Dynamic quantization support from GPU with XMX

Authors

Alexander Suslov, Alexander Kozlov,  Nikolay Lyalyushkin,  Nikita Savelyev, Souvikk Kundu, Andrey Anufriev, Pablo Munoz, Liubov Talamanova, Daniil Lyakhov, Yury Gorbachev, Nilesh Jain, Maxim Proshin, Evangelos Georganas

‍

Summary

This quarter marked a major shift towards efficiency in large-scale AI, driven by the unsustainable computational and memory costs of current architectures. The focus is now on making models dramatically faster and more hardware-friendly, especially for demanding long-context and multimodal tasks. 🚀 There is a growing adoption of dynamic, data-aware techniques like dynamic sparse attention and token pruning, which intelligently reduce computation by focusing only on the most critical information. Furthermore, optimization is increasingly tailored to new hardware through ultra-low precision; quantization is being pushed to the extreme, with native 1-bit (BitNet) inference and 4-bit (FP4) training becoming viable by aligning directly with new GPU capabilities.

A parallel trend is the creation of simple, readable frameworks like Nano-vLLM, whose lightweight design aims to lower the barrier to entry for developers and researchers.

‍

Highlights

• MMInference: Accelerating Pre-filling for Long-Context Visual Language Models via Modality-Aware Permutation Sparse Attention (https://arxiv.org/pdf/2502.02631). The authors introduce MMInference (Multimodality Million tokens Inference), a dynamic sparse attention method that accelerates the prefilling stage for long-context multi-modal inputs. The core ideas stemfrom analyzing the attention patterns specific to multi-modal inputs in VLMs: (1) Visual inputs exhibit strong temporal and spatial locality, leading to a unique sparse pattern the authors term the "Grid pattern".(2) Attention patterns differ significantly within a modality versus across modalities. The authors introduces the permutation-based method for offline searching the optimal sparse patterns for each head based on the input andoptimized kernels to compute attention much faster. MMInference speeds up the VLM pre-filling stage by up to 8.3x (at 1 million tokens) without losing accuracy and without needing any model retraining.The paper demonstrates maintained performance across various multi-modal benchmarks (like Video QA and Captioning) using state-of-the-art models (LongVila, LlavaVideo, VideoChat-Flash, Qwen2.5-VL). The code is available at https://aka.ms/MMInference.

• Beyond Text-Visual Attention: Exploiting Visual Cues for Effective Token Pruning in VLMs (https://arxiv.org/pdf/2412.01818). The authors introduce VisPruner, a training-free method for compressing visual token sequences in VLMs, dramatically reducing computational overhead. Unlike prior approaches that rely on text-visual attention scores - often biased and dispersed - VisPruner leverages visual cues directly from the visual encoder. They identify two key flaws in attention-based pruning: (1) attention shift: positional bias causes attention to favor lower image regions (tokens closer to the text in sequence); (2) attention dispersion: attention is spread too uniformly, making it hard to identify important tokens. VisPruner first selects a small set of important tokens using [CLS] attention (typically focused on foreground objects), then complements them with diverse tokens selected via similarity-based filtering to preserve background and contextual information. This visual-centric pruning strategy avoids reliance on language model internals and is compatible with fast attention mechanisms like FlashAttention. VisPruner outperforms finetuning-free baselines like FastV, SparseVLM, and VisionZip across 13 benchmarks—including high-resolution and video tasks—even when retaining as little as 5% of the original visual tokens. It achieves up to 95% FLOPs reduction and 75% latency reduction.

• OuroMamba: A Data-Free Quantization Framework for Vision Mamba Models (https://www.arxiv.org/pdf/2503.10959). The authors present OuroMamba, the first data-free post-training quantization (DFQ) method for vision Mamba-based models (VMMs). The authors identify two key challenges in enabling DFQ for VMMs, (1) VMM’s recurrent state transitions restricts capturing of long-range interactions and leads to semantically weak synthetic data,(2) VMM activation exhibit dynamic outlier variations across time-steps, rendering existing static PTQ techniques ineffective. To address these challenges, OuroMamba presents a two-stage framework: (1) OuroMamba-Gen to generate semantically rich and meaningful synthetic data. It applies constructive learning on patch level VMM features generated through neighborhood interactions in the latent state space, (2) OuroMamba-Quant to employ mixed-precision quantization with lightweight dynamic outlier detection during inference. In specific, the paper presents a threshold based outlier channel selection strategy for activation that gets updated every time-step. Extensive experiments across vision and generative tasks show that our data-free OuroMamba surpasses existing data-driven PTQ techniques, achieving state-of-the-art performance across diverse quantization settings. Additionally, the authors demonstrate the efficacy via implementation of efficient GPU kernels to achieve practical latency speedup of up to 2.36×.

• TailorKV: A Hybrid Framework for Long-Context Inference via Tailored KV Cache Optimization (https://arxiv.org/pdf/2505.19586). TailorKV is a novel framework designed to optimize the KV cache in LLMs for long-context inference, significantly reducing GPU memory usage and latency without sacrificing model performance.Instead of applying a one-size-fits-all compression strategy, TailorKV intelligently tailors compression based on the characteristics of each Transformer layer. The authors look at how each layer distributes its attention across tokens: (1) If a layer spreads attention broadly across many tokens, it’s considered to be dense. These layers are good candidates for quantization, because compressing them doesn’t significantly harm performance (usually shallow layers). (2) If a layer focuses attention on just a few tokens, it’s considered to be sparse. These layers are better suited for sparse retrieval, where only the most important tokens are kept in memory (deeper layers). To make this decision, they compute a score for each layer that reflects how concentrated or spread out the attention is. If the score is above a certain threshold, the layer is labeled quantization-friendly; otherwise, it’s considered sparsity-friendly. This classification is done offline, meaning it’s calculated once before inference, so it doesn’t affect runtime performance. TailorKV drastically reduces memory usage by quantizing 1 to 2 layers to 1-bit precision and loading only 1% to 3% of the tokens for the remaining layers.Maintains high accuracy across diverse tasks and datasets, outperforming state-of-the-art methods like SnapKV, Quest, and PQCache on LongBench. Code is available at: https://github.com/ydyhello/TailorKV.

• Log-Linear Attention (https://arxiv.org/pdf/2506.04761). The authors present Log-Linear Attention, a general framework that extends linear attention and state-space models by introducing a logarithmic growing memory structure for efficient long-context modeling. The paper identifies two key limitations in prior linear attention architectures: (1) the use of fixed-size hidden states restricts their ability to model multi-scale temporal dependencies, and (2) their performance degrades on long sequences due to the lack of hierarchical context aggregation.To address these challenges, Log-Linear Attention places a particular structure on the attention mask, enabling the compute cost to be log-linear and the memory cost to be logarithmic in sequence length (O(TlogT) training time,O(logT) inference time and memory). Conceptually, it uses a Fenwick tree–based scheme to hierarchically partition the input into power-of-two-sized segments. Each query attends to a logarithmic number of hidden states, summarizing increasingly coarse ranges of past tokens. This design emphasizes recent context with finer granularity, while efficiently compressing distant information.The framework is instantiated on top of two representative models: Mamba-2 and Gated DeltaNet, resulting in Log-Linear Mamba-2 and Log-Linear Gated DeltaNet. These variants inherit the expressive recurrence structures of their linear counterparts but benefit from logarithmic memory growth and sub-quadratic training algorithms via a custom chunk-wise parallel scan implementation in Triton.Experiments across language modeling, long-context retrieval, and in-context reasoning benchmarks show that Log-Linear Attention consistently improves long-range recall while achieving competitive or better throughput than FlashAttention-2 at longer sequence lengths (>8K). The code is available at https://github.com/HanGuo97/log-linear-attention.

• The Sparse Frontier: Sparse Attention Trade-offs in Transformer LLMs (https://arxiv.org/pdf/2504.17768). The authors introduce SparseFrontier, a systematic evaluation of dynamic sparse attention methods aimed at accelerating inference in LLMs for long-context inputs (up to 128K tokens). The core ideas stem from an extensive analysis of sparse attention trade-offs across different inference stages, model scales, and task types: (1) Sparse attention during decoding tolerates higher sparsity than during prefilling, particularly in larger models, due to differences in memory and compute bottlenecks.(2) No single sparse pattern is optimal across all tasks - retrieval, aggregation, and reasoning tasks each require different units of sparsification (e.g., blocks vs. tokens) and budget strategies. During prefilling, the best sparsification structure (e.g., blocks or verticals and slashes) is task-dependent, with uniform allocation across layers performing comparably to dynamic allocation. During decoding, page-level Quest excels by preserving the KV cache structure, avoiding the performance degradation associated with token pruning during generation. Their FLOPS analysis shows that for long context, large sparse models outperform smaller dense ones at the same compute cost. They also establish scaling laws predicting accuracy from model size, sequence length, and compression ratio.The code is available at: https://github.com/PiotrNawrot/sparse-frontier.

‍

Papers with notable results

Quantization

• SeedLM: Compressing LLM Weights into Seeds of Pseudo-Random Generators (https://arxiv.org/pdf/2410.10714). This paper introduces SeedLM, a novel data-free post-training compression method for Large Language Models (LLMs) that uses seeds of pseudo-random generators and some coefficients to recreate model weights. SeedLM aims to reduce memory access and leverage idle compute cycles during inference, effectively speeding up memory-bound tasks by trading compute for fewer memory accesses.The method generalizes well across diverse tasks, achieving better zero-shot accuracy retention at 4- and 3-bit compression compared to OmniQuant, AWQ and QuIP#. Additionally, FPGA-based tests demonstrate close to 4x speedup for memory-bound tasks such as generation for 4bit per value over an FP16 Llama baseline.

• LoTA-QAF: Lossless Ternary Adaptation for Quantization-Aware Fine-Tuning (https://arxiv.org/pdf/2505.18724). LoTA-QAF is a quantization-aware fine-tuning method for LLMs designed for efficient edge deployment. Its key innovation is a ternary adaptation approach, where ternary adapter matrices can only increment, decrement, or leave unchanged each quantized integer weight (+1, −1, or 0) within the quantization grid during fine-tuning. This tightly restricts the amount each quantized value can change, ensuring the adapters do not make large modifications to weights. The method enables lossless merging of adaptation into the quantized model, preserving computational efficiency and model performance with no quantization-induced accuracy loss at merge. The method uses a novel ternary signed gradient descent (t-SignSGD) optimizer to efficiently update these highly constrained ternary weights. Evaluated on the Llama-3.1/3.3 and Qwen-2.5 families, LoTA-QAF consistently outperforms previous quantization-aware fine-tuning methods such as QA-LoRA, especially at very low bit-widths (2-bit and 3-bit quantization), recovering up to 5.14% more accuracy on MMLU compared to LoRA under 2-bit quantization, while also being 1.7x–2x faster at inference after merging. Task-specific fine-tuning shows LoTA-QAF improves on other quantization-aware methods, though it slightly lags behind full-precision LoRA in those scenarios.The code is available at: https://github.com/KingdalfGoodman/LoTA-QAF.

• SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-bit Training (https://arxiv.org/abs/2505.11594). The authors introduce SageAttention3, a novel FP4 micro-scaling quantization technique for Transformer attention designed to achieve a 5x speedup in inference on NVIDIA GPUs and an 8-bit novel training approach that preserves model accuracy during finetuning while reducing memory demands. The method applies FP4 quantization to the two main attention matrix multiplications, using a microscaling strategy with a group size of 16 elements per scale factor. This fine granularity limits the impact of outlier values that can otherwise cause significant quantization error. To address issues with quantizing the attention map, the authors propose a two-level quantization scheme. First, each row of attention map is scaled into the range[0, 448 × 6], which ensures the FP8 scaling factor (required by hardware) fully utilizes its representation range. Then, FP4 quantization is applied at the block level. This two-step process significantly reduces quantization error compared to direct quantization. Empirical results show that SageAttention3 delivers substantial inference speedups with minimal quality loss on language, image, and video generation benchmarks. The code is available at: https://github.com/thu-ml/SageAttention.‍

• MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods (https://arxiv.org/abs/2501.13484). This paper tackles the challenge of post-training quantization for Mamba architectures. Standard quantization techniques adapted from large language models result in substantial accuracy loss when applied to Mamba models, largely due to extreme outliers and inconsistent variances across different channels in weights and activations. To address these issues, the authors propose MambaQuant, introducing two variance alignment techniques: KLT-Enhanced and Smooth-Fused rotations. These methods effectively equalize channel variances, resulting in more uniform data distributions before quantization. Experimental results show that MambaQuant enables Mamba models to be quantized to 8 bits for both weights and activations with less than 1% loss in accuracy, markedly surpassing previous approaches on both vision and language tasks.
• APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (https://arxiv.org/pdf/2504.02508). APHQ-ViT is a PTQ method designed to address the challenges of quantizing Vision Transformers, particularly under ultra-low bit settings. Traditional reconstruction-based PTQ methods, effective for Convolutional Neural Networks, often fail with ViTs due to inaccurate estimation of output importance and significant accuracy degradation when quantizing post-GELU activations. To overcome these issues, APHQ-ViT introduces an improved Average Perturbation Hessian (APH) loss for better importance estimation. Additionally, it proposes an MLP Reconstruction technique that replaces the GELU activation function with ReLU in the MLP modules and reconstructs them using the APH loss on a small unlabeled calibration set. Experiments demonstrate that APHQ-ViT, utilizing linear quantizers, outperforms existing PTQ methods by substantial margins in 3-bit and 4-bit quantization across various vision tasks. The source code for APHQ-ViT is available at https://github.com/GoatWu/APHQ-ViT.
• DL-QAT: Weight-Decomposed Low-Rank Quantization-Aware Training for Large Language Models (https://arxiv.org/abs/2504.09223). DL-QAT is a quantization-aware training (QAT) technique for LLMs that achieves high efficiency by updating less than 1% of parameters. It introduces group-specific quantization magnitudes and uses LoRA-based low-rank adaptation within the quantization space. Tested on LLaMA and LLaMA2, DL-QAT outperforms previous state-of-the-art methods—including QA-LoRA and LLM-QAT - by up to 4.2% on MMLU benchmarks for 3-bit models, while greatly reducing memory and training costs.
• BitNet b1.58 2B4T Technical Report (https://arxiv.org/abs/2504.09223). Microsoft Research released the weights for BitNet b1.58 2B4T, the first open-source, native 1-bit Large Language Model (LLM) at the 2-billion parameter scale and inference framework bitnet.cpp. The new 2B model demonstrates performance comparable to the Qwen 2.5 1.5B on benchmarks, while operating at 2x the speed and consuming 12x less energy.
• Quartet: Native FP4 Training Can Be Optimal for Large Language Models (https://arxiv.org/pdf/2505.14669). The authors introduced a new method "Quarter" for the stable 4-bit floating-point (FP4) training. There is specifically designed for the native FP4 hardware in NVIDIA's new Blackwell GPUs and achieved a nearly 2x speedup on the most intensive training computations compared to 8-bit techniques, all while maintaining "near-lossless" accuracy. The method outlines to perform a forward pass that minimizes MSE (based on QuEST) together with a backwardpass that is unbiased (based on Stochastic Rounding). The code of extremely efficient GPU-aware implementation https://github.com/IST-DASLab/Quartet
• InfiJanice: Joint Analysis and In-situ Correction Engine for Quantization-Induced Math Degradation in Large Language Models (https://arxiv.org/pdf/2505.11574). The authours investigates how quantization significantly harms the mathematical reasoning abilities of LLMs. The study reveals that quantization can degrade reasoning accuracy by up to 69.81% on complex benchmarks, with smaller models being more severely affected. Authors developed an automated pipeline to analyze and categorize the specific errors introduced by quantization. Based on these findings, they created a compact, targeted dataset named "Silver Bullet." The most notable result is that fine-tuning a quantized model on as few as 332 of these curated examples for just 3–5 minutes on a single GPU is sufficient to restore its mathematical reasoning accuracy to the level of the original, full-precision model.

‍

Pruning/Sparsity

• Token Sequence Compression for Efficient Multimodal Computing (https://arxiv.org/pdf/2504.17892). The authors introduce a training-free method for compressing visual token sequences in visual language models (VLMs), significantly reducing computational costs. Instead of relying on attention-based “saliency”—a measure of how much attention a model gives to each token—they use simple clustering to group similar visual tokens and aggregate them. Their “Cluster & Aggregate” approach outperforms prior finetuning-free methods like VisionZip and SparseVLM across 8+ benchmarks, even when retaining as little as 11% of the original tokens. Surprisingly, random and spatial sampling also perform competitively, revealing high redundancy in visual encodings.

• Beyond 2:4: exploring V:N:M sparsity for efficient transformer inference on GPUs (https://arxiv.org/abs/2410.16135). This paper introduces and systematically studies V:N:M sparsity as a more efficient and flexible alternative to the industry-standard 2:4 sparsity for accelerating Transformer inference on GPUs. In the V:N:M approach, weight matrices are divided into V×M blocks; within each block, most columns are pruned, and 2:4 sparsity is then applied to the remaining columns. This scheme enables significantly higher and more adaptable sparsity ratios, while remaining compatible with existing GPU sparse tensor core acceleration. The authors propose a comprehensive framework for creating V:N:M-sparse Transformers: it features a heuristic method for selecting V and M values to optimize the accuracy-speedup trade-off, a V:N:M-specific channel permutation method for improving accuracy in low-budget training scenarios, and a three-stage LoRA training process for memory-efficient fine-tuning. Experimental results show that V:N:M-sparse Transformers can achieve much higher sparsity levels - such as 75% parameter reduction, while maintaining nearly lossless accuracy on downstream tasks, and outperform 2:4 sparsity in both speed and flexibility.

• TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (https://arxiv.org/pdf/2503.18278v2). The authors introduce a training-free, optimization-based framework for reducing visual token redundancy in VLMs. Visual tokens often dominate the input sequence—up to 95% in some models. TopV addresses this by pruning unimportant visual tokens once during the prefilling stage, before decoding begins.Instead of relying on attention scores, TopV estimates the importance of each visual token by solving an optimal transport problem. In this setup: (1) Source tokens are the input visual tokens entering a specific transformer layer. (2) Target tokens are the output visual tokens after that layer has processed the input—specifically, the output after the Post-LN sub-layer. TopV calculates how much each input token contributes to the output using the Sinkhorn algorithm, guided by a cost function that considers: (1) How similar the tokens are in content (feature similarity), (2) How close they are in the image (spatial proximity), (3) How central they are in the image (centrality). To prevent visual collapse—especially in detail-sensitive tasks like OCR and captioning—TopV includes a lightweight recovery step. From the discarded tokens, TopV uniformly samples a subset at regular intervals (e.g., every 4th or 6th token) and reinserts them into the token sequence alongside the top-k tokens, ensuring spatial diversity and semantic coverage without significant overhead.TopV performs pruning once after the prompt and image are processed. The pruned visual token set remains fixed throughout decoding, enabling efficient and consistent inference.
• SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (https://arxiv.org/pdf/2410.04417). SparseVLM introduces a lightweight, training-free framework for visual token sparsification in vision-language models (VLMs). Unlike text-agnostic approaches, it leverages cross-attention to identify text-relevant visual tokens (“raters”) and adaptively prunes others based on the rank of the attention matrix. Crucially, SparseVLM doesn’t discard all pruned tokens—instead, it recycles the most informative ones (those with high attention relevance scores). These are grouped using a density peak clustering algorithm, and each cluster is compressed into a single representative token. The reconstructed tokens are then reinserted into the model, replacing the larger set of pruned tokens with a compact, information-rich representation. Applied to LLaVA, SparseVLM achieves a 4.5× compression rate with only a 0.9% accuracy drop, reduces CUDA latency by 37%, and saves 67% memory. The code is available at https://github.com/Gumpest/SparseVLMs.

‍

Other

• Hogwild! Inference: Parallel LLM Generation via Concurrent Attention (https://arxiv.org/pdf/2504.06261). Hogwild! Inference introduces a novel paradigm for parallel inference for reasoning tasks that departs significantly from prior structured approaches by enabling dynamic, parallel collaboration. The method runs multiple LLM "workers" concurrently, allowing them to interact in real-time through a shared Key-Value (KV) cache. This shared workspace lets workers see each other's progress as it happens, fostering emergent teamwork without rigid, pre-planned coordination. A key innovation is the efficient use of Rotary Position Embeddings (RoPE) to synchronize the workers' views of the shared cache with minimal computational overhead. Empirical results show significant wall-clock speedups—up to 3.6x with 4 workers—on complex reasoning tasks. This is achieved "out of the box" on existing models without requiring fine-tuning and can be stacked with another optimization methods such as speculative decoding. The technique fundamentally improves the speed-cost-quality trade-off for inference, shifting the paradigm from sequential "chains of thought" to collaborative "teams of thought". The code is available at https://github.com/eqimp/hogwild_llm.

• Parallel Scaling Law for Language Models (https://arxiv.org/pdf/2505.10475). The authors introduce a novel "parallel" scaling method for LLMs (ParScale), distinct from traditional parameter (Dense, MoE) or inference-time (CoT) scaling. The technique processes a single input through 'P' parallel streams, each modified by a unique, learnable prefix vector. These streams are run concurrently on the same base model, and their outputs are intelligently aggregated by a small network. This method yields a quality improvement equivalent to increasing the model size by a factor of log(P), without actually expanding the core parameter count. For example, 8 parallel streams can match the performance of a model three times larger. ParScale is highly efficient for local inference, where memory bandwidth is the main bottleneck. Compared to direct parameter scaling for similar quality, it can require up to 22x less additional RAM and add 6x less latency. The approach can be applied for pretrained models, even with frozen weight, fine-tune only perscale components. The code is available at https://github.com/QwenLM/ParScale.

• Packing Input Frame Context in Next-Frame Prediction Models for Video Generation (https://arxiv.org/pdf/2504.12626). FramePack is a framework for next-frame prediction video generators that enables long-duration video synthesis with a constant computational cost (O(1)), regardless of length. It circumvents growing context windows by maintaining a fixed-size token buffer and codes input frames as shown in the figure below. To maintain temporal consistency and mitigate error accumulation, the system employs a bi-directional sampling scheme, alternating between forward and backward prediction passes. This efficiency allows a 13-billion parameter model to generate over 1800 frames (1 minute @ 30 fps) on a GPU with only 6GB of VRAM. The O(1) complexity in memory and latency makes FramePack a practical solution for generating minute-long videos on consumer hardware, with generation speeds of ~1.5 seconds per frame reported on an RTX 4090.The code is available at https://github.com/lllyasviel/FramePack.

• MoDM: Efficient Serving for Image Generation via Mixture-of-Diffusion Models (https://arxiv.org/pdf/2503.11972). Diffusion-based text-to-image generation models trade latency for quality: small models are fast but generate lower quality images, while large models produce better images but are slow. This paper presents MoDM, a novel caching-based serving system for diffusion models that dynamically balances latency and quality through a mixture of diffusion models.Unlike prior approaches that rely on model-specific internal features, MoDM caches final images, allowing seamless retrieval and reuse across multiple diffusion model families.This design enables adaptive serving by dynamically balancing latency and image quality: using smaller models for cache-hit requests to reduce latency while reserving larger models for cache-miss requests to maintain quality. Small model image quality is preserved using retrieved cached images. MoDM has a global monitor that optimally allocates GPU resources and balances inference workload, ensuring high throughput while meeting Service-Level Objectives (SLOs) under varying request rates. Extensive evaluations show that MoDM significantly reduces an average serving time by 2.5× while retaining image quality, making it a practical solution for scalable and resource-efficient model deployment.
• ‍Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding (https://arxiv.org/abs/2505.22618). Fast-dLLM is a training-free method to accelerate diffusion-based large language models by introducing a block-wise KV Cache and confidence-aware parallel decoding. The block-wise KV Cache reuses more than 90% of attention activations with bidirectional (prefix and suffix) caching, delivering throughput improvements ranging from 8.1x to 27.6x while keeping accuracy loss under 2%. Confidence-aware parallel decoding selectively generates tokens that exceed a set confidence threshold (like 0.9), achieving up to 13.3x speedup and preserving output coherence thanks to theoretical guarantees. Experimentally, Fast-dLLM achieves up to 27.6× end-to-end speedup on 1024-token sequences (e.g., LLaDA, 8-shot) and keeps accuracy within 2% of the baseline across major reasoning and code benchmarks including GSM8K, MATH, HumanEval, and MBPP.
• SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation (https://arxiv.org/pdf/2503.09641). SANA-Sprint is a highly efficient text-to-image diffusion model designed for ultra-fast generation. Its core innovation is a hybrid distillation framework that combines continuous-time consistency models (sCM) with latent adversarial diffusion distillation (LADD). This approach drastically reduces inference requirements from over 20 steps to just 1-4. Key performance benchmarks establish a new state-of-the-art. In a single step, SANA-Sprint generates a 1024x1024 image with FID of 7.59. This is achieved with a latency of just 0.1 seconds on an NVIDIA H100 GPU and 0.31 seconds on a consumer RTX 4090. This makes it approximately 10 times faster than its competitor, FLUX-schnell, while also delivering higher image quality.The code is available at https://github.com/NVlabs/Sana.

‍

Software

• FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (https://arxiv.org/abs/2412.07752). FlashRNN extends traditional RNNs - such as LSTMs and GRUs - by introducing a parallelization scheme where the hidden state is divided into multiple smaller blocks, allowing for parallel computation similar to the head-wise processing in Transformers. The authors develop and open-source custom fused CUDA and Triton kernels that leverage the GPU memory hierarchy efficiently for both forward and backward passes, together with an automatic hardware-aware optimization framework. This approach achieves up to 50x speedup over vanilla PyTorch implementations, making RNNs competitive with Transformer-like models on modern GPUs. The code is available at: https://github.com/NX-AI/flashrnn.
• Nano-vLLM (https://github.com/GeeeekExplorer/nano-vllm). A lightweight vLLM implementation built from scratch.Key Features: (1)🚀 Fast offline inference - Comparable inference speeds to vLLM (2)📖 Readable codebase - Clean implementation in ~ 1,200 lines of Python code (3)⚡ Optimization Suite - Prefix caching, Tensor Parallelism, Torch compilation, CUDA graph, etc.
• NeMo-Inspector: A Visualization Tool for LLM Generation Analysis (https://arxiv.org/pdf/2505.00903). The authors introduce NeMo-Inspector, an open-source tool designed to simplify the analysis of synthetic datasets with integrated inference capabilities.
• FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving (https://arxiv.org/pdf/2501.01005). The authors present FlashInfer: a customizable and efficient attention engine for LLM serving. FlashInfer tackles KV-cache storage heterogeneityusing block-sparse format and composable formats to optimize memory access and reduce redundancy, supports JIT compilation and load-balanced scheduling algorithm adjusts to dynamism of user requests while maintaining compatibility with CUDAGraph which requires static configuration. FlashInfer achieve29-69% inter-token-latency reduction compared to Triton, 28-30% latency reduction for long-context inference, and 13-17% speedup for LLM serving with parallel generation.The code is available at https://github.com/flashinfer-ai/flashinfer.

Authors: Ivan Novoselov, Alexandra Sidorova, Vladislav Golubev, Dmitry Gorokhov

Introduction

Deep learning (DL) has become a powerful tool for addressing challenges in various domains like computer vision, generative AI, and natural language processing. Industrial applications of deep learning often require performing inference in resource-constrained environments or in real time. That’s why it’s essential to optimize inference of DL models for particular use cases, such as low-latency, high-throughput or low-memory environments. Thankfully, there are several frameworks designed to make this easier, and OpenVINO stands out as a powerful tool for achieving these goals.

OpenVINO is an open-source toolkit for optimization and deployment of DL models. It demonstrates top-tier performance across a variety of hardware including CPU (x64, ARM), AI accelerators (Intel NPU) and Intel GPUs. OpenVINO supports models from popular AI frameworks and delivers out-of-the box performance improvements for diverse applications (you are welcome to explore demo notebooks). With ongoing development and a rapidly growing community, OpenVINO continues to evolve as a versatile solution for high-performance AI deployments.

The primary objective of OpenVINO is to maximize performance for a given DL model. To do that, OpenVINO applies a set of hardware-dependent optimizations The optimizations are typically performed by replacing a target group of operations with a custom operation that can be executed more efficiently. In the standard approach, these custom operations are executed using handcrafted implementations. This approach is highly effective when optimizing a few patterns of operations. On the other hand, it lacks scalability and thus requires too much effort when dozens of similar patterns should be supported.

To address this limitation and build a more flexible optimization engine, OpenVINO introduced Snippets, an integrated Just-In-Time (JIT) compiler for computational graphs. Snippets provide a flexible and scalable approach for operation fusions and enablement. The graph compiler automatically identifies subgraphs of operations that can benefit from fusion and combines them into a single node, referred to as “Subgraph”. Snippets then apply a series of optimizing transformations to the subgraph and JIT compile an executable that efficiently performs the computations defined by the subgraph.

One of the most common examples of such subgraphs is Scaled Dot-Product Attention (SDPA) pattern. SDPA it is a cornerstone of transformer-based architectures which dominate most of the state-of-the-art models. There are numerous SDPA pattern flavours and variations dictated by model-specific adjustments or optimizations. Thanks to compiler-based design, Snippets support most of these configurations. Fig. 1 illustrates the general structure of the SDPA pattern supported by Snippets, highlighting its adaptability to different model requirements:

Note that SDPA has quadratic time and memory complexity with respect to sequence length. It means that by fusing SDPA-like patterns, Snippets significantly reduce memory consumption and accelerate transformer models, especially for large sequence lengths.

Snippets effectively optimized SDPA patterns but had a key limitation: they did not support dynamic shapes. In other words, input shapes must be known at the model compilation stage and can’t be changed in runtime. This limitation reduced the applicability of Snippets to many real-world scenarios where input shapes are not known in advance. While it is technically possible to JIT-compile a new binary for each unique set of input shapes, this approach introduces significant recompilation overheads, often negating the performance gains from SDPA fusion.

Fortunately, this static-shape limitation is not inherent to Snippets design. They can be modified to support dynamic shapes internally and generate shape-agnostic binaries. In this post, we discuss Snippets architecture and the challenges we faced during this dynamism enablement.

Architecture

The first step of the Snippets pipeline is called Tokenization. It is applied to an ov::Model, which represents OpenVINO Intermediate Representation (IR). It’s a standard IR in the OV Runtime you can read more about it here or here. The purpose of this stage is to identify parts of the initial model that can be lowered by Snippets efficiently. We are not going to discuss Tokenization in detail because this article is mostly focused on the dynamism implementation. A more in-depth description of the Tokenization process can be found in the Snippets design guide. The key takeaway for us here is that the subsequent lowering is performed on a part of the initial ov::Model. We will call this part Subgraph, and the Subgraph at first is also represented as an ov::Model.

Now let’s have a look at the lowering pipeline, its schematic representation is shown on Fig.2a. As can be seen from the picture, the lowering process consists of three main phases: Data Flow Optimizations, Control Flow Optimizations and Binary Code Generation. Let’s briefly discuss each of them.

Lowering Pipeline

The first stage is the Data Flow optimizations. As we mentioned above, this stage’s input is a part of the initial model represented as an ov::Model. This representation is very convenient for high-level transformations such as opset conversion, operations’ fusion/decomposition and precision propagation. Here are some examples of the transformations performed on this stage:

1. ConvertPowerToPowerStatic — operation Power with scalar exponent input is converted to PowerStatic operation from the Snippets opset. The PowerStatic ops then use the values of the exponents to produce more optimal code.
2. FuseTransposeBrgemm — Transpose operations that can be executed in-place with Brgemm blocks are fused into the Brgemm operations.
3. PrecisionPropagation pass automatically inserts Converts operations between the operations that don’t natively support the desired execution precision.

The next stage of the lowering process is Control Flow optimizations (or simply CFOs). Note that the ov::Model is designed to primarily describe data flow, so is not very convenient for CFOs. Therefore, we had to develop our own IR (called Linear IR or simply LIR) that explicitly represents both control and data flows, you can read more about LIR here. So the ov::Model IR is converted to LIR before the start of the CFOs.

As you can see from the Fig.2a, the Control Flow optimization pipeline could roughly be divided into three main blocks. The first one is called Loop Generation and Optimization. This block includes all loop-related optimizations such as automatic generation of loops based on the input tensors’ dimensions, loop fusion and blocking loops generation.

The second block of Control Flow optimizations is called Utility Ops Insertion. We need this block of transformations here to insert utility operations that depend on loop control structures, specifically on their entry and exit points locations. For example, operations like Load, Store, MemoryBuffer, LoopBegin and LoopEnd are inserted during this stage.

The last step of CFO is the Memory Usage Optimizations block. These transformations determine required sizes of internal memory buffers, and analyze how much of that memory can be reused. A graph coloring algorithm is employed to minimize memory consumption.

Now all Control Flow optimizations are performed, and we are ready to proceed to the next stage of the lowering pipeline — Binary Code Generation (BCG). As one can see from Fig.2a, this stage consists of three substages. The first one is Register Assignment. We use a fairly standard approach here: calculate live intervals first and use the linear scan algorithm to assign abstract registers that are later mapped to physical ones.

The next BSG substage is Loop Expansion. To better understand its purpose, let’s switch gears for a second and think about loops in general. Sometimes it’s necessary to process the first or the last iteration of a loop in a special way. For example, to initialize a variable or to process blocking loops’ tails. The Loop Expansion pass unrolls these special iterations (usually the first or the last one) and explicitly injects them into the IR. This is needed to facilitate subsequent code emission.

The final step of the BCG stage is Code Emission. At this stage, every operation in our IR is mapped to a binary code emitter, which is then used to produce a piece of executable code. As a result, we produce an executable that performs calculations described by the initial input ov::Model.

Dynamic Shapes Support

Note that some stages of the lowering pipeline are inherently shape-sensitive, i.e. they rely on specific values of input shapes to perform optimizations. These stages are schematically depicted on the Fig. 2b.

As can be seen from the picture, shapes are used to determine loops’ work amounts and pointer increments that should be performed on every iteration. These parameters are later baked into the executable during Code Emission. Another example is Memory Usage Optimizations, since input shapes are needed to calculate memory consumption. Loop Expansion also relies on input shapes, since it needs to understand if tail processing is required for a particular loop. Note also that Snippets use compute primitives from third-party libraries, BRGEMM block from OneDNN for example. These primitives should as well be compiled with appropriate parameters that are also shape-sensitive.

One way to address these challenges is to rerun the lowering pipeline for every new set of input shapes, and to employ caching to avoid processing the same shapes twice. However, preliminary analysis indicated that this approach is too slow. Since this re-lowering needs to be performed in runtime, the performance benefit provided by Snippets is essentially eliminated by the recompilation overheads.

The performed experiments thus indicate that we can afford to run the whole lowering pipeline only once during the model compilation stage. Only some minor adjustments can be made at runtime. In other words, we need to remove all shape-sensitive logic from the lowering pipeline and perform the compilation without it. The remaining shape-sensitive transformations should be performed at runtime. Of course, we would also need to share this runtime context with the compiled shape-agnostic kernel. The idea behind this approach is schematically represented on Fig. 2c.

As one can see from the picture, all the shape-sensitive transformations are now performed by a new entity called Runtime Configurator. It’s probably easier to understand its purpose in some examples.

Imagine that we need to perform a unary operation get_result(X) on an input tensor X — for example, apply an activation function. To do this, we need to load some input data from memory into registers, perform the necessary computations and write the results back to memory. Of course this read-compute-write sequence should be done in a loop since we need to process the entire input tensor. These steps are described in more detail in Fig. 3 using pseudocode. Fig. 3a corresponds to a static kernel while Fig. 3b represents a dynamic one.

Let’s consider the static kernel as a starting point. As the first step, we need to load pointers to input and output memory blobs to general-purpose registers (or simply GPRs) denoted G_IN and G_OUT on the picture. Then we initialize another GPR that stores the loop work amount (G_WA). Note that the loop is used to traverse the input tensor, so the loop’s work amount is fixed because the tensor’s dimensions are also known at the BCG stage. The next six steps in the picture (3 to 8) are in the loop’s body.

In step 3, we load input data into a vector register V_0, note that the appropriate pointer is already loaded to G_IN, and offset_in is fixed because the input tensor is static. Next, we apply our get_result function to the data in V_0 and place the result in a spare vector register V_1. Now we need to store V_1 back to memory, which is done on step 5. Note that offset_out is also known in the static case. This brings us almost to the end of the loop’s body, and the last few things we need to do are to increment data pointers (step 6), decrement loop counter (step 7), and jump to the beginning of the body, if needed (step 8).

Finally, we need to reset data pointers to their initial values after the loop is finished, which is done using finalization offsets on step 9. Note that this step could be omitted in our simplified example, but it’s often needed for more complicated use cases, such as when the data pointers are used by subsequent loops.

Now that we understand the static kernel, let’s consider the dynamic one, which is shown in Fig. 3b. Unsurprisingly, the dynamic kernel performs essentially the same steps as the static one, but with additional overhead due to loading shape-dependent parameters from the extended runtime arguments. Take step 1 as an example, we need to load not only memory pointers (to G_IN and G_OUT), but also a pointer to the runtime arguments prepared by the runtime configurator (to G_ARG).

Next, we need to load a pointer to the appropriate loop descriptor (a structure that stores loops’ parameters) to a temporary register G_TMP, and only then we can initialize the loop’s work amount register G_WA (step 2). Similarly, in order to load data to V_0, we need to load a runtime-calculated offset from the runtime arguments in step 3. The computations in step 4 are the same as in the static case, since they don’t depend on the input shapes. Storing the results to memory (step 5) requires reading a dynamic offset from the runtime arguments again. Next, we need to shift the data pointers, and again we have to load the increments from the corresponding loop descriptor in G_ARG because they are also shape-dependent, as the input tensor can be strided. The following two steps 7 and 8 are the same as in the static case, but the finalization offsets are also dynamic, so we have to load them from G_ARG yet again.

As one can see from Fig. 3, dynamic kernels incorporate additional overhead due to reading the extended runtime parameters provided by the runtime configurator. However, this overhead could be acceptable as long as the input tensor is large enough (Load/Store operations would take much longer than reading runtime arguments from L1) and the amount of computation is sufficient (get_results is much larger than the overhead). Let’s consider the performance of this design in the Results section to see if these conditions are met in practical use cases.

Results

We selected three platforms to evaluate the dynamic pipeline’s performance in Snippets. These platforms represent different market segments: the Intel Core machine is designed for high-performance user and professional tasks. While the Intel Xeon is a good example of enterprise-level hardware often used in data centers and cloud computing applications. The information about the platforms is described in the table below:

As discussed in the Introduction, Snippets support various SDPA-like patterns which form the backbone of Transformer models. These models often work with input data of arbitrary size (for example, sequence length in NLP). Thus, dynamic shapes support in Snippets can efficiently accelerate many models based on Transformer architecture with dynamic inputs.

We selected 43 different Transformer-models from HuggingFace to measure how the enablement of dynamic pipeline in Snippets affects performance. The models were downloaded and converted to OpenVINO IRs using Optimum Intel. These models represent different domains and were designed to solve various tasks in natural language processing, text-to-image image generation and speech recognition (see full model list at the end of the article). What unifies all these models is that they all contain SDPA subgraph and thus can be accelerated by Snippets.

Let’s take a closer look at the selected models. The 37 models of them solve different tasks in natural language processing. Their performance was evaluated using a list of 2000 text sequences with different lengths, which also mimics the real-word scenario. The total processing time of all the sequences were measured in every experiment. Note that the text sequences were converted to model inputs using model-specific tokenizers prior to the benchmarking. The lengths’ distribution of the tokenized sequences is shown on Fig. 4. As can be seen from the picture, the distribution is close to normal with the mean length of 31 tokens.

The other 6 models of the selected model scope solve tasks in text-to-image image generation (Stable Diffusion) and speech recognition (Whisper). These models decompose into several smaller models after export to OpenVINO representation using Optimum Intel. Stable Diffusion topology is decomposed into Encoder, Diffuser and Decoder. The most interesting model here is Diffuser because it’s the one responsible for denoising of the latent image representation. This generation stage is repeated several times, so it is the most computationally intensive, which mostly effects on the generation time of the image. Whisper is also decomposed into Encoder and Decoder, which also contain SDPA patterns. The Encoder encodes the spectrogram from the feature extractor to form a sequence of encoder hidden states. Then, the decoder autoregressively predicts text tokens, conditional on both the previous tokens and the encoder hidden states. Currently, Snippets support efficient execution of SDPA only in Whisper Encoder, while Decoder is a subject for future support. To evaluate the inference performance of Stable Diffusion and Whisper models, we collect generation time of image/speech using LLM Benchmark from openvino.genai. This script provides a unified approach to estimate performance for GenAI workloads in OpenVINO.

Performance Improvements

Note that the main goal of these experiments is to estimate the impact of Snippets on the performance of the dynamic pipeline. To do that, we performed two series of experiments for every model. The first version of experiments is with disabled Snippets tokenization. In this case, all operations from the SDPA pattern are performed on the CPU plugin side as stand-alone operations. The second variant of experiments — with enabled Snippets tokenization. The relative difference between numbers collected on these two series of experiments is our performance metric — speedup, the higher the better. Firstly, let’s take a closer look at the resulting speedups for the BERT models which are depicted on Fig. 5.

The speedups on RPL range from 3 to 18%, while on average the models are accelerated by 7%. The ARL-S speedups are somewhat higher and reach 20–25% for some models, the average acceleration factor is around 9%. The most affected platform is SPR, it has the highest average speedup of 15 %.

One can easily see from these numbers that both average and maximum speedups depend on the platform. To understand the reason for this variation, we should recall that the main optimizations delivered by Snippets are vertical fusion and tiling. These optimizations improve cache locality and reduce the memory access overheads. Note SPR has the largest caches among the examined platforms. It also uses BF16 precision that takes two times less space per data element compared to F32 used on ARL-S and RPL. Finally, SPR has AMX ISA extension that allows it to perform matrix multiplications much faster. As a result, SDPA execution was more memory bound on SPR, so this platform benefited the most from the Snippets enablement. At the same time, the model speedups on ARL-S and RPL are almost on the same level. These platforms use FP32 inference precision while SPR uses BF16, and they have less cache size than SPR.

Now, let’s consider Stable Diffusion and Whisper topologies and compare their speedups with some of BERT-like models. As can be seen from the Fig. 6, the most accelerated Stable Diffusion topology is StableDiffusion-3-medium — almost 33% on ARL-S and 40% on SPR. The most accelerated model in this Stable Diffusion pipeline is Diffuser. This model has made a great contribution to speeding up the entire image generation time. The reason the Diffuser benefits more from Snippets enablement is that they use larger sequence lengths and embedding sizes. It means that their attention blocks process more data and are more memory constrained compared to BERT-like models. As a result, the Diffuser models in Stable Diffusion benefit more from the increased cache locality provided by Snippets. This effect is more pronounced on the SPR than on the ARL-S and RPL for the reasons discussed above (cache sizes, BF16, AMX).

The second most accelerated model is whisper-large-v3–30% on SPR. This model has more parameters than base and tiny models and process more Mel spectrogram frequency bins than they. It means that Encoder of whisper-large-v3 attention blocks processes more data, like Diffuser part of Stable Diffusion topologies. By the same reasons, whisper-large-v3 benefits more (increased cache locality provided by Snippets).

Memory Consumption Improvements

Another important improvement from Snippets using is reduction of memory consumption. Snippets use vertical fusion and various optimizations from Memory Usage Optimizations block (see the paragraph “Lowering Pipeline” in “Architecture” above for more details). Due to this fact, Subgraphs tokenized by Snippets consumes less memory than the same operations performed as stand-alone in CPU Plugin.

Let’s take a look at the Fig. 7 where we can see improvements in image generation memory consumption using Stable Diffusion pipelines from Snippets usage. As discussed above, the attention blocks in the Diffuser models from these pipelines process more data and consume more memory. Because of that, the greatest impact on memory consumption from using Snippets is seen on Stable Diffusion pipelines. For example, memory consumption of image generation is reduced by 25–50% on RPL and ARL-S platforms with FP32 inference precision and by 15–30% on SPR with BF16 inference precision.

Thus, one of the major improvements from using Snippets is memory consumption reduction. It allows extending the range of platforms which are capable to infer such memory-intensive models as Stable Diffusion.

Conclusion

Snippets is a JIT compiler used by OpenVINO to optimize performance-critical subgraphs. We briefly discussed Snippets’ lowering pipeline and the modifications made to enable dynamism support. After these changes, Snippets generate shape-agnostic kernels that can be used for various input shapes without recompilation.

This design was tested on realistic use cases across several platforms. As a result, we demonstrate that Snippets can accelerate BERT-like models by up to 25%, Stable Diffusion and Whisper pipelines up to 40%. Additionally, Snippets can significantly reduce memory consumption by several tens of percent. Notably, these improvements result from more optimal hardware utilization, so the models’ accuracy remains unaffected.

References

Model list can be found here: https://gist.github.com/a-sidorova/2c2aa584922389138e5ccbad0e0c773b

Intel Xeon Gold 6430L: https://www.intel.com/content/www/us/en/products/sku/231737/intel-xeon-gold-6430-processor-60m-cache-2-10-ghz/specifications.html

Intel Core i9-13900K: https://www.intel.com/content/www/us/en/products/sku/230496/intel-core-i913900k-processor-36m-cache-up-to-5-80-ghz/specifications.html

Intel Core Ultra 7 265K: https://www.intel.com/content/www/us/en/products/sku/241063/intel-core-ultra-7-processor-265k-30m-cache-up-to-5-50-ghz/specifications.html

Notices & Disclaimers

Performance varies by use, configuration, and other factors. Learn more on the Performance Index site.

No product or component can be absolutely secure. Your costs and results may vary. Intel technologies may require enabled hardware, software or service activation.

© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries.

‍

Authors

Alexander Kozlov,  Nikolay Lyalyushkin,  Nikita Savelyev, Souvikk Kundu, Andrey Anufriev, Pablo Munoz, Alexander Suslov, Liubov Talamanova, Daniil Lyakhov, Yury Gorbachev, Nilesh Jain, Maxim Proshin, Evangelos Georganas

‍

Summary

This quarter we noticed a significant effort and progress on optimizing LLMs for long-context tasks. The current trend is that each and every LLM is published with the extended (usually interpolated) context which is usually 128K and above. The idea is to naturally process large amount of data within the model instead of preprocess it the way RAG systems do it. It inevitably increases computational complexity specifically of ScaledDotProductAttention operation which gets dominant on long contexts. Thus, many works devoted to the optimization of rather prefill with special computation patterns (A-shape, Tri-shape, XAttention) or using Sparse Attention at the decoding stage.

‍

Highlights

• ParetoQ: Scaling Laws in Extremely Low-bit LLM Quantization by Meta (https://arxiv.org/pdf/2502.02631). The paper presents a unified framework that facilitates comparisons across 1-bit, 1.58-bit, 2-bit, 3-bit, and 4-bit quantization settings. The findings reveal a notable learning transition between 2 and 3 bits: For 3-bits and above, the fine-tuned models stay close to their original pre-trained distributions, whereas for learning 2-bit networks or below, the representations change drastically. By optimizing training schemes and refining quantization functions, the ternary 600M-parameter model even outperforms the previous SoTA ternary 3B-parameter model in accuracy, using only one-fifth of the parameters.

• QuEST: Stable Training of LLMs with 1-Bit Weights and Activations by ISTA and Red Hat AI (https://arxiv.org/pdf/2502.05003). The paper introduces quantization method that allows stable training with 1-bit weights and activations. It achieves this by improving two key aspects of QAT methods: (1) accurate and fast quantization of the (continuous) distributions of weights and activations via Hadamard normalization and MSE-optimal fitting; (2) a new trust gradient estimator based on the idea of explicitly minimizing the error between the noisy gradient computed over quantized states and the “true” (but unknown) full-precision gradient. Experiments on Llama-type architectures show that the method induces stable scaling laws across the entire range of hardware-supported precisions, and can be extended to sparse representations. The code is available at https://github.com/IST-DASLab/QuEST.

• Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention by Deepseek-AI, Peking University, University of Washington (https://arxiv.org/pdf/2502.11089). The paper presents a method with hardware-aligned optimizations to achieve efficient long-context modeling. It employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. The approach advances sparse attention design with two key features: (1) Authors achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) They enable end-to-end training, reducing pretraining computation without sacrificing model performance. Experiments show the model pretrained with the proposed method maintains or exceeds Full Attention models across general benchmarks, long-context tasks, and instruction-based reasoning. It achieves substantial speedups over Full Attention on 64k-length sequences across decoding, forward propagation, and backward propagation. Non-official implementations are available on GitHub.

• LSERVE: EFFICIENT LONG-SEQUENCE LLM SERVING WITH UNIFIED SPARSE ATTENTION by MIT, SJTU, Nvidia (https://arxiv.org/pdf/2502.14866). The paper introduces a system that accelerates long-sequence LLM serving via hybrid sparse attention. This method unifies different hardware-friendly, structured sparsity patterns for both prefilling and decoding attention into a single framework, where computations on less important tokens are skipped block-wise. It demonstrates the compatibility of static and dynamic sparsity in long-context LLM attention. Authors convert half of the attention heads to nearly free streaming heads in both the prefilling and decoding stages. Additionally, we they that only a constant number of KV pages is required to preserve long-context capabilities, irrespective of context length. They then design a hierarchical KV page selection policy that dynamically prunes KV pages based on query-centric similarity. The method accelerates LLM prefilling by up to 2.9x and decoding by 1.3-2.1x over vLLM, maintaining long-context accuracy. Code is released at https://github.com/mit-han-lab/omniserve.

• XAttention: Block Sparse Attention with Antidiagonal Scoring by Tsinghua University, MIT, SJTU, and NVIDIA (https://arxiv.org/pdf/2503.16428). The paper introduces XAttention method that significantly accelerates long-context inference in Transformers models using sparse attention. XAttention’s key innovation is the insight that the sum of antidiagonal values (i.e., from the lower-left to upper-right) in the attention matrix provides a powerful proxy for block importance. This allows for precise identification and pruning of non-essential blocks, resulting in high sparsity and dramatically accelerated inference. On RULER and LongBench for language, VideoMME for video understanding, and VBench for video generation—XAttention achieves accuracy comparable to full attention while delivering substantial computational gains. It shows up to 13.5x acceleration in attention computation. The code is available at https://github.com/mit-han-lab/x-attention.

‍

Papers with notable results

Quantization

• Optimizing Large Language Model Training Using FP4 Quantization by Microsoft and University of Science and Technology of China (https://arxiv.org/pdf/2501.17116). The work introduces the FP4 training framework for LLMs, addressing quantization challenges with two key ideas: a differentiable quantization estimator for precise weight updates and an outlier clamping and compensation strategy to prevent activation collapse. To ensure stability, the framework integrates a mixed-precision training scheme and vector-wise quantization. Experimental results demonstrate that our FP4 framework achieves accuracy comparable to BF16 and FP8, with minimal degradation, scaling effectively to 13B-parameter LLMs trained on up to 100B.
• MQuant: Unleashing the Inference Potential of Multimodal Large Language Models via Full Static Quantization by Houmo AI, Southeast University, and Xi’an Jiaotong University (https://arxiv.org/pdf/2502.00425). The work focuses on the problems of VLM quantization with a coarse scale granularity. It proposes several techniques to tackle the quantization problems, namely: Modality-Specific Static Quantization (MSQ), assigning distinct static scales for visual vs. textual tokens; Attention-Invariant Flexible Switching (AIFS), reordering tokens to preserve casual attention while eliminating expensive token-wise scale computations;  Rotation Magnitude Suppression (RMS), mitigating weight outliers arising from online Hadamard rotations. On five mainstream VLMs (including Qwen-VL, MiniCPM-V, CogVLM2), the method achieves near-floating-point accuracy under W4A8 setting. The code is planned to be published.
• An Empirical Study of LLaMA3 Quantization: From LLMs to MLLMs by The University of Hong Kong, Beihang University, and ETH Zurich (https://arxiv.org/pdf/2404.14047). Authors assessed the performance of the LLaMA3-based LLaVA-Next-8B model under 2-4 ultra-low bits with post-training quantization methods. Experimental results indicate that LLaMA3 still suffers from non-negligible degradation in linguistic and visual contexts, particularly under ultra-low bit widths. This highlights the significant performance gap at low bit-width that needs to be addressed in future developments. The code is available at: https://github.com/Macaronlin/LLaMA3-Quantization.
• Nanoscaling Floating-Point (NxFP): NanoMantissa, Adaptive Microexponents, and Code Recycling for Direct-Cast Compression of Large Language Models by Harvard University (https://arxiv.org/pdf/2412.19821). This paper profiles modern LLMs and identifies three main challenges of low-bit Microscaling format, i.e., inaccurate tracking of outliers, vacant quantization levels, nd wasted binary code. In response, Nanoscaling (NxFP) proposes three techniques, i.e., NanoMantissa, Adaptive Microexponent, and Code Recycling to enable better accuracy and smaller memory footprint than state-of-the-art MxFP. Experimental results on direct-cast inference across various modern LLMs demonstrate that the proposed methods outperform MxFP by up to 0.64 in perplexity and by up to 30% in accuracy on MMLU benchmarks.
• RoSTE: An Efficient Quantization-Aware Supervised Fine-Tuning Approach for Large Language Models by University of Minnesota and The Chinese University of Hong Kong (https://arxiv.org/pdf/2502.09003). The paper introduces a fine-tuning based method that directly optimizes quantized weights and rotation matrices within a single model architecture. It proposes a bilevel optimization formulation, where upper level subproblem optimizes weight matrices, while lower level subproblem employs a surrogate loss to guide the selection of rotation matrix. Authors designed an algorithm which alternates between (i) a QAT subroutine incorporating a rotation-enabled straightthrough-estimator (STE) update, and (ii) a low complexity heuristic for selecting rotation matrices based on the random Walsh-Hadamard matrix. They provide a theoretical analysis of the benefits of rotation-enabled quantization in QA-SFT by examining the prediction error resulted from the QAT stage of RoSTE. This analysis directly motivates the use of quantization error based surrogate loss and justifies the adoption.
• NESTQUANT: NESTED LATTICE QUANTIZATION FOR MATRIX PRODUCTS AND LLMS by MIT and Hebrew University of Jerusalem (https://arxiv.org/pdf/2502.09720). The paper proposes a PTQ scheme for weights and activations that is based on self-similar nested lattices. Recent work has mathematically shown such quantizers to be information-theoretically optimal for low-precision matrix multiplication. We implement a practical low-complexity version based on Gosset lattice, making it a drop-in quantizer for any matrix multiplication step (e.g., in self-attention, MLP etc). For example, the method quantizes weights, KV-cache, and activations of Llama-3-8B to 4 bits, achieving perplexity of 6.6 on wikitext2.
• ViM-VQ: Efficient Post-Training Vector Quantization for Visual Mamba by Zhejiang University and vivo Mobile (https://arxiv.org/pdf/2503.09509). A practical study of vector quantization method for Visual Mamba networks (ViMs). Authors identify several key challenges: 1) The weights of Mamba-based blocks in ViMs contain numerous outliers, significantly amplifying quantization errors. 2) When applied to ViMs, the latest VQ methods suffer from excessive memory consumption, lengthy calibration procedures, and suboptimal performance in the search for optimal codewords. They propose a post-training vector quantization method tailored for ViMs. It consists of two components: 1) a fast convex combination optimization algorithm that updates both the convex combinations and the convex hulls to search for optimal codewords, and 2) an incremental vector quantization strategy that incrementally confirms optimal codewords to mitigate truncation errors. The results demonstrate that the method achieves stateof-the-art performance in low-bit quantization across various visual tasks.
• SSVQ: Unleashing the Potential of Vector Quantization with Sign-Splitting by Zhejiang University and vivo Mobile (https://arxiv.org/pdf/2503.08668). The paper proposes the vector quantization approach which decouples the sign bit of weights from the codebook. It involves extracting the sign bits of uncompressed weights and performing clustering and compression on all-positive weights. Authors also introduce latent variables for the sign bit and jointly optimize both the signs and the codebook. Additionally, they implement a progressive freezing strategy for the learnable sign to ensure training stability. Experiments on modern models and tasks demonstrate that the method achieves a good compression-accuracy trade-off compared to conventional VQ. Authors also validate the algorithm on a hardware accelerator, showing that SSVQ achieves a 3× speedup over the 8-bit compressed model by reducing memory access.
• MergeQuant: Accurate 4-bit Static Quantization of Large Language Models by Channel-wise Calibration (https://arxiv.org/pdf/2503.07654). The paper introduces per-channel static quantization method. It integrates the per-channel quantization steps with the corresponding scalings and linear mappings through a Quantization Step Migration method, eliminating the quantization overheads before and after matrix multiplication. Authors also propose dimensional reconstruction and adaptive clipping to address the nonuniformity of quantization scale factors and redistribute the channel variations to the subsequent modules to balance the parameter distribution under QSM. They evaluate method on Llama 2 and Llama 3 models in W4A4 setting.
• QuantCache: Adaptive Importance-Guided Quantization with Hierarchical Latent and Layer Caching for Video Generation by Shanghai Jiao Tong University, MGTV, Shanhai Academy (https://arxiv.org/pdf/2503.06545). Authors propose a training-free inference acceleration framework that jointly optimizes hierarchical latent caching, adaptive importance-guided quantization, and structural redundancy-aware pruning. It achieves an end-to-end latency speedup of 6.72x on OpenSora with minimal loss in generation quality. experiments across multiple video generation benchmarks demonstrate the effectiveness of our method for DiT inference. The code and models will be available at https://github.com/JunyiWuCode/QuantCache.
• Matryoshka Quantization by Google DeepMind (https://arxiv.org/pdf/2502.06786). Practitioners are often forced to maintain multiple models with different quantization levels or serve a single model that best satisfies the quality-latency trade-off. On the other hand, integer data types, such as int8, inherently possess a nested (Matryoshka) structure where smaller bit-width integers, like int4 or int2, are nested within the most significant bits. In this paper, the authors propose Matryoshka Quantization (MatQuant), a multi-scale quantization technique that alleviates the aforementioned challenge. It allows us to train and maintain a single quantized model but serve it with the precision demanded by the deployment. Furthermore, leveraging MatQuant’s co-training and co-distillation regularization, int2 precision models extracted by MatQuant outperform standard int2 quantization by up to to 4% and 7% with OmniQuant and QAT as base algorithms respectively. Finally, authors demonstrate that by using an extra bit to represent outliers, a model with an effective precision of 2.05-bit gives an additional 6% improvement with OmniQuant as the base algorithm.

‍

Pruning/Sparsity

• Mamba-Shedder: Post-Transformer Compression for Efficient Selective Structured State Space Models by Intel Labs (https://arxiv.org/pdf/2501.17088v1).  This paper explores the compression of SSM-based models, particularly Mamba and its hybrids. The authors discuss the sensitivity of these models to the removal of selected components at different granularities to reduce the model size and computational overhead, thus improving their efficiency while maintaining accuracy. The proposed solutions, collectively referred to as Mamba-Shedder, achieve a speedup of up to 1.4x during inference, demonstrating that model efficiency can be improved by eliminating several redundancies with minimal impact on the overall model performance. The code is available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.
• DeepSeekAI, The Sparsity Revolution That Shook the AI Market: https://www.linkedin.com/pulse/deepseek-ai-sparsity-revolution-shook-market-jabar-riaz-hwvbf. The blogpost discusses two types of sparsity to train efficient and competitive LLM models, namely data sparsity and model sparsity.

‍

Other

• TPU Scaling Book from Google - a series of blog posts on how to optimize LLMs for Google TPUv5: https://jax-ml.github.io/scaling-book/applied-inference.
• The Ultra-Scale Playbook: Training LLMs on GPU Clusters https://huggingface.co/spaces/nanotron/ultrascale-playbook. A tutorial from HuggingFace on the basics of multi-GPU training and how to scale it.
• Qwen2.5-1M Technical Report by Alibaba (https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5-1M/Qwen2_5_1M_Technical_Report.pdf). Authors introduce Qwen2.5-1M, a series of models that extend the context length to 1 million tokens. Compared to the previous 128K version, the Qwen2.5-1M series has significantly enhanced long-context capabilities through long-context pretraining and post-training. To reduce inference costs, authors implement a sparse attention method along with chunked prefill optimization for deployment scenarios and a sparsity refinement method to improve precision. Additionally, they detail optimizations in the inference engine, including kernel optimization, pipeline parallelism, and scheduling optimization, which significantly enhance overall inference performance. Qwen2.5-1M models achieve a remarkable 3x to 7x prefill speedup in scenarios with 1 million tokens of context.
• WaferLLM: A Wafer-Scale LLM Inference System by University of Edinburgh and Microsoft (https://arxiv.org/pdf/2502.04563). The paper introduces LLM inference system that is guided by a device model that captures the unique hardware characteristics of wafer-scale architectures. It proposes MeshGEMM and MeshGEMV, the GEMM and GEMV implementations designed to scale effectively on wafer-scale accelerator. Authors focus on four principles when designing the implementation: Massive Parallel cores, Highly non-uniform memory access Latency, Constrained local Memory, and Limited hardware-assisted Routing. Evaluations show that the method achieves 200× better wafer-scale accelerator utilization than state-of-the-art systems. On a commodity wafer-scale accelerator, it delivers 606× faster and 22× more energy-efficient GEMV compared to an advanced GPU. One of the limitations of the method is a limited model size due to a need to replicate memory over the computational units to increase the latency.
• EmbBERT-Q: Breaking Memory Barriers in Embedded NLP by Politecnico di Milano (https://arxiv.org/pdf/2502.10001). The paper proposes a new LM model specifically designed for tiny devices, combining efficiency and effectiveness. Authors analytically evaluate the memory usage and computational complexity of the model and its components, providing a tool to evaluate the weights and activations of memory trade-offs required to operate within tiny device constraints. They also release all code, scripts, and model checkpoints at https://github.com/RiccardoBravin/tiny-LLM.
• M2R2: MIXTURE OF MULTI-RATE RESIDUALS FOR EFFICIENT TRANSFORMER INFERENCE by Apple (https://arxiv.org/pdf/2502.02040). The paper introduce Mixture of Multi-rate Residuals, a framework that dynamically modulates the velocity of residual transformations to optimize early residual alignment. This modification improves inference efficiency by better aligning intermediate representations at earlier stages. Authors show the efficacy of the technique in diverse optimization setups such as dynamic computing, speculative decoding, and MoE Ahead-of-Time. In self-speculative decoding setups, M2R2 achieves up to 2.8X speedups on MT-Bench under lossless conditions. In Mixture-of-Experts architectures, they enhance decoding speed by coupling early residual alignment with ahead-of-time expert loading into high-bandwidth memory. This enables concurrent memory access and computation, reducing the latency bottlenecks inherent in expert switching during decoding. Empirical results show that the method delivers a speedup of 2.9X in MoE architectures.
• Extending Language Model Context Up to 3 Million Tokens on a Single GPU by KAIST and DeepAuto.ai (https://arxiv.org/pdf/2502.08910). To enable efficient and practical long-context utilization, authors introduce an LLM inference framework that accelerates processing by dynamically eliminating irrelevant context tokens through a modular hierarchical token pruning algorithm. The method also allows generalization to longer sequences by selectively applying various RoPE adjustment methods according to the internal attention patterns within LLMs. They also offload the key-value cache to host memory during inference, significantly reducing GPU memory pressure. As a result, the method enables the processing of up to 3 million tokens on a single L40s 48GB GPU without any permanent loss of context information. The framework achieves an 18.95x.
• KernelBench: Can LLMs Write Efficient GPU Kernels? by Stanford University and Princeton University (https://arxiv.org/pdf/2502.10517). The paper introduces KernelBench, an open-source framework for evaluating LMs’ ability to write fast and correct kernels on a suite of 250 carefully selected PyTorch ML workloads. KernelBench represents a real-world engineering environment and making progress on the introduced benchmark directly translates to faster practical kernels. Auhors introduce a new evaluation metric fastp, which measures the percentage of generated kernels that are functionally correct and offer a speedup greater than an adjustable threshold p over baseline. Experiments across various models and test-time methods show that frontier reasoning models perform the best out of the box but still fall short overall, matching the PyTorch baseline in less than 20% of the cases.
• Investigating the Impact of Quantization Methods on the Safety and Reliability of Large Language Models by Skolkovo Institute, Artificial Intelligence Research Institute, HSE University (https://arxiv.org/pdf/2502.15799). Authors introduce OpenSafetyMini, a openended safety dataset designed to better distinguish between models. They evaluate 4 state-ofthe-art quantization techniques across LLaMA and Mistral models using 4 benchmarks, including human evaluations. Findings reveal that the optimal quantization method varies for 4-bit precision, while vector quantization techniques deliver the best safety and trustworthiness performance at 2-bit precision, providing foundation for future research. The dataset and reproduces available at: https://github.com/On-Point-RND/OpenSafetyMini-Investigating-the-Impact-of-Quantization-Methods-on-the-Safety-and-Reliability-of-LLM.
• MOBA: MIXTURE OF BLOCK ATTENTION FOR LONG-CONTEXT LLMS by Moonshot AI, Tsinghua University, and Zhejiang University (https://arxiv.org/pdf/2502.13189v1). In this work, authors propose a solution that adheres to the “less structure” principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. They introduce Mixture of Block Attention (MoBA), an approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. It is based on block partitioning and routing strategy within Multi-Head Self-Attention. The code is available at https://github.com/MoonshotAI/MoBA.
• JUDGE DECODING: FASTER SPECULATIVE SAMPLING REQUIRES GOING BEYOND MODEL ALIGNMENT by Meta GenAI and ETH Zurich (https://openreview.net/pdf?id=mtSSFiqW6y). The paper demonstrates through a series of experiments how the decision mechanism in speculative decoding rejects many high-quality tokens, identifying a key limitation of the technique. Authors adapt verification using ideas from LLM-as-a-judge, eliciting the same versatile rating capability in the target by adding a simple linear layer that can be trained in under 1.5 hours. Using a Llama 8B/70B-Judge, the proposed approach obtains speedups of 9x over standard decoding, achieving an unprecedented 129 tokens/s, while maintaining the quality of Llama-405B on a range of benchmarks.

‍

Software

• FlashMLA by Deepseek: https://github.com/deepseek-ai/FlashMLA. FlashMLA is an efficient MLA decoding kernel for Hopper GPUs, optimized for variable-length sequences serving.
• DeepSeek releases DeepGEMM is a library designed for clean and efficient FP8 General Matrix Multiplications (GEMMs) with fine-grained scaling: https://github.com/deepseek-ai/DeepGEMM.
• SVDQuant Meets NVFP4: 4× Smaller and 3× Faster FLUX with 16-bit Quality on NVIDIA Blackwell GPUs: https://hanlab.mit.edu/blog/svdquant-nvfp4.
• M3 Ultra Runs DeepSeek R1 With 671 Billion Parameters Using 448GB Of Unified Memory, Delivering High Bandwidth Performance At Under 200W Power Consumption, With No Need For A Multi-GPU Setup: https://wccftech.com/m3-ultra-chip-handles-deepseek-r1-model-with-671-billion-parameters/.
• NVIDIA TensorRT Model Optimizer: https://github.com/NVIDIA/TensorRT-Model-Optimizer. A Library to Quantize and Compress Deep Learning Models for Optimized Inference on GPUs.

‍