UE5M3 as a scaling factor potentially offers a good compromise between range and precision with manageable computational overhead.

Running large language models on consumer devices such as laptops and desktops is challenging because model weights often exceed GPU memory capacity, making offloading inference necessary to extend effective model capacity with CPU memory. Existing offloading systems, however, typically rely on coarse layer-level or expert-level scheduling, which overlooks substantial heterogeneity among tensors within the same layer and adapts poorly to changing hardware load conditions on such devices. This paper presents ATSInfer, a hybrid CPU-GPU inference system for consumer devices that performs offloading at tensor granularity. ATSInfer combines static tensor placement with load-aware dynamic transfer, and introduces asynchronous CPU-GPU coordination to efficiently schedule hardware storage, data movement, and computation across heterogeneous backends. We implement ATSInfer and evaluate it on representative consumer platforms using both dense and MoE models. Compared with existing systems, ATSInfer improves prefill throughput by up to 1.94times and decode throughput by up to 3.29times, while also increasing GPU utilization and making more effective use of PCIe bandwidth. These results show that ATSInfer can substantially improve the user experience of local LLM deployment on personal consumer devices.

Deploying deep learning on microcontrollers (MCUs) requires fitting both the model weights and runtime activations within tens to hundreds of kilobytes of SRAM and Flash. Existing approaches either use an interpreter with a pre-written kernel library or combine a Python code generator with hand-written C kernels. This incurs dispatch overhead and requires framework dependencies. None produces self-contained, dependency-free C code specialized to a given model. In this paper, we present te-comp, the first complete MLIR-based compiler that takes a standard model in .tflite or .safetensors format and emits all C source files needed for deployment: the inference engine (kernels, weight arrays, memory layout) and the RTOS integration layer with zero runtime dependencies. On 10 MCUNet CNN benchmarks, te-comp outperforms MCUNet TinyEngine by using 38% less SRAM (up to 5.9times) and 15% less Flash, while generating complete C code rather than linking a pre-written kernel library. To demonstrate, we use te-comp to compile an 8.3M-parameter GPT-Neo transformer (TinyStories-v2) and deploy it on a single-core Cortex-M7 at 160MHz. The full firmware uses 249KB of 512KB SRAM and 3.62MB of 4MB Flash and achieves approximately 1 token/s interactive text generation via Q4_0 weight quantization, a stateful int4 KV cache, and vocabulary pruning -- a capability no existing MCU inference framework supports.

Deep learning (DL) compilers rely on cost models and auto-tuning to optimize tensor programs for target hardware. However, existing approaches depend on large offline datasets, incurring high collection costs and offering suboptimal transferability across platforms. In this paper, we introduce TCL, a novel efficient and transferable compiler framework for fast tensor program optimization across diverse hardware platforms to address these challenges. Specifically, TCL is built on three core enablers: (1) the RDU Sampler, a data-efficient active learning strategy that selects only 10% of tensor programs by jointly optimizing Representativeness, Diversity, and Uncertainty, substantially reducing data collection costs while maintaining near-original model accuracy; (2) a new Mamba-based cost model that efficiently captures long-range schedule dependencies while achieving a favorable trade-off between prediction accuracy and computational cost through reduced parameterization and lightweight sequence modeling; and (3) a continuous knowledge distillation framework that effectively and progressively transfers knowledge across multiple hardware platforms while avoiding the parameter explosion and data dependency issues typically caused by traditional multi-task learning. Extensive experiments validate the effectiveness of each individual enabler and the holistic TCL framework. When optimizing a range of mainstream DL models on both CPU and GPU platforms, TCL achieves, on average, 16.8x and 12.48x faster tuning time, and 1.20x and 1.13x lower inference latency, respectively, compared to Tenset-MLP.

The rapid growth of large-language models (LLMs) is driving a new wave of specialized hardware for inference. This paper presents the first workload-centric, cross-architectural performance study of commercial AI accelerators, spanning GPU-based chips, hybrid packages, and wafer-scale engines. We compare memory hierarchies, compute fabrics, and on-chip interconnects, and observe up to 3.7x performance variation across architectures as batch size and sequence length change. Four scaling techniques for trillion-parameter models are examined; expert parallelism offers an 8.4x parameter-to-compute advantage but incurs 2.1x higher latency variance than tensor parallelism. These findings provide quantitative guidance for matching workloads to accelerators and reveal architectural gaps that next-generation designs must address.

On NVIDIA Blackwell B200, tensor-core throughput outpaces special-function exponential throughput by more than two orders of magnitude, exposing exponential evaluation in fused attention kernels. A pretrained Transformer, however, may not need it evaluated accurately at every element. We characterize what a pretrained model does need by approximating softmax at inference in ten frozen decoder-only models (0.5B-72B). The number of positions the softmax map assigns probability to and within-row resolution can be cut substantially, yet uniform weighting of the same positions is damaging. Where a fixed resolution budget is placed matters as much as its size, with resolution near the row maximum consistently favored. Perturbations matched on scalar distortion produce model-dependent responses of opposite sign. These findings motivate Rowmax-PoT, a coarse logarithmic weight representation anchored at each row maximum, and Rowmax-H15, its hardware specialization in FlashAttention-4. On B200, the patched FP8 attention forward is 12.4% faster at causal 8K and 25.8% faster at non-causal 8K in host-side call-latency measurements; board energy per forward falls by 8.4% at causal 16K. Measured separately on the BF16 kernel path at 2K, Rowmax-H15 increases perplexity by 0.091-0.492% across five models from three families.

Collocating deep learning training tasks improves GPU utilization but causes drastic slowdowns due to resource contention and risks Out-of-Memory (OOM) failures. Accurate memory estimation is essential for robust collocation, while GPU utilization -- a key proxy for resource contention -- enables interference-aware scheduling to reduce slowdowns and improve throughput. Existing GPU memory estimators span three paradigms -- analytical models, CPU-side libraries, and ML-based estimators -- each with distinct limitations: dependence on detailed model specifications, intrusive integration, poor generalization, and varying latency overhead. GPU heterogeneity further complicates estimation, as identical tasks can exhibit markedly different memory footprints across hardware generations. GPU utilization remains comparatively understudied, further complicated by the non-additive nature of utilization metrics and hardware sensitivity. We conduct a systematic analysis of representative estimators from each paradigm -- Horus, PyTorch FakeTensor, and our lightweight ML-based estimator -- evaluating accuracy, generalizability, and practical overhead. We construct a synthetic dataset spanning MLPs, CNNs, and Transformers with controlled architectural variations, and train MLP- and Transformer-based estimators for memory prediction. We further experiment with utilization estimation on the same dataset. Our evaluation reveals key tradeoffs and validates estimators against real-world unseen models. Significant challenges remain: analytical models are hardware-dependent, CPU-side libraries impose intrusive integration costs, and ML-based estimators struggle with cross-architecture generalization. We release all datasets, tools, and artifacts to support further research.

Large Transformer networks are increasingly used in settings where low inference latency can improve the end-user experience and enable new applications. However, autoregressive inference is resource intensive and requires parallelism for efficiency. Parallelism introduces collective communication that is both expensive and represents a phase when hardware resources are underutilized. Towards mitigating this, Kraken is an evolution of the standard Transformer architecture that is designed to complement existing tensor parallelism schemes for efficient inference on multi-device systems. By introducing a fixed degree of intra-layer model parallelism, the architecture allows collective operations to be overlapped with compute, decreasing latency and increasing hardware utilization. When trained on OpenWebText, Kraken models reach a similar perplexity as standard Transformers while also preserving their language modeling capabilities when evaluated on the SuperGLUE benchmark. Importantly, when tested on multi-GPU systems using TensorRT-LLM engines, Kraken speeds up Time To First Token by a mean of 35.6% across a range of model sizes, context lengths, and degrees of tensor parallelism.

In the realm of Large Language Model (LLM) inference, the inherent structure of transformer models coupled with the multi-GPU tensor parallelism strategy leads to a sequential execution of computation and communication. This results in substantial underutilization of computing resources during the communication phase. To mitigate this inefficiency, various techniques have been developed to optimize the use of computational power throughout the communication process. These strategies primarily involve overlapping matrix computations and communications, as well as interleaving micro-batches across different requests. Nonetheless, these approaches either fall short of achieving ideal overlap or impose certain limitations on their application. To overcome these challenges, this paper introduces a novel strategy for computation-communication overlap that operates at the sequence level. This method not only enhances the degree of overlap but also minimizes the constraints on its applicability. Experimental evaluations conducted using 30b/70b models have demonstrated significant improvements in efficiency. Specifically, the proposed technique has been shown to reduce time consumption by approximately 35% on 4090 GPU and by roughly 15% on A800 GPU during the prefill stage of LLM inference.

Tuning tensor program generation involves searching for various possible program transformation combinations for a given program on target hardware to optimize the tensor program execution. It is already a complex process because of the massive search space and exponential combinations of transformations make auto-tuning tensor program generation more challenging, especially when we have a heterogeneous target. In this research, we attempt to address these problems by learning the joint neural network and hardware features and transferring them to the new target hardware. We extensively study the existing state-of-the-art dataset, TenSet, perform comparative analysis on the test split strategies and propose methodologies to prune the dataset. We adopt an attention-inspired approach for tuning the tensor programs enabling them to embed neural network and hardware-specific features. Our approach could prune the dataset up to 45\% of the baseline without compromising the Pairwise Comparison Accuracy (PCA). Further, the proposed methodology can achieve on-par or improved mean inference time with 25%-40% of the baseline tuning time across different networks and target hardware.

Web applications are increasingly becoming the primary platform for AI service delivery, making in-browser deep learning (DL) inference more prominent. However, current in-browser inference systems fail to effectively utilize advanced web programming techniques and customize kernels for various client devices, leading to suboptimal performance. To address the issues, this paper presents the first in-browser inference system, nn-JIT.web, which enables just-in-time (JIT) auto-generation of optimized kernels for both CPUs and GPUs during inference. The system achieves this by using two novel web programming techniques that can significantly reduce kernel generation time, compared to other tensor compilers such as TVM, while maintaining or even improving performance. The first technique, Tensor-Web Compiling Co-Design, lowers compiling costs by unifying tensor and web compiling and eliminating redundant and ineffective compiling passes. The second technique, Web-Specific Lite Kernel Optimization Space Design, reduces kernel tuning costs by focusing on web programming requirements and efficient hardware resource utilization, limiting the optimization space to only dozens. nn-JIT.web is evaluated for modern transformer models on a range of client devices, including the mainstream CPUs and GPUs from ARM, Intel, AMD and Nvidia. Results show that nn-JIT.web can achieve up to 8.2x faster within 30 seconds compared to the baselines across various models.

With the increasing adoption of graph neural networks (GNNs) in the machine learning community, GPUs have become an essential tool to accelerate GNN training. However, training GNNs on very large graphs that do not fit in GPU memory is still a challenging task. Unlike conventional neural networks, mini-batching input samples in GNNs requires complicated tasks such as traversing neighboring nodes and gathering their feature values. While this process accounts for a significant portion of the training time, we find existing GNN implementations using popular deep neural network (DNN) libraries such as PyTorch are limited to a CPU-centric approach for the entire data preparation step. This "all-in-CPU" approach has negative impact on the overall GNN training performance as it over-utilizes CPU resources and hinders GPU acceleration of GNN training. To overcome such limitations, we introduce PyTorch-Direct, which enables a GPU-centric data accessing paradigm for GNN training. In PyTorch-Direct, GPUs are capable of efficiently accessing complicated data structures in host memory directly without CPU intervention. Our microbenchmark and end-to-end GNN training results show that PyTorch-Direct reduces data transfer time by 47.1% on average and speeds up GNN training by up to 1.6x. Furthermore, by reducing CPU utilization, PyTorch-Direct also saves system power by 12.4% to 17.5% during training. To minimize programmer effort, we introduce a new "unified tensor" type along with necessary changes to the PyTorch memory allocator, dispatch logic, and placement rules. As a result, users need to change at most two lines of their PyTorch GNN training code for each tensor object to take advantage of PyTorch-Direct.

As machine learning tasks continue to evolve, the trend has been to gather larger datasets and train increasingly larger models. While this has led to advancements in accuracy, it has also escalated computational costs to unsustainable levels. Addressing this, our work aims to strike a delicate balance between computational efficiency and model accuracy, a persisting challenge in the field. We introduce a novel method that employs core subset selection for reweighting, effectively optimizing both computational time and model performance. By focusing on a strategically selected coreset, our approach offers a robust representation, as it efficiently minimizes the influence of outliers. The re-calibrated weights are then mapped back to and propagated across the entire dataset. Our experimental results substantiate the effectiveness of this approach, underscoring its potential as a scalable and precise solution for model training.

Large Language Models (LLMs) based on Mixture-of-Experts (MoE) architecture are showing promising performance on various tasks. However, running them on resource-constrained settings, where GPU memory resources are not abundant, is challenging due to huge model sizes. Existing systems that offload model weights to CPU memory suffer from the significant overhead of frequently moving data between CPU and GPU. In this paper, we propose Fiddler, a resource-efficient inference engine with CPU-GPU orchestration for MoE models. The key idea of Fiddler is to use the computation ability of the CPU to minimize the data movement between the CPU and GPU. Our evaluation shows that Fiddler can run the uncompressed Mixtral-8x7B model, which exceeds 90GB in parameters, to generate over 3 tokens per second on a single GPU with 24GB memory, showing an order of magnitude improvement over existing methods. The code of Fiddler is publicly available at https://github.com/efeslab/fiddler

Autoregressive models, despite their commendable performance in a myriad of generative tasks, face challenges stemming from their inherently sequential structure. Inference on these models, by design, harnesses a temporal dependency, where the current token's probability distribution is conditioned on preceding tokens. This inherent characteristic severely impedes computational efficiency during inference as a typical inference request can require more than thousands of tokens, where generating each token requires a load of entire model weights, making the inference more memory-bound. The large overhead becomes profound in real deployment where requests arrive randomly, necessitating various generation lengths. Existing solutions, such as dynamic batching and concurrent instances, introduce significant response delays and bandwidth contention, falling short of achieving optimal latency and throughput. To address these shortcomings, we propose Flover -- a temporal fusion framework for efficiently inferring multiple requests in parallel. We deconstruct the general generation pipeline into pre-processing and token generation, and equip the framework with a dedicated work scheduler for fusing the generation process temporally across all requests. By orchestrating the token-level parallelism, Flover exhibits optimal hardware efficiency and significantly spares the system resources. By further employing a fast buffer reordering algorithm that allows memory eviction of finished tasks, it brings over 11x inference speedup on GPT and 16x on LLAMA compared to the cutting-edge solutions provided by NVIDIA FasterTransformer. Crucially, by leveraging the advanced tensor parallel technique, Flover proves efficacious across diverse computational landscapes, from single-GPU setups to distributed scenarios, thereby offering robust performance optimization that adapts to variable use cases.

We introduce GraphNet, a dataset of 2.7K real-world deep learning computational graphs with rich metadata, spanning six major task categories across multiple deep learning frameworks. To evaluate tensor compiler performance on these samples, we propose the benchmark metric Speedup Score S(t), which jointly considers runtime speedup and execution correctness under tunable tolerance levels, offering a reliable measure of general optimization capability. Furthermore, we extend S(t) to the Error-aware Speedup Score ES(t), which incorporates error information and helps compiler developers identify key performance bottlenecks. In this report, we benchmark the default tensor compilers, CINN for PaddlePaddle and TorchInductor for PyTorch, on computer vision (CV) and natural language processing (NLP) samples to demonstrate the practicality of GraphNet. The full construction pipeline with graph extraction and compiler evaluation tools is available at https://github.com/PaddlePaddle/GraphNet .

The demand for inference on extremely large scale LLMs has seen enormous growth in the recent months. It made evident the colossal shortage of dedicated hardware capable of efficient and fast processing of the involved compute and memory movement. The problem is aggravated by the exploding raise in the lengths of the sequences being processed, since those require efficient on-chip storage of the KV-cache of size proportional to the sequence length. To make the required compute feasible and fit the involved data into available memory, numerous quantization techniques have been proposed that allow accurate quantization for both weights and activations. One of the main recent breakthroughs in this direction was introduction of the family of Block Floating Point (BFP) formats characterized by a block of mantissas with a shared scale factor. These enable memory- power-, and compute- efficient hardware support of the tensor operations and provide extremely good quantization accuracy. The main issues preventing widespread application of block formats is caused by the presence of outliers in weights and activations since those affect the accuracy of the other values in the same block. In this paper, we focus on the most critical problem of limited KV-cache storage. We propose a novel approach enabling usage of low precision BFP formats without compromising the resulting model accuracy. We exploit the common channel-wise patterns exhibited by the outliers to rearrange them in such a way, that their quantization quality is significantly improved. The methodology yields 2x savings in the memory footprint without significant degradation of the model's accuracy. Importantly, the rearrangement of channels happens at the compile time and thus has no impact on the inference latency.

Large language models make weight storage and memory traffic major inference costs, motivating low-precision formats that represent each weight with only a few bits. Such formats use a scale to map floating-point values into a small codebook; NVFP4 improves local range utilization by letting every 16 E2M1 weights share an E4M3 block scale. Choosing that scale is difficult in GPTQ because quantizing one column updates those that follow, so evaluating a block independently can misestimate its final reconstruction error. Large models pose a second challenge: full-precision weights, calibration activations, and second-order state cannot all remain on one accelerator, while assigning complete layers to devices leaves each time-consuming layer solve serial. We introduce Schur Replay, a scale-selection algorithm that reproduces the GPTQ updates caused by each block scale and scores the resulting block error after accounting for compensation from unquantized columns. Separately, our execution infrastructure keeps only the active layer resident, tiers activations across device, host, and disk, retires full-precision layers after export, and distributes independent output rows across tensor-parallel ranks. Together, the algorithm and infrastructure attain 99.35% and 100.84% question-weighted recovery from BF16 across seven benchmarks on Qwen3.5-397B-A17B and Llama-3.3-70B-Instruct. On the 397B model, the infrastructure reduces measured per-layer time by 15.17times over ModelOpt and 23.14times over LLM Compressor, with lower memory used per GPU.

We present Nautilus, a novel tensor compiler that moves toward fully automated math-to-kernel optimization. Nautilus compiles a high-level algebraic specification of tensor operators into efficient tiled GPU kernels. Nautilus's successive lowering design allows high-level optimizations, expression rewrites, and tile optimizations to be jointly applied in a single end-to-end system. Nautilus presents a novel auto-scheduler that discovers sequences of high-level optimizations, while preserving the regular program structure needed by tile optimizers. Nautilus's auto-scheduler captures complex interactions and trade-offs in the high-level optimizations, including aggressive global transformations like advanced reduction fusion. Nautilus is the first end-to-end tensor compiler capable of starting from a math-like description of attention and automatically discovering FlashAttention-3-like kernels, offloading the entire burden of optimization from the programmer to the compiler. Across five transformer-based models and 150 evaluation configurations on NVIDIA GH200 and RTX 5090 GPUs, Nautilus achieves up to 23% higher throughput than state-of-the-art compilers on GH200 and up to 42% on RTX 5090, while matching or exceeding manually written cuDNN kernels on many long-sequence configurations.

Recently, foundation models based on Vision Transformers (ViTs) have become widely available. However, their fine-tuning process is highly resource-intensive, and it hinders their adoption in several edge or low-energy applications. To this end, in this paper we introduce an efficient fine-tuning method for ViTs called ALaST (Adaptive Layer Selection Fine-Tuning for Vision Transformers) to speed up the fine-tuning process while reducing computational cost, memory load, and training time. Our approach is based on the observation that not all layers are equally critical during fine-tuning, and their importance varies depending on the current mini-batch. Therefore, at each fine-tuning step, we adaptively estimate the importance of all layers and we assign what we call ``compute budgets'' accordingly. Layers that were allocated lower budgets are either trained with a reduced number of input tokens or kept frozen. Freezing a layer reduces the computational cost and memory usage by preventing updates to its weights, while discarding tokens removes redundant data, speeding up processing and reducing memory requirements. We show that this adaptive compute allocation enables a nearly-optimal schedule for distributing computational resources across layers, resulting in substantial reductions in training time (up to 1.5x), FLOPs (up to 2x), and memory load (up to 2x) compared to traditional full fine-tuning approaches. Additionally, it can be successfully combined with other parameter-efficient fine-tuning methods, such as LoRA.

The performance of deep learning models at scale relies heavily on how effectively high-level mathematical operations are mapped to underlying physical hardware. While high-level tensor frameworks provide flexible abstractions, their execution models inherently lack the whole-graph visibility required to maximize hardware utilization, often forcing a reliance on opaque, hand-written kernel libraries for complex operations like Attention. To bridge this gap, we present the next iteration of Nova, an automated end-to-end JIT compiler that achieves absolute control over hardware mapping by synthesizing fine-grained kernels directly from the computation's structure. In this work, we extend Nova's compilation pipeline to natively support full Transformer architectures. By capturing eager executions and unifying forward and backward passes into a single value-semantic dialect, Nova unlocks aggressive whole-graph optimizations. Rather than relying on rigid, pre-compiled library calls, Nova focuses on extensive cross-operator fusions, collapsing complex causal attention sub-graphs, element-wise operations, and memory-bound normalizations directly into single fused kernels to drastically reduce global memory roundtrips. In our evaluations training a full GPT-2 architecture on Ada 6000 GPUs, Nova demonstrates superior end-to-end throughput, averaging 441K tokens/second compared to 406K for our own eager execution and 405K for torch.compile. By drastically reducing memory-bound overheads through compiler-native fusion, Nova enables efficient full LLM compilation on modern hardware while strictly maintaining numerical parity.

Lossless model compression holds tremendous promise for alleviating the memory and bandwidth bottlenecks in bit-exact Large Language Model (LLM) serving. However, existing approaches often result in substantial inference slowdowns due to fundamental design mismatches with GPU architectures: at the kernel level, variable-length bitstreams produced by traditional entropy codecs break SIMT parallelism; at the system level, decoupled pipelines lead to redundant memory traffic. We present ZipServ, a lossless compression framework co-designed for efficient LLM inference. ZipServ introduces Tensor-Core-Aware Triple Bitmap Encoding (TCA-TBE), a novel fixed-length format that enables constant-time, parallel decoding, together with a fused decompression-GEMM (ZipGEMM) kernel that decompresses weights on-the-fly directly into Tensor Core registers. This "load-compressed, compute-decompressed" design eliminates intermediate buffers and maximizes compute intensity. Experiments show that ZipServ reduces the model size by up to 30%, achieves up to 2.21x kernel-level speedup over NVIDIA's cuBLAS, and expedites end-to-end inference by an average of 1.22x over vLLM. ZipServ is the first lossless compression system that provides both storage savings and substantial acceleration for LLM inference on GPUs.

Graphics processing units (GPUs) are continually evolving to cater to the computational demands of contemporary general-purpose workloads, particularly those driven by artificial intelligence (AI) utilizing deep learning techniques. A substantial body of studies have been dedicated to dissecting the microarchitectural metrics characterizing diverse GPU generations, which helps researchers understand the hardware details and leverage them to optimize the GPU programs. However, the latest Hopper GPUs present a set of novel attributes, including new tensor cores supporting FP8, DPX, and distributed shared memory. Their details still remain mysterious in terms of performance and operational characteristics. In this research, we propose an extensive benchmarking study focused on the Hopper GPU. The objective is to unveil its microarchitectural intricacies through an examination of the new instruction-set architecture (ISA) of Nvidia GPUs and the utilization of new CUDA APIs. Our approach involves two main aspects. Firstly, we conduct conventional latency and throughput comparison benchmarks across the three most recent GPU architectures, namely Hopper, Ada, and Ampere. Secondly, we delve into a comprehensive discussion and benchmarking of the latest Hopper features, encompassing the Hopper DPX dynamic programming (DP) instruction set, distributed shared memory, and the availability of FP8 tensor cores. The microbenchmarking results we present offer a deeper understanding of the novel GPU AI function units and programming features introduced by the Hopper architecture. This newfound understanding is expected to greatly facilitate software optimization and modeling efforts for GPU architectures. To the best of our knowledge, this study makes the first attempt to demystify the tensor core performance and programming instruction sets unique to Hopper GPUs.

Mixture-of-Experts (MoE) architectures allow frontier language models to scale to trillions of parameters, but their deployment is constrained by massive memory footprints and memory-bandwidth limitations. Although modern accelerators provide Sparse Tensor Cores (SpTCs) that reduce weight storage and increase throughput through low-precision semi-structured sparsity, exploiting them for MoEs remains challenging because of substantial model-quality degradation and the lack of grouped sparse GEMM primitives. We present an end-to-end hardware-software co-design framework that compresses expert weights into hardware-native, low-precision sparse representations and accelerates their execution on SpTCs. Algorithmically, our framework relaxes discrete semi-structured support selection through continuous reparameterization, enabling differentiable joint optimization with quantized weights under a router-weighted reconstruction objective and scalable expert-parallel compression. Systemically, we develop a custom grouped sparse GEMM kernel tailored to low-precision sparse MoE inference on SpTCs. Across MoE models ranging from 30 billion to one trillion parameters, our framework improves state-of-the-art joint sparse-quantization accuracy by up to 4.35 percentage points while preserving 96.09% of the original model's performance. On NVIDIA B200 GPUs, our kernel outperforms the vendor baseline by up to 1.65times, increasing serving throughput by 1.18times and reducing end-to-end latency by up to 4.03times. These results establish hardware-software co-design as a practical path toward scalable and efficient MoE deployment.

Deep neural network (DNN) inference has become an important part of many data-center workloads. This has prompted focused efforts to design ever-faster deep learning accelerators such as GPUs and TPUs. However, an end-to-end DNN-based vision application contains more than just DNN inference, including input decompression, resizing, sampling, normalization, and data transfer. In this paper, we perform a thorough evaluation of computer vision inference requests performed on a throughput-optimized serving system. We quantify the performance impact of server overheads such as data movement, preprocessing, and message brokers between two DNNs producing outputs at different rates. Our empirical analysis encompasses many computer vision tasks including image classification, segmentation, detection, depth-estimation, and more complex processing pipelines with multiple DNNs. Our results consistently demonstrate that end-to-end application performance can easily be dominated by data processing and data movement functions (up to 56% of end-to-end latency in a medium-sized image, and sim 80% impact on system throughput in a large image), even though these functions have been conventionally overlooked in deep learning system design. Our work identifies important performance bottlenecks in different application scenarios, achieves 2.25times better throughput compared to prior work, and paves the way for more holistic deep learning system design.

Efficient inference with Large Language Models (LLMs) increasingly relies on Key-Value (KV) caches to store previously computed key and value vectors at each layer. These caches are essential to minimize redundant computation during autoregressive token generation, lowering computational complexity from quadratic to linear. However, the growth of KV caches has posed significant system-level challenges, particularly as model sizes increase, context lengths grow, and concurrent requests compete for limited memory resources. Even though several recent frameworks for KV cache management have emerged, their comparative trade-offs in memory consumption and inference performance have not been fully understood, especially under varying request sizes and model configurations. In this work, we conduct an empirical study of three state-of-the-art KV cache management frameworks: vLLM, InfiniGen, and H2O. These frameworks employ techniques such as tensor offloading, token eviction heuristics, and speculative scheduling to balance memory usage and performance. We evaluate their performance in terms of a range of metrics such as latency, throughput, and memory usage across a spectrum of key parameters including request rates, model sizes, and sparsity levels. Our results pinpoint the conditions for each framework to perform the best, revealing the most suitable selection and configuration of KV cache strategies under memory and performance constraints.

Large language models (LLMs) have been widely applied but face challenges in efficient inference. While quantization methods reduce computational demands, ultra-low bit quantization with arbitrary precision is hindered by limited GPU Tensor Core support and inefficient memory management, leading to suboptimal acceleration. To address these challenges, we propose a comprehensive acceleration scheme for arbitrary precision LLMs. At its core, we introduce a novel bipolar-INT data format that facilitates parallel computing and supports symmetric quantization, effectively reducing data redundancy. Building on this, we implement an arbitrary precision matrix multiplication scheme that decomposes and recovers matrices at the bit level, enabling flexible precision while maximizing GPU Tensor Core utilization. Furthermore, we develop an efficient matrix preprocessing method that optimizes data layout for subsequent computations. Finally, we design a data recovery-oriented memory management system that strategically utilizes fast shared memory, significantly enhancing kernel execution speed and minimizing memory access latency. Experimental results demonstrate our approach's effectiveness, with up to 2.4\times speedup in matrix multiplication compared to NVIDIA's CUTLASS. When integrated into LLMs, we achieve up to 6.7\times inference acceleration. These improvements significantly enhance LLM inference efficiency, enabling broader and more responsive applications of LLMs.

Modern natural language processing models have achieved unprecedented scale, yet the tools for their evaluation often remain a computational bottleneck, limiting the pace of research. This is particularly acute for in-training evaluation metrics, such as per-sentence reward signals in Reinforcement Learning, which must operate efficiently on batches of token IDs directly on the GPU. In this paper, we introduce TensorBLEU, a novel implementation of the BLEU metric designed from the ground up for this specific use case. Our approach is fully vectorized for GPU-accelerated, per-sentence computation within PyTorch and introduces a memory-efficient counting mechanism. By creating a compact, batch-specific dictionary of n-grams using torch.unique, our method avoids the prohibitive memory costs of traditional hashing-based vectorization, making it practical for large-vocabulary models. We benchmark TensorBLEU against NLTK, the standard library for token-ID-based BLEU calculation on the CPU. Experiments show that TensorBLEU provides speedups of over 13x on consumer-grade GPUs (NVIDIA T4) and exceeding 40x on data-center-class hardware (NVIDIA A100). This performance transforms a significant bottleneck into a negligible part of the training loop. By clearly defining its role as a "Token-ID BLEU" for development purposes and open-sourcing our implementation, we provide a powerful tool for accelerating research in areas like RL-based model fine-tuning.

We rethink test-time scaling laws from a practical efficiency perspective, revealing that the effectiveness of smaller models is significantly overestimated. Prior work, grounded in compute-optimality, overlooks critical memory access bottlenecks introduced by inference-time strategies (e.g., Best-of-N, long CoTs). Our holistic analysis, spanning models from 0.6B to 32B parameters, reveals a new Kinetics Scaling Law that better guides resource allocation by incorporating both computation and memory access costs. Kinetics Scaling Law suggests that test-time compute is more effective when used on models above a threshold than smaller ones. A key reason is that in TTS, attention, rather than parameter count, emerges as the dominant cost factor. Motivated by this, we propose a new scaling paradigm centered on sparse attention, which lowers per-token cost and enables longer generations and more parallel samples within the same resource budget. Empirically, we show that sparse attention models consistently outperform dense counterparts, achieving over 60 points gains in low-cost regimes and over 5 points gains in high-cost regimes for problem-solving accuracy on AIME, encompassing evaluations on state-of-the-art MoEs. These results suggest that sparse attention is essential for realizing the full potential of test-time scaling because, unlike training, where parameter scaling saturates, test-time accuracy continues to improve through increased generation. The code is available at https://github.com/Infini-AI-Lab/Kinetics.

Large neural networks spend most computation on floating point tensor multiplications. In this work, we find that a floating point multiplier can be approximated by one integer adder with high precision. We propose the linear-complexity multiplication L-Mul algorithm that approximates floating point number multiplication with integer addition operations. The new algorithm costs significantly less computation resource than 8-bit floating point multiplication but achieves higher precision. Compared to 8-bit floating point multiplications, the proposed method achieves higher precision but consumes significantly less bit-level computation. Since multiplying floating point numbers requires substantially higher energy compared to integer addition operations, applying the L-Mul operation in tensor processing hardware can potentially reduce 95% energy cost by element-wise floating point tensor multiplications and 80% energy cost of dot products. We calculated the theoretical error expectation of L-Mul, and evaluated the algorithm on a wide range of textual, visual, and symbolic tasks, including natural language understanding, structural reasoning, mathematics, and commonsense question answering. Our numerical analysis experiments agree with the theoretical error estimation, which indicates that L-Mul with 4-bit mantissa achieves comparable precision as float8_e4m3 multiplications, and L-Mul with 3-bit mantissa outperforms float8_e5m2. Evaluation results on popular benchmarks show that directly applying L-Mul to the attention mechanism is almost lossless. We further show that replacing all floating point multiplications with 3-bit mantissa L-Mul in a transformer model achieves equivalent precision as using float8_e4m3 as accumulation precision in both fine-tuning and inference.

Large language models (LLMs) have demonstrated remarkable success as foundational models, benefiting various downstream applications through fine-tuning. Recent studies on loss scaling have demonstrated the superior performance of larger LLMs compared to their smaller counterparts. Nevertheless, training LLMs with billions of parameters poses significant challenges and requires considerable computational resources. For example, training a one trillion parameter GPT-style model on 20 trillion tokens requires a staggering 120 million exaflops of computation. This research explores efficient distributed training strategies to extract this computation from Frontier, the world's first exascale supercomputer dedicated to open science. We enable and investigate various model and data parallel training techniques, such as tensor parallelism, pipeline parallelism, and sharded data parallelism, to facilitate training a trillion-parameter model on Frontier. We empirically assess these techniques and their associated parameters to determine their impact on memory footprint, communication latency, and GPU's computational efficiency. We analyze the complex interplay among these techniques and find a strategy to combine them to achieve high throughput through hyperparameter tuning. We have identified efficient strategies for training large LLMs of varying sizes through empirical analysis and hyperparameter tuning. For 22 Billion, 175 Billion, and 1 Trillion parameters, we achieved GPU throughputs of 38.38%, 36.14%, and 31.96%, respectively. For the training of the 175 Billion parameter model and the 1 Trillion parameter model, we achieved 100% weak scaling efficiency on 1024 and 3072 MI250X GPUs, respectively. We also achieved strong scaling efficiencies of 89% and 87% for these two models.

We introduce a unified theoretical framework for the rigorous analysis and systematic construction of deep neural networks (DNNs). This framework addresses a gap in existing theory by explicitly modeling the structure of tensor operations -- lower level information that is often abstracted. Our framework enables two novel objectives: (1) analysis of the evolution of architectural complexity over deep learning history, and (2) automatic construction of novel architectures based on new types of tensor operations. Our study of DNNs introduced over the past 40 years reveals a connection between groundbreaking architectures and increases in different types of architectural complexity. Moreover, we identify several large classes of higher complexity architectures that have not yet been explored. We then collect a dataset of 3,000+ higher complexity architectures, which we publicly release at: https://github.com/combinatoriallabs/ArchitecturalComplexity.

Serving large-scale machine learning (ML) models efficiently and with low latency has become challenging owing to increasing model size and complexity. Quantizing models can simultaneously reduce memory and compute requirements, facilitating their widespread access. However, for large models not all layers are equally amenable to the same numerical precision and aggressive quantization can lead to unacceptable loss in model accuracy. One approach to prevent this accuracy degradation is mixed-precision quantization, which allows different tensors to be quantized to varying levels of numerical precision, leveraging the capabilities of modern hardware. Such mixed-precision quantiztaion can more effectively allocate numerical precision to different tensors `as needed' to preserve model accuracy while reducing footprint and compute latency. In this paper, we propose a method to efficiently determine quantization configurations of different tensors in ML models using post-training mixed precision quantization. We analyze three sensitivity metrics and evaluate them for guiding configuration search of two algorithms. We evaluate our method for computer vision and natural language processing and demonstrate latency reductions of up to 27.59% and 34.31% compared to the baseline 16-bit floating point model while guaranteeing no more than 1% accuracy degradation.

Scaling language models to handle longer input sequences typically necessitates large key-value (KV) caches, resulting in substantial memory overhead during inference. In this paper, we propose Tensor Product Attention (TPA), a novel attention mechanism that uses tensor decompositions to represent queries, keys, and values compactly, significantly shrinking KV cache size at inference time. By factorizing these representations into contextual low-rank components (contextual factorization) and seamlessly integrating with RoPE, TPA achieves improved model quality alongside memory efficiency. Based on TPA, we introduce the Tensor ProducT ATTenTion Transformer (T6), a new model architecture for sequence modeling. Through extensive empirical evaluation of language modeling tasks, we demonstrate that T6 exceeds the performance of standard Transformer baselines including MHA, MQA, GQA, and MLA across various metrics, including perplexity and a range of renowned evaluation benchmarks. Notably, TPAs memory efficiency enables the processing of significantly longer sequences under fixed resource constraints, addressing a critical scalability challenge in modern language models. The code is available at https://github.com/tensorgi/T6.

Large language models (LLMs) have shown excellent performance on various tasks, but the astronomical model size raises the hardware barrier for serving (memory size) and slows down token generation (memory bandwidth). In this paper, we propose Activation-aware Weight Quantization (AWQ), a hardware-friendly approach for LLM low-bit weight-only quantization. Our method is based on the observation that weights are not equally important: protecting only 1% of salient weights can greatly reduce quantization error. We then propose to search for the optimal per-channel scaling that protects the salient weights by observing the activation, not weights. AWQ does not rely on any backpropagation or reconstruction, so it can well preserve LLMs' generalization ability on different domains and modalities, without overfitting to the calibration set; it also does not rely on any data layout reordering, maintaining the hardware efficiency. AWQ outperforms existing work on various language modeling, common sense QA, and domain-specific benchmarks. Thanks to better generalization, it achieves excellent quantization performance for instruction-tuned LMs and, for the first time, multi-modal LMs. We also implement efficient tensor core kernels with reorder-free online dequantization to accelerate AWQ, achieving a 1.45x speedup over GPTQ and is 1.85x faster than the cuBLAS FP16 implementation. Our method provides a turn-key solution to compress LLMs to 3/4 bits for efficient deployment.

Equivariant Graph Neural Networks (EGNNs) have become a widely used approach for modeling 3D atomistic systems. However, mainstream architectures face critical scalability bottlenecks due to the explicit construction of geometric features or dense tensor products on every edge. To overcome this, we introduce E2Former-V2, a scalable architecture that integrates algebraic sparsity with hardware-aware execution. We first propose Equivariant Axis-Aligned Sparsification (EAAS). EAAS builds on Wigner-6j convolution by exploiting an SO(3) rightarrow SO(2) change of basis to transform computationally expensive dense tensor contractions into efficient, sparse parity re-indexing operations. Building on this representation, we introduce On-the-Fly Equivariant Attention, a fully node-centric mechanism implemented via a custom fused Triton kernel. By eliminating materialized edge tensors and maximizing SRAM utilization, our kernel achieves a 20times improvement in TFLOPS compared to standard implementations. Extensive experiments on the SPICE and OMol25 datasets demonstrate that E2Former-V2 maintains comparable predictive performance while notably accelerating inference. This work demonstrates that large equivariant transformers can be trained efficiently using widely accessible GPU platforms. The code is avalible at https://github.com/IQuestLab/UBio-MolFM/tree/e2formerv2.

Since Large Language Models or LLMs have demonstrated high-quality performance on many complex language tasks, there is a great interest in bringing these LLMs to mobile devices for faster responses and better privacy protection. However, the size of LLMs (i.e., billions of parameters) requires highly effective compression to fit into storage-limited devices. Among many compression techniques, weight-clustering, a form of non-linear quantization, is one of the leading candidates for LLM compression, and supported by modern smartphones. Yet, its training overhead is prohibitively significant for LLM fine-tuning. Especially, Differentiable KMeans Clustering, or DKM, has shown the state-of-the-art trade-off between compression ratio and accuracy regression, but its large memory complexity makes it nearly impossible to apply to train-time LLM compression. In this paper, we propose a memory-efficient DKM implementation, eDKM powered by novel techniques to reduce the memory footprint of DKM by orders of magnitudes. For a given tensor to be saved on CPU for the backward pass of DKM, we compressed the tensor by applying uniquification and sharding after checking if there is no duplicated tensor previously copied to CPU. Our experimental results demonstrate that \prjname can fine-tune and compress a pretrained LLaMA 7B model from 12.6 GB to 2.5 GB (3bit/weight) with the Alpaca dataset by reducing the train-time memory footprint of a decoder layer by 130times, while delivering good accuracy on broader LLM benchmarks (i.e., 77.7% for PIQA, 66.1% for Winograde, and so on).

Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale. We introduce MiniMax Sparse Attention (MSA), a blockwise sparse attention built upon Grouped Query Attention (GQA). A lightweight Index Branch scores key-value blocks and independently selects a Top-k subset for each GQA group, enabling group-specific sparse retrieval while maintaining efficient block-level execution; the Main Branch then performs exact block-sparse attention over only the selected blocks. Designed around a principle of simplicity and scalability, MSA is deliberately streamlined, making it straightforward to deploy efficiently across a broad range of GPUs. To translate sparsity into practical speedups, we co-design MSA with a GPU execution path that uses exp-free Top-k selection and KV-outer sparse attention to improve tensor-core utilization under block-granular access. On a 109B-parameter model with native multimodal training, MSA performs on par with GQA while reducing per-token attention compute by 28.4x at 1M context. Paired with our co-designed kernel, MSA achieves 14.2x prefill and 7.6x decoding wall-clock speedups on H800. Our inference kernel is available at: https://github.com/MiniMax-AI/MSA. A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3.

High-performance sparse matrix-matrix (SpMM) multiplication is paramount for science and industry, as the ever-increasing sizes of data prohibit using dense data structures. Yet, existing hardware, such as Tensor Cores (TC), is ill-suited for SpMM, as it imposes strict constraints on data structures that cannot be met by unstructured sparsity found in many applications. To address this, we introduce (S)parse (Ma)trix Matrix (T)ensor Core-accelerated (SMaT): a novel SpMM library that utilizes TCs for unstructured sparse matrices. Our block-sparse library leverages the low-level CUDA MMA (matrix-matrix-accumulate) API, maximizing the performance offered by modern GPUs. Algorithmic optimizations such as sparse matrix permutation further improve performance by minimizing the number of non-zero blocks. The evaluation on NVIDIA A100 shows that SMaT outperforms SotA libraries (DASP, cuSPARSE, and Magicube) by up to 125x (on average 2.6x). SMaT can be used to accelerate many workloads in scientific computing, large-model training, inference, and others.

Transformer training systems are built around dense linear algebra, yet a nontrivial fraction of end-to-end time is spent on surrounding memory-bound operators. Normalization, activations, residual updates, reductions, and related computations repeatedly move large intermediate tensors through global memory while performing little arithmetic, making data movement an increasingly important bottleneck in otherwise highly optimized training stacks. We introduce CODA, a GPU kernel abstraction that expresses these computations as GEMM-plus-epilogue programs. CODA is based on the observation that many Transformer operators exposed as separate framework kernels can be algebraically reparameterized to execute while a GEMM output tile remains on chip, before it is written to memory. The abstraction fixes the GEMM mainloop and exposes a small set of composable epilogue primitives for scaling, reductions, pairwise transformations, and accumulation. This constrained interface preserves the performance structure of expert-written GEMMs while remaining expressive enough to cover nearly all non-attention computation in the forward and backward pass of a standard Transformer block. Across representative Transformer workloads, both human- and LLM-authored CODA kernels achieve high performance, suggesting that GEMM-plus-epilogue programming offers a practical path toward combining framework-level productivity with hardware-level efficiency.

Every public LLM cost calculator we surveyed treats GPU utilization as a fixed input -- entered by the user, baked in as a preset, or silently assumed at 100% -- never measured against the operator's actual load. We show that this assumption is the dominant source of error: on identical H100 hardware, effective cost spans \0.21 to 15.25 per million output tokens, an underutilization penalty of 2.5-24x across low-to-moderate enterprise loads (1-10 rps) and up to 36.3x near idle -- driven by one operator-controlled variable, offered request rate lambda, which sets in-flight concurrency via Little's Law and which no open-source calculator exposes. Because calculators take utilization as a user-supplied input, any utilization-naive estimate understates true cost by exactly 1/U, systematically mispricing self-hosting -- most severely over-selling it for low-traffic workloads. We propose a measurement methodology that parameterizes the relationship as C_eff = f(H, M, Q, lambda, L), validate it with 42 benchmarks across dense, ultra-sparse MoE, and sparse MoE models, and release vllm-cost-meter, an open-source cost meter that attaches to a live vLLM server and reports real \$/M-tokens against the operator's own traffic. We further show that FP8 quantization benefits the MoE architectures we tested roughly 2.2-2.4x more than the dense model (+69 to +74% vs. +31% peak throughput; n=3, broader validation needed), and our data are consistent with active parameter count, not total model size, being a primary predictor of saturation economics. To rule out single-hardware confounding we repeat the core sweep on A100 80GB PCIe (56 runs): the load-driven spread reproduces at 7.0-11.4x, the active-parameters ordering survives at FP8, and the dense-FP8 advantage inverts on silicon without native FP8 tensor cores -- a hardware-conditional caveat the framework already accommodates.

It is widely acknowledged that large models have the potential to deliver superior performance across a broad range of domains. Despite the remarkable progress made in the field of machine learning systems research, which has enabled the development and exploration of large models, such abilities remain confined to a small group of advanced users and industry leaders, resulting in an implicit technical barrier for the wider community to access and leverage these technologies. In this paper, we introduce PyTorch Fully Sharded Data Parallel (FSDP) as an industry-grade solution for large model training. FSDP has been closely co-designed with several key PyTorch core components including Tensor implementation, dispatcher system, and CUDA memory caching allocator, to provide non-intrusive user experiences and high training efficiency. Additionally, FSDP natively incorporates a range of techniques and settings to optimize resource utilization across a variety of hardware configurations. The experimental results demonstrate that FSDP is capable of achieving comparable performance to Distributed Data Parallel while providing support for significantly larger models with near-linear scalability in terms of TFLOPS.

Deep Learning has been one of the most disruptive technological advancements in recent times. The high performance of deep learning models comes at the expense of high computational, storage and power requirements. Sensing the immediate need for accelerating and compressing these models to improve on-device performance, we introduce Deeplite Neutrino for production-ready optimization of the models and Deeplite Runtime for deployment of ultra-low bit quantized models on Arm-based platforms. We implement low-level quantization kernels for Armv7 and Armv8 architectures enabling deployment on the vast array of 32-bit and 64-bit Arm-based devices. With efficient implementations using vectorization, parallelization, and tiling, we realize speedups of up to 2x and 2.2x compared to TensorFlow Lite with XNNPACK backend on classification and detection models, respectively. We also achieve significant speedups of up to 5x and 3.2x compared to ONNX Runtime for classification and detection models, respectively.

Long-context inference in large language models is bottlenecked by Key--Value (KV) cache loading during the decoding stage, where the sequential nature of generation requires repeatedly transferring the KV cache from off-chip High-Bandwidth Memory (HBM) to on-chip Static Random-Access Memory (SRAM) at each step. While Multi-Head Latent Attention (MLA) significantly reduces the total KV cache size, it suffers from a sharding bottleneck during distributed decoding via Tensor Parallelism (TP). Since its single latent head cannot be partitioned, each device is forced to redundantly load the complete KV cache for every token, consuming excessive memory traffic and diminishing TP benefits like weight sharding. In this work, we propose Multi-Head Low-Rank Attention (MLRA), which enables partitionable latent states for efficient 4-way TP decoding. Extensive experiments show that MLRA achieves state-of-the-art perplexity and downstream task performance, while also delivering a 2.8times decoding speedup over MLA. Code is available at https://github.com/SongtaoLiu0823/MLRA. Pretrained weights, along with the training and evaluation data, are available at https://huggingface.co/Soughing/MLRA.

The rapid growth of Large Transformer-based models, specifically Large Language Models (LLMs), now scaling to trillions of parameters, has necessitated training across thousands of GPUs using complex hybrid parallelism strategies (e.g., data, tensor, and pipeline parallelism). Checkpointing this massive, distributed state is critical for a wide range of use cases, such as resilience, suspend-resume, investigating undesirable training trajectories, and explaining model evolution. However, existing checkpointing solutions typically treat model state as opaque binary blobs, ignoring the ``3D heterogeneity'' of the underlying data structures--varying by memory location (GPU vs. Host), number of ``logical'' objects sharded and split across multiple files, data types (tensors vs. Python objects), and their serialization requirements. This results in significant runtime overheads due to blocking device-to-host transfers, data-oblivious serialization, and storage I/O contention. In this paper, we introduce DataStates-LLM, a novel checkpointing architecture that leverages State Providers to decouple state abstraction from data movement. DataStates-LLM exploits the immutability of model parameters during the forward and backward passes to perform ``lazy'', non-blocking asynchronous snapshots. By introducing State Providers, we efficiently coalesce fragmented, heterogeneous shards and overlap the serialization of metadata with bulk tensor I/O. We evaluate DataStates-LLM on models up to 70B parameters on 256 A100-40GB GPUs. Our results demonstrate that DataStates-LLM achieves up to 4times higher checkpointing throughput and reduces end-to-end training time by up to 2.2times compared to state-of-the-art solutions, effectively mitigating the serialization and heterogeneity bottlenecks in extreme-scale LLM training.

Large language model (LLM) decoding is latency-sensitive and often bottlenecked by fragmented operator execution and repeated off-chip materialization of intermediate tensors. Prior work expands fusion scope by leveraging thread-block clusters and on-chip inter-block collectives to fuse attention-side operators such as QKV projection, attention, and output projection. We develop ClusterFusion++, a CUDA-level extension that broadens fusion to the full Transformer decoder block for GPT-NeoX/Pythia models: LayerNorm -> QKV -> RoPE -> decode attention -> output projection -> Post-LN -> MLP -> residual. We additionally engineer a CUDA-Graph-compatible execution mode with persistent Tensor Memory Accelerator (TMA) descriptors to reduce per-step overhead. On an NVIDIA RTX 5090-class GPU, ClusterFusion++ improves throughput by 1.34x for Pythia-2.8B and yields similar gains for Pythia-6.9B, while maintaining high output fidelity (near-token-identical generation, with minor non-determinism from FP16 atomics).

The growth rate of the GPU memory capacity has not been able to keep up with that of the size of large language models (LLMs), hindering the model training process. In particular, activations -- the intermediate tensors produced during forward propagation and reused in backward propagation -- dominate the GPU memory use. This leads to high training overhead such as high weight update cost due to the small micro-batch size. To address this challenge, we propose SSDTrain, an adaptive activation offloading framework to high-capacity NVMe SSDs. SSDTrain reduces GPU memory usage without impacting performance by fully overlapping data transfers with computation. SSDTrain is compatible with popular deep learning frameworks like PyTorch, Megatron, and DeepSpeed, and it employs techniques such as tensor deduplication and forwarding to further enhance efficiency. We extensively experimented with popular LLMs like GPT, BERT, and T5. Results demonstrate that SSDTrain reduces 47% of the activation peak memory usage. Meanwhile, SSDTrain perfectly overlaps the I/O with the computation and incurs negligible overhead. Compared with keeping activations in GPU memory and layerwise full recomputation, SSDTrain achieves the best memory savings with negligible throughput loss. We further analyze how the reduced activation memory use may be leveraged to increase throughput by increasing micro-batch size and reducing pipeline parallelism bubbles.

Training large language models requires distributing computation across many accelerators, yet practitioners select parallelism strategies (data, tensor, pipeline, ZeRO) through trial and error because no unified systematic framework predicts their behavior. We introduce placement semantics: each strategy is specified by how it places four training states (parameters, optimizer, gradients, activations) across devices using five modes (replicated, sharded, sharded-with-gather, materialized, offloaded). From placement alone, without implementation details, we derive memory consumption and communication volume. Our predictions match published results exactly: ZeRO-3 uses 8x less memory than data parallelism at 1.5x communication cost, as reported in the original paper. We prove two conditions (gradient integrity, state consistency) are necessary and sufficient for distributed training to match single-device results, and provide composition rules for combining strategies safely. The framework unifies ZeRO Stages 1-3, Fully Sharded Data Parallel (FSDP), tensor parallelism, and pipeline parallelism as instances with different placement choices.

In this paper, we delve into the foundational principles of tensor categories, harnessing the universal property of the tensor product to pioneer novel methodologies in deep network architectures. Our primary contribution is the introduction of the Tensor Attention and Tensor Interaction Mechanism, a groundbreaking approach that leverages the tensor category to enhance the computational efficiency and the expressiveness of deep networks, and can even be generalized into the quantum realm.

The large language model era urges faster and less costly inference. Prior model compression works on LLMs tend to undertake a software-centric approach primarily focused on the simulated quantization performance. By neglecting the feasibility of deployment, these approaches are typically disabled in real practice. They used to drastically push down the quantization bit range for a reduced computation which might not be supported by the mainstream hardware, or involve sophisticated algorithms that introduce extra computation or memory access overhead. We argue that pursuing a hardware-centric approach in the construction of quantization algorithms is crucial. In this regard, we are driven to build our compression method on top of hardware awareness, eliminating impractical algorithm choices while maximizing the benefit of hardware acceleration. Our method, OdysseyLLM, comes with a novel W4A8 kernel implementation called FastGEMM and a combined recipe of quantization strategies. Extensive experiments manifest the superiority of our W4A8 method which brings the actual speed boosting up to 4times compared to Hugging Face FP16 inference and 2.23times vs. the state-of-the-art inference engine TensorRT-LLM in FP16, and 1.45times vs. TensorRT-LLM in INT8, yet without substantially harming the performance.

Vision backbone networks play a central role in modern computer vision. Enhancing their efficiency directly benefits a wide range of downstream applications. To measure efficiency, many publications rely on MACs (Multiply Accumulate operations) as a predictor of execution time. In this paper, we experimentally demonstrate the shortcomings of such a metric, especially in the context of edge devices. By contrasting the MAC count and execution time of common architectural design elements, we identify key factors for efficient execution and provide insights to optimize backbone design. Based on these insights, we present LowFormer, a novel vision backbone family. LowFormer features a streamlined macro and micro design that includes Lowtention, a lightweight alternative to Multi-Head Self-Attention. Lowtention not only proves more efficient, but also enables superior results on ImageNet. Additionally, we present an edge GPU version of LowFormer, that can further improve upon its baseline's speed on edge GPU and desktop GPU. We demonstrate LowFormer's wide applicability by evaluating it on smaller image classification datasets, as well as adapting it to several downstream tasks, such as object detection, semantic segmentation, image retrieval, and visual object tracking. LowFormer models consistently achieve remarkable speed-ups across various hardware platforms compared to recent state-of-the-art backbones. Code and models are available at https://github.com/altair199797/LowFormer/blob/main/Beyond_MACs.md.

The Mixture of Experts (MoE) model becomes an important choice of large language models nowadays because of its scalability with sublinear computational complexity for training and inference. However, existing MoE models suffer from two critical drawbacks, 1) tremendous inner-node and inter-node communication overhead introduced by all-to-all dispatching and gathering, and 2) limited scalability for the backbone because of the bound data parallel and expert parallel to scale in the expert dimension. In this paper, we systematically analyze these drawbacks in terms of training efficiency in the parallel framework view and propose a novel MoE architecture called Pipeline MoE (PPMoE) to tackle them. PPMoE builds expert parallel incorporating with tensor parallel and replaces communication-intensive all-to-all dispatching and gathering with a simple tensor index slicing and inner-node all-reduce. Besides, it is convenient for PPMoE to integrate pipeline parallel to further scale the backbone due to its flexible parallel architecture. Extensive experiments show that PPMoE not only achieves a more than 1.75times speed up compared to existing MoE architectures but also reaches 90% throughput of its corresponding backbone model that is 20times smaller.

In the classical transformer attention scheme, we are given three n times d size matrices Q, K, V (the query, key, and value tokens), and the goal is to compute a new n times d size matrix D^{-1} exp(QK^top) V where D = diag( exp(QK^top) {bf 1}_n ). In this work, we study a generalization of attention which captures triple-wise correlations. This generalization is able to solve problems about detecting triple-wise connections that were shown to be impossible for transformers. The potential downside of this generalization is that it appears as though computations are even more difficult, since the straightforward algorithm requires cubic time in n. However, we show that in the bounded-entry setting (which arises in practice, and which is well-studied in both theory and practice), there is actually a near-linear time algorithm. More precisely, we show that bounded entries are both necessary and sufficient for quickly performing generalized computations: bullet On the positive side, if all entries of the input matrices are bounded above by o(sqrt[3]{log n}) then we show how to approximate the ``tensor-type'' attention matrix in n^{1+o(1)} time. bullet On the negative side, we show that if the entries of the input matrices may be as large as Omega(sqrt[3]{log n}), then there is no algorithm that runs faster than n^{3-o(1)} (assuming the Strong Exponential Time Hypothesis from fine-grained complexity theory). We also show that our construction, algorithms, and lower bounds naturally generalize to higher-order tensors and correlations. Interestingly, the higher the order of the tensors, the lower the bound on the entries needs to be for an efficient algorithm. Our results thus yield a natural tradeoff between the boundedness of the entries, and order of the tensor one may use for more expressive, efficient attention computation.

We present HyperQuant (Hadamard, optimallY Packing, Entropy Rice-coding), a unified post-training quantization pipeline for the weights and the KV cache of large language and diffusion transformers. Across a suite of self-contained experiments (Table 1), HyperQuant outperforms the recent HIGGS scheme at every operating point from 3 to 5 bits per scalar (bps) on weights, and beats both TurboQuant and OCTOPUS on KV quantization down to 1.7 bps. Beyond the LLM setting, HyperQuant quantizes the 19B-parameter LTX-2 DiT video model with no observable per-frame artifacts. End-to-end on an H100 at 4 bps, HyperQuant compresses the linear weights ~3.9x and the KV cache ~3.79x at near-lossless quality. HyperQuant combines four known ideas into a single construction: (i) a per-tile Randomized Hadamard Transform that makes the per-coordinate distribution of weights and activations approximately Gaussian; (ii) quantization to a low-dimensional optimal lattice (E8, D4, A2, or Z); (iii) lossless bit-stripping and near-entropy-optimal variable-length Rice coding of the lattice indices; and (iv) bias-correction methods for the KV cache that keep the reconstruction unbiased under inner products, preserving attention semantics. We further integrate the pipeline with 8-bit and 4-bit Tensor-Core MMA paths (fp8-e4m3, int8, nvfp4, mxfp4), and find that int8 beats fp8 on the post-RHT lattice output. Project page: https://moonmath.ai/hyperquant/

Diffusion Transformers deliver state-of-the-art video generation, but their long spatiotemporal sequences make attention the dominant deployment cost, and a deployable low-bit kernel must be accurate and fast. Accuracy is limited by outliers: a block's quantization scale is set by its largest entries, leaving typical entries confined to a narrow range of representable values. Prior work smooths queries and keys, but value outliers follow no fixed channel or spatiotemporal structure and remain the dominant source of output error. Speed is limited by softmax: low-bit Tensor Cores accelerate only the two matrix multiplications, so the high-precision exponential between them becomes the longest pipeline stage on datacenter GPUs. We propose VC-Attention, a training-free low-bit attention framework that addresses both by pairing Value smoothing with a fused probability Cast. V-Smooth reorders value tokens by lightweight online clustering, so the tokens in a hardware block quantize well together. It quantizes only the residual after subtracting the block mean, and restores that mean from the row sum the online softmax already maintains. ExpCast-FP8 maps log-domain scores directly to E4M3 probability codes with one fused multiply-add, eliminating the FP32 exponential and the format conversion. We implement VC-Attention for B200, B300, H200, RTX PRO 6000, and RTX 5090. Across Wan2.2, LongCat-Video, HunyuanVideo-1.5, and MiniMax-H3, VC-Attention improves fidelity over low-bit baselines, speeds up the attention kernel over BF16 FlashAttention-4 by 1.46-1.59x on datacenter Blackwell and Hopper and by 2.3-3.6x on workstation cards, and generates a clip 1.13-1.19x and 1.36-1.70x faster end to end.

Stable 4-bit floating-point (FP4) pretraining is difficult because the E2M1 payload represents only a narrow range of magnitudes. NVIDIA's Transformer Engine recipe addresses this with current-tensor scaling, a randomized Hadamard transform (RHT), and bfloat16 (BF16) final layers, adding work outside the FP4 matrix multiplications. We instead pair E2M1 payloads with unsigned E5M3 () block scales. Their wider range permits periodic tensor scaling, while our recipe applies selective stochastic rounding to backward gradients, omits RHT, and uses FP4 in all eligible internal linears. We pretrain a Nemotron-H 8B model for nearly 190 billion tokens. Compared with Transformer Engine , the proposed block-16 recipe finishes with lower final-window training loss and, under their respective quantized-inference policies, lower validation loss measured as held-out negative log-likelihood. Its quantized-inference downstream point estimates are also higher on all three reported aggregates. A native execution ablation that jointly removes RHT and the BF16 final-block exemption increases measured model-body token throughput by 21.2\%. These results demonstrate end-to-end software-emulated pretraining with a simpler recipe and motivate native support for block scaling.

During early optimization passes, compilers must make predictions for machine-dependent characteristics such as execution unit utilization, number of register spills, latency, throughput etc. to generate better code. Often a hand-written static/analytical hardware cost model is built into the compiler. However, the need for more sophisticated and varied predictions has become more pronounced with the development of deep learning compilers which need to optimize dataflow graphs. Such compilers usually employ a much higher level MLIR form as an IR representation before lowering to traditional LLVM-IR. A static/analytical cost model in such a scenario is cumbersome and error prone as the opcodes represent very high level algebraic/arithmetic operations. Hence, we develop a machine learning-based cost model for high-level MLIR which can predict different target variables of interest such as CPU/GPU/xPU utilization, instructions executed, register usage etc. By considering the incoming MLIR as a text input a la NLP models we can apply well-known techniques from modern NLP research to help predict hardware characteristics more accurately. We expect such precise ML-driven hardware cost models to guide our deep learning compiler in graph level optimizations around operator fusion, local memory allocation, kernel scheduling etc. as well as in many kernel-level optimizations such as loop interchange, LICM and unroll. We report early work-in -progress results of developing such models on high-level MLIR representing dataflow graphs emitted by Pytorch/Tensorflow-like frameworks as well as lower-level dialects like affine. We show that these models can provide reasonably good estimates with low error bounds for various hardware characteristics of interest and can be a go-to mechanism for hardware cost modelling in the future.

Most state of the art deep neural networks are overparameterized and exhibit a high computational cost. A straightforward approach to this problem is to replace convolutional kernels with its low-rank tensor approximations, whereas the Canonical Polyadic tensor Decomposition is one of the most suited models. However, fitting the convolutional tensors by numerical optimization algorithms often encounters diverging components, i.e., extremely large rank-one tensors but canceling each other. Such degeneracy often causes the non-interpretable result and numerical instability for the neural network fine-tuning. This paper is the first study on degeneracy in the tensor decomposition of convolutional kernels. We present a novel method, which can stabilize the low-rank approximation of convolutional kernels and ensure efficient compression while preserving the high-quality performance of the neural networks. We evaluate our approach on popular CNN architectures for image classification and show that our method results in much lower accuracy degradation and provides consistent performance.

Mixture of Experts (MoE) models enhance neural network scalability by dynamically selecting relevant experts per input token, enabling larger model sizes while maintaining manageable computation costs. However, efficient training of large-scale MoE models across thousands of GPUs presents significant challenges due to limitations in existing parallelism strategies. We introduce an end-to-end training framework for large-scale MoE models that utilizes five-dimensional hybrid parallelism: Tensor Parallelism, Expert Parallelism, Context Parallelism, Data Parallelism, and Pipeline Parallelism. Central to our approach is MoE Parallel Folding, a novel strategy that decouples the parallelization of attention and MoE layers in Transformer models, allowing each layer type to adopt optimal parallel configurations. Additionally, we develop a flexible token-level dispatcher that supports both token-dropping and token-dropless MoE training across all five dimensions of parallelism. This dispatcher accommodates dynamic tensor shapes and coordinates different parallelism schemes for Attention and MoE layers, facilitating complex parallelism implementations. Our experiments demonstrate significant improvements in training efficiency and scalability. We achieve up to 49.3% Model Flops Utilization (MFU) for the Mixtral 8x22B model and 39.0% MFU for the Qwen2-57B-A14B model on H100 GPUs, outperforming existing methods. The framework scales efficiently up to 1,024 GPUs and maintains high performance with sequence lengths up to 128K tokens, validating its effectiveness for large-scale MoE model training. The code is available in Megatron-Core.

In this paper, we present a novel method that reduces model inference latency during distributed deployment of Large Language Models (LLMs). Our contribution is an optimized inference deployment scheme that address the current limitations of state-of-the-art quantization kernels when used in conjunction with Tensor Parallel (TP). Our method preserves data locality in GPU memory access patterns and exploits a priori knowledge of TP to reduce global communication. We demonstrate an up to 1.81x speedup over existing methods for Llama-70B and up to 1.78x speedup for IBM WatsonX's Granite-20B MLP layer problem sizes on A100 and H100 NVIDIA DGX Systems for a variety of TP settings.

Large transformer models display promising performance on a wide range of natural language processing (NLP) tasks. Although the AI community has expanded the model scale to the trillion parameter level, the practical deployment of 10-100 billion parameter models is still uncertain due to the latency, throughput, and memory constraints. In this paper, we proposed EnergonAI to solve the challenges of the efficient deployment of 10-100 billion parameter transformer models on single- or multi-GPU systems. EnergonAI adopts a hierarchy-controller system architecture to coordinate multiple devices and efficiently support different parallel patterns. It delegates the execution of sub-models to multiple workers in the single-controller style and applies tensor parallelism and pipeline parallelism among the workers in a multi-controller style. Upon the novel architecture, we propose three techniques, i.e. non-blocking pipeline parallelism, distributed redundant computation elimination, and peer memory pooling. EnergonAI enables the users to program complex parallel code the same as a serial one. Compared with the FasterTransformer, we have proven that EnergonAI has superior performance on latency and throughput. In our experiments, EnergonAI can achieve 37% latency reduction in tensor parallelism, 10% scalability improvement in pipeline parallelism, and it improves the model scale inferred on a single GPU by using a larger heterogeneous memory space at cost of limited performance reduction.

Training dense LLMs requires enormous amounts of data and centralized compute, which introduces fundamental bottlenecks and ever-growing costs for large models. Several studies aim to reduce this dependency on centralization by reducing the communication overhead of training dense models. Taking this idea of reducing communication overhead to a natural extreme, by training embarrassingly parallelizable ensembles of small independent experts, has been shown to outperform large dense models trained in traditional centralized settings. However, existing studies do not take into account underlying differences amongst data domains and treat them as monolithic, regardless of their underlying complexity, size, or distribution. In this paper, we explore the effects of introducing heterogeneity to these ensembles of domain expert models. Specifically, by allowing models within the ensemble to vary in size--as well as the number of training steps taken depending on the training data's domain--we study the effect heterogeneity has on these ensembles when evaluated against domains included in, and excluded from, the training set. We use the same compute budget to train heterogeneous ensembles and homogeneous baselines for comparison. We show that the heterogeneous ensembles achieve the lowest perplexity scores in 20 out of the 21 data domains used in the evaluation. Our code is available at https://github.com/gensyn-ai/hdee.

The deployment and training of neural networks on edge computing devices pose many challenges. The low memory nature of edge devices is often one of the biggest limiting factors encountered in the deployment of large neural network models. Tensor rematerialization or recompute is a way to address high memory requirements for neural network training and inference. In this paper we consider the problem of execution time minimization of compute graphs subject to a memory budget. In particular, we develop a new constraint programming formulation called Moccasin with only O(n) integer variables, where n is the number of nodes in the compute graph. This is a significant improvement over the works in the recent literature that propose formulations with O(n^2) Boolean variables. We present numerical studies that show that our approach is up to an order of magnitude faster than recent work especially for large-scale graphs.

Deep learning increasingly relies on massive data with substantial storage, annotation, and training costs. To reduce costs, coreset selection finds a representative subset of data to train models while ideally performing on par with the full data training. To maximize performance, current state-of-the-art coreset methods select data using dataset-specific ground truth labels and training. However, these methodological requirements prevent selection at scale on real-world, unlabeled data. To that end, this paper addresses the selection of coresets that achieve state-of-the-art performance but without using any labels or training on candidate data. Instead, our solution, Zero-Shot Coreset Selection via Iterative Subspace Sampling (ZCore), uses previously-trained foundation models to generate zero-shot, high-dimensional embedding spaces to interpret unlabeled data. ZCore then iteratively quantifies the relative value of all candidate data based on coverage and redundancy in numerous subspace distributions. Finally, ZCore selects a coreset sized for any data budget to train downstream models. We evaluate ZCore on four datasets and outperform several state-of-the-art label-based methods, especially at low data rates that provide the most substantial cost reduction. On ImageNet, ZCore selections for 10% training data achieve a downstream validation accuracy of 53.99%, which outperforms prior label-based methods and removes annotation and training costs for 1.15 million images. Our paper's code is publicly available at https://github.com/voxel51/zcore.

EdgeAI systems are increasingly employing computer vision applications to enable intelligent, on-device decision-making in real-time. However, these deployments face highly dynamic operational conditions, with fluctuating constraints on latency, power availability, and memory resources. Deep Neural Networks (DNN), which follow fixed computational execution flows, lack the flexibility to adapt to such variability, resulting in inefficient and suboptimal performance in edge scenarios. This underscores the need for architectures that are not only efficient but also dynamically scalable at runtime. In this paper, we propose Elastoformer: A framework that transforms conventional neural networks (NN) into Elastic NN capable of real-time elastic inference. Unlike the conventional bag-of-models approach, which requires maintaining multiple independent models for different operating conditions, Elastoformer offers a single, modular solution that dynamically switches between multiple modes of operation at runtime, adapting efficiently to the changing computational budgets of edge devices without the overhead of managing separate models. Experiments reveal that our framework achieves up to 85% reduction in computation FLOPs, 50% reduction in latency and 76% reduction in memory overhead, while showcasing the architecture agnostic nature of the framework across both Vision Transformers and CNNs. Our code is available at https://github.com/sudaksh14/Elastoformer.

Many researchers believe that ConvNets perform well on small or moderately sized datasets, but are not competitive with Vision Transformers when given access to datasets on the web-scale. We challenge this belief by evaluating a performant ConvNet architecture pre-trained on JFT-4B, a large labelled dataset of images often used for training foundation models. We consider pre-training compute budgets between 0.4k and 110k TPU-v4 core compute hours, and train a series of networks of increasing depth and width from the NFNet model family. We observe a log-log scaling law between held out loss and compute budget. After fine-tuning on ImageNet, NFNets match the reported performance of Vision Transformers with comparable compute budgets. Our strongest fine-tuned model achieves a Top-1 accuracy of 90.4%.

The challenge of mapping AI architectures to GPU hardware is creating a critical bottleneck in AI progress. Despite substantial efforts, hand-written custom kernels fail to meet their theoretical performance thresholds, even on well-established operations like linear attention. The diverse hardware capabilities of GPUs might suggest that we need a wide variety of techniques to achieve high performance. However, our work explores whether a small number of key abstractions can drastically simplify the process. We present ThunderKittens (TK), a framework for writing performant AI kernels while remaining easy to use and maintain. Our abstractions map to the three levels of the GPU hierarchy: (1) at the warp-level, we provide 16x16 matrix tiles as basic data structures and PyTorch-like parallel compute operations over tiles, (2) at the thread-block level, we provide a template for overlapping asynchronous operations across parallel warps, and (3) at the grid-level, we provide support to help hide the block launch and tear-down, and memory costs. We show the value of TK by providing kernels that match or outperform prior kernels for a range of AI operations. We match CuBLAS and FlashAttention-3 on GEMM and attention inference performance and outperform the strongest baselines by 10-40% on attention backwards, 8times on state space models, and 14times on linear attention.

LLMs have seen rapid adoption in all domains. They need to be trained on high-end high-performance computing (HPC) infrastructures and ingest massive amounts of input data. Unsurprisingly, at such a large scale, unexpected events (e.g., failures of components, instability of the software, undesirable learning patterns, etc.), are frequent and typically impact the training in a negative fashion. Thus, LLMs need to be checkpointed frequently so that they can be rolled back to a stable state and subsequently fine-tuned. However, given the large sizes of LLMs, a straightforward checkpointing solution that directly writes the model parameters and optimizer state to persistent storage (e.g., a parallel file system), incurs significant I/O overheads. To address this challenge, in this paper we study how to reduce the I/O overheads for enabling fast and scalable checkpointing for LLMs that can be applied at high frequency (up to the granularity of individual iterations) without significant impact on the training process. Specifically, we introduce a lazy asynchronous multi-level approach that takes advantage of the fact that the tensors making up the model and optimizer state shards remain immutable for extended periods of time, which makes it possible to copy their content in the background with minimal interference during the training process. We evaluate our approach at scales of up to 180 GPUs using different model sizes, parallelism settings, and checkpointing frequencies. The results show up to 48times faster checkpointing and 2.2times faster end-to-end training runtime compared with the state-of-art checkpointing approaches.

Sequence parallelism (SP), which divides the sequence dimension of input tensors across multiple computational devices, is becoming key to unlocking the long-context capabilities of generative AI models. This paper investigates the state-of-the-art SP approaches, i.e. DeepSpeed-Ulysses and Ring-Attention, and proposes a unified SP approach, which is more robust to transformer model architectures and network hardware topology. This paper compares the communication and memory cost of SP and existing parallelism, including data/tensor/zero/expert/pipeline parallelism, and discusses the best practices for designing hybrid 4D parallelism involving SP. We achieved 86% MFU on two 8xA800 nodes using SP for sequence length 208K for the LLAMA3-8B model. Our code is publicly available on https://github.com/feifeibear/long-context-attention.

The key-value (KV) cache accelerates LLMs decoding by storing KV tensors from previously generated tokens. It reduces redundant computation at the cost of increased memory usage. To mitigate this overhead, existing approaches compress KV tensors into lower-bit representations; however, quantization errors can accumulate as more tokens are generated, potentially resulting in undesired outputs. In this paper, we introduce SQuat (Subspace-orthogonal KV cache quantization). It first constructs a subspace spanned by query tensors to capture the most critical task-related information. During key tensor quantization, it enforces that the difference between the (de)quantized and original keys remains orthogonal to this subspace, minimizing the impact of quantization errors on the attention mechanism's outputs. SQuat requires no model fine-tuning, no additional calibration dataset for offline learning, and is grounded in a theoretical framework we develop. Through numerical experiments, we show that our method reduces peak memory by 2.17 to 2.82, improves throughput by 2.45 to 3.60, and achieves more favorable benchmark scores than existing KV cache quantization algorithms.

Transformers are mostly relying on softmax attention, which introduces quadratic complexity with respect to sequence length and remains a major bottleneck for efficient inference. Prior work on linear or hybrid attention typically replaces softmax attention uniformly across all layers, often leading to significant performance degradation or requiring extensive retraining to recover model quality. This work proposes LayerBoost, a layer-aware attention reduction method that selectively modifies the attention mechanism based on the sensitivity of individual transformer layers. It first performs a systematic sensitivity analysis on a pretrained model to identify layers that are critical for maintaining performance. Guided by this analysis, three distinct strategies can be applied: retaining standard softmax attention in highly sensitive layers, replacing it with linear sliding window attention in moderately sensitive layers, and removing attention entirely in layers that exhibit low sensitivity. To recover performance after these architectural modifications, we introduce a lightweight distillation-based healing phase requiring only 10M additional training tokens. LayerBoost reduces inference latency and improves throughput by up to 68% at high concurrency, while maintaining competitive model quality. It matches base model performance on several benchmarks, exhibits only minor degradations on others, and significantly outperforms state-of-the-art attention linearization methods. These efficiency gains make our method particularly well-suited for high-concurrency serving and hardware-constrained deployment scenarios, where inference cost and memory footprint are critical bottlenecks.

Despite extensive progress on image generation, common deep generative model architectures are not easily applied to lossless compression. For example, VAEs suffer from a compression cost overhead due to their latent variables. This overhead can only be partially eliminated with elaborate schemes such as bits-back coding, often resulting in poor single-sample compression rates. To overcome such problems, we establish a new class of tractable lossless compression models that permit efficient encoding and decoding: Probabilistic Circuits (PCs). These are a class of neural networks involving |p| computational units that support efficient marginalization over arbitrary subsets of the D feature dimensions, enabling efficient arithmetic coding. We derive efficient encoding and decoding schemes that both have time complexity O (log(D) cdot |p|), where a naive scheme would have linear costs in D and |p|, making the approach highly scalable. Empirically, our PC-based (de)compression algorithm runs 5-40 times faster than neural compression algorithms that achieve similar bitrates. By scaling up the traditional PC structure learning pipeline, we achieve state-of-the-art results on image datasets such as MNIST. Furthermore, PCs can be naturally integrated with existing neural compression algorithms to improve the performance of these base models on natural image datasets. Our results highlight the potential impact that non-standard learning architectures may have on neural data compression.

Tensor networks, originally developed for quantum many-body physics, are promising models for machine learning. We derive stochastic Riemannian optimizers for tree tensor networks (TTNs) on both their parameter and quotient manifolds, including adaptive and learning-rate-free schemes suitable for minibatch training. Using a hybrid CNN-TTN architecture, we evaluate the methods on Fashion-MNIST, CIFAR10, and Imagenette. The proposed optimizers achieve predictive performance comparable to unconstrained optimization while enabling numerically stable downstream compression.

Weight-only post-training quantization is the cheapest way to shrink a retrieval embedder, and the received advice for applying it -- protect the embedding table, allocate bits by module sensitivity, prefer a ranking-aware objective over weight reconstruction -- was carried into LLM quantization largely intact. We test that advice on retrieval embedders directly, quantizing five checkpoints from four architecture families across a grid of bit widths and group sizes, and isolating the embedding, attention and feed-forward blocks at each width. Every heuristic fails to transfer as stated. The embedding table never emerges as the dominant isolated protection priority in any family, despite being the largest tensor in several of them. Module sensitivity does not survive as a transferable ordering: at INT4/g16 the spread between modules is too small to allocate against, at INT3 the ordering becomes family-dependent and joint damage stops being the sum of its parts, and at INT2 comparable reconstruction error accompanies retention ranging from 1.3 to 65.9 percent of full precision. A cheap reconstruction proxy is useful for screening uniform bit widths but substantially less reliable for choosing which tensors to protect; its apparent strength across the whole grid is a range-extension artifact. A distilled 109M student at INT3 holds 78.04 NDCG@10 in 68.4 MB and dominates the extreme-PTQ arm of its own 0.6B teacher, 297.9 MB at 64.46, on both size and quality -- but only inside the task it was distilled for. Sizes are byte counts of files that exist rather than arithmetic estimates, and the measurement repository carries the byte provenance for every one of them.

Concurrent local LLM serving on unified-memory desktops must preserve memory headroom and output fidelity, which speed-only rankings overlook. We introduce SiliconBench, which evaluates nine Apple Silicon serving engines through three lenses: speed, memory, and fidelity. We evaluate chat and agent serving on Qwen3, Qwen3.5, and Gemma 4. We use a classification task to check for quality regressions against an NVIDIA reference. DGX Spark provides a complementary serving-performance reference. Three desiderata guide interpretation: serving architecture readiness, memory discipline, and multi-node scaling. On Qwen3-0.6B, vllm-metal alone more than doubles throughput on both workloads from concurrency 1 to 16. CUDA vLLM and SGLang show stronger concurrency scaling on the same prompts. Explicit memory budgets do not guarantee memory headroom: two stacks complete every request while memory use approaches physical capacity and throughput declines. The newer model architectures have narrower engine support. Their evaluated implementations match the fidelity reference. Only three stacks satisfy the completion, fidelity, and model-coverage gates. Comparisons on larger dense and MoE models reinforce the importance of scheduling prompt processing alongside ongoing generation: vllm-metal's packed prefill-decode path maintains lower first-token latency than omlx under concurrent load. In the tested two-machine configurations, tensor parallelism over Thunderbolt RDMA scales while pipeline parallelism over TCP regresses. We release benchmark code, per-run results, and maintenance journals, supported by a workflow combining bounded agent fixes with human review.

Attention, as a core layer of the ubiquitous Transformer architecture, is the bottleneck for large language models and long-context applications. FlashAttention elaborated an approach to speed up attention on GPUs through minimizing memory reads/writes. However, it has yet to take advantage of new capabilities present in recent hardware, with FlashAttention-2 achieving only 35% utilization on the H100 GPU. We develop three main techniques to speed up attention on Hopper GPUs: exploiting asynchrony of the Tensor Cores and TMA to (1) overlap overall computation and data movement via warp-specialization and (2) interleave block-wise matmul and softmax operations, and (3) block quantization and incoherent processing that leverages hardware support for FP8 low-precision. We demonstrate that our method, FlashAttention-3, achieves speedup on H100 GPUs by 1.5-2.0times with FP16 reaching up to 740 TFLOPs/s (75% utilization), and with FP8 reaching close to 1.2 PFLOPs/s. We validate that FP8 FlashAttention-3 achieves 2.6times lower numerical error than a baseline FP8 attention.

GPU collective communication is typically optimized for bandwidth, yet many emerging workloads are increasingly limited by latency. Long-context decode-heavy large language model (LLM) inference is a prime example, where serving large models requires multiple GPUs, and many small collectives lie directly on the critical path of token generation. Therefore, even microsecond of overhead can impact performance and cost. In this work, we study how to approach the hardware Speed-of-Light (SoL) lower bound for GPU collectives within a scale-up network. We identify key principles for near-optimal designs, including barrier-free synchronization and efficient use of symmetric memory and multicast. Building on NCCL's device-side API, we develop low-latency interfaces for constructing custom collective kernels and use them to implement new symmetric collectives in NCCL. Microbenchmarks show substantial latency reductions for small and medium messages, reducing overhead to within 7% of the absolute SoL lower bound. When integrated into real applications, these kernels improve inter-token latency and throughput in LLM inference and accelerate cuSOLVERMp, demonstrating benefits for both AI inference and traditional HPC workloads.

As deep learning models scale, their training cost has surged significantly. Due to both hardware advancements and limitations in current software stacks, the need for data efficiency has risen. Data efficiency refers to the effective hiding of data access latency and the avoidance of unnecessary data movements. Major challenges arise from the growing disparity between GPU memory bandwidth and computational throughput, imminent GPU memory capacity limitations, and inefficiencies in the PyTorch software stack, including a lack of device-specific PCIe transfer optimizations and high-level domain-specific abstractions. To effectively mitigate these data inefficiencies for deep learning training, this dissertation analyzes data inefficiency in representative deep training tasks, specifically in graph neural networks (GNNs) and large language models (LLMs). It then proposes novel runtime and code generation techniques to mitigate these challenges and implements these optimizations seamlessly within the PyTorch stack while maintaining strong programmability and interoperability. First, PyTorch-Direct is devised to incorporate the GPU-centric PCIe data transfer paradigm in PyTorch for GNN training. Next, Hector intermediate representation (IR) and its code generator are proposed to introduce domain-specific high-level abstraction and systematically address memory-intensive performance challenges for relational GNNs. Finally, in LLM training, the throughput has been increasingly constrained by GPU memory capacity. To mitigate this, the SSDTrain offloading framework is designed and implemented. Together, these contributions show that code generation and runtime techniques can systematically mitigate the data management bottlenecks in deep learning training, which stem from the data-intensive nature of workloads and the oversimplification inherent in the deep learning training software stack.

Neural architecture search (NAS) has a great impact by automatically designing effective neural network architectures. However, the prohibitive computational demand of conventional NAS algorithms (e.g. 10^4 GPU hours) makes it difficult to directly search the architectures on large-scale tasks (e.g. ImageNet). Differentiable NAS can reduce the cost of GPU hours via a continuous representation of network architecture but suffers from the high GPU memory consumption issue (grow linearly w.r.t. candidate set size). As a result, they need to utilize~proxy tasks, such as training on a smaller dataset, or learning with only a few blocks, or training just for a few epochs. These architectures optimized on proxy tasks are not guaranteed to be optimal on the target task. In this paper, we present ProxylessNAS that can directly learn the architectures for large-scale target tasks and target hardware platforms. We address the high memory consumption issue of differentiable NAS and reduce the computational cost (GPU hours and GPU memory) to the same level of regular training while still allowing a large candidate set. Experiments on CIFAR-10 and ImageNet demonstrate the effectiveness of directness and specialization. On CIFAR-10, our model achieves 2.08\% test error with only 5.7M parameters, better than the previous state-of-the-art architecture AmoebaNet-B, while using 6times fewer parameters. On ImageNet, our model achieves 3.1\% better top-1 accuracy than MobileNetV2, while being 1.2times faster with measured GPU latency. We also apply ProxylessNAS to specialize neural architectures for hardware with direct hardware metrics (e.g. latency) and provide insights for efficient CNN architecture design.