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.

‍

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

‍

Summary

What a quarter! Tons of works for Transformer model optimization in Q4’24 including fundamental ones such as “scaling lows for quantized LLMs“. Such a huge effort can indicate a growing adoption of LLMs and AI in general and the need for a further cost reduction. We had to extend the Highlights to six papers this time considering the amount of work being done.

‍

Highlights

• Scaling Laws for Precision by Harvard, Stanford, MIT, Carnegie Mellon Universities, and Databricks (https://arxiv.org/pdf/2411.04330). In this work, authors devise “precision-aware” scaling laws for both training and inference. They propose that training in lower precision reduces the model’s effective parameter count, allowing predicting the additional loss incurred from training in low precision and post-train quantization. For inference, they find that the degradation introduced by post-training quantization increases as models are trained on more data, eventually making additional pretraining data actively harmful. For training, their scaling laws allow predicting the loss of a model with different parts in different precisions and suggest that training larger models in lower precision may be compute optimal. Authors unify the scaling laws for post and pretraining quantization to arrive at a single functional form that predicts degradation from training and inference in varied precisions. They fit on over 465 pretraining runs and validate our predictions on model sizes up to 1.7B parameters trained on up to 26B tokens.

• Low-Bit Quantization Favors Undertrained LLMs: Scaling Laws for Quantized LLMs with100T Training Tokens by University of Virginia, Tencent AI Lab Seattle (https://arxiv.org/pdf/2411.17691).Authors propose a perspective that one can use to measure an LLM’s training levels and determine the number of training tokens required for fully training LLMs of various sizes. Moreover, authors use the scaling laws to predict the quantization performance of different-sized LLMs trained with 100 trillion tokens. Our projection shows that the low-bit quantization performance of future models, which are expected to be trained with over 100 trillion tokens, may NOT be desirable. This poses a potential challenge for low-bit quantization in the future and highlights the need for awareness of a model’s training level when evaluating low-bit quantization research. Checkpoints are available at: https://huggingface.co/Xu-Ouyang.

• Hymba: A Hybrid-head Architecture for Small Language Models by Nvidia, Georgia Institute of Technology, and HKUST (https://www.arxiv.org/abs/2411.13676).The paper introduces a family of small language models featuring a hybrid-head parallel architecture that integrates transformer attention mechanisms with state space models (SSMs) for enhanced efficiency. Additionally, authors introduce learnable meta tokens that are prepended to prompts, storing critical information. This model is further optimized by incorporating cross-layer key-value (KV) sharing and partial sliding window attention, resulting in a compact cache size. Hymba-1.5B-Base model surpasses all sub-2B public models in performance and even outperforms Llama-3.2-3B with1.32% higher average accuracy, an 11.67× cache size reduction, and 3.49×throughput. Models are available on the Hugging Face Hub.

• THE SUPER WEIGHT IN LARGE LANGUAGE MODELS by Apple and University of Notre Dame (https://arxiv.org/pdf/2411.07191). This work presents a finding that pruning single parameters can destroy an LLM’s ability to generate text – increasing perplexity by 3 orders of magnitude and reducing zero-shot accuracy to guessing. It proposes a data-free method for identifying such parameters, termed super weights, using a single forward pass through the model. Authors find that these super weights induce correspondingly rare and large activation outliers, termed super activations. When preserved with high precision, super activations can improve simple round-to-nearest quantization to become competitive with state-of-the-art methods. For weight quantization, they similarly find that by preserving the super weight and clipping other weight outliers, round-to-nearest quantization can scale to much larger block sizes than previously considered. The code is available at n https://github.com/mengxiayu/LLMSuperWeight.

• Pushing the Limits of Large Language Model Quantization via the Linearity Theorem by Yandex, HSE University, ISTA, GenAI CoE, KAUST, Neural Magic (https://arxiv.org/pdf/2411.17525). The paper presents a “linearity theorem” establishing a direct relationship between the layer-wise ℓ2 reconstruction error and the model perplexity increase due to quantization. This enables two novel applications: (1) a simple data-free LLM quantization method using Hadamard rotations and MSE-optimal grids, dubbed HIGGS, which outperforms all prior data-free approaches such as the extremely popular NF4 quantized format, and (2) an optimal solution to the problem of finding non-uniform per-layer quantization levels which match a given compression constraint in the medium-bit width regime, obtained by reduction to dynamic programming. Authors demonstrate improved accuracy-compression trade-offs on Llama-3.1 and 3.2- family models, as well as on Qwen-family models.

• SANA:EFFICIENT HIGH-RESOLUTION IMAGE SYNTHESIS WITH LINEAR DIFFUSION TRANSFORMERS by NVIDIA, MIT, Tsinghua University (https://arxiv.org/pdf/2410.10629). Authors introduce Sana, a text-to-image frame work that can generate images up to 4096×4096 resolution. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8×,authors trained an AE that can compress images 32×, effectively reducing the number of latent tokens. (2) Linear DiT: they replace all vanilla attention in DiT with linear attention (3) Decoder-only text encoder: they replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: they propose Flow-DPM-Solver to reduce sampling steps. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20times smaller and 100+ times faster in measured throughput. Project web page with code: https://nvlabs.github.io/Sana/.

‍

Papers with notable results

Quantization

• VPTQ: EXTREME LOW-BIT VECTOR POST-TRAINING QUANTIZATION FOR LARGE LANGUAGE MODELS by Microsoft and University of Science and Technology of China (https://arxiv.org/abs/2409.17066). The authors introduce Vector Post-Training Quantization and use Second-Order Optimization to formulate the LLM VQ problem and guide the algorithm design by solving the optimization. They further refine the weights using Channel-Independent Second-Order Optimization for a granular VQ. In addition, by decomposing the optimization problem, authors propose a brief codebook initialization algorithm and extend VPTQ to support residual and outlier quantization, which enhances model accuracy and further compresses the model. The method achieves good results on llama-2 and llama-3 model families, resulting in a 1.6-1.8× increase in inference throughput compared to SOTA. The code is available at https://github.com/microsoft/VPTQ.
• ADDITION IS ALL YOU NEED FOR ENERGY-EFFICIENT LANGUAGE MODELS by BitEnergy AI (https://arxiv.org/pdf/2410.00907). Authors propose the linear-complexity multiplication 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 which can potentially reduce 95% energy cost by elementwise floating point tensor multiplications and 80% energy cost of dot products. A numerical analysis and experiments indicate that the method with 4-bit mantissa achieves comparable precision as float8 e4m3 multiplications, and 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.
• BitNet a4.8: 4-bit Activations for 1-bit LLMs by Microsoft and University of Chinese Academy of Sciences (https://arxiv.org/pdf/2411.04965). In this work, authots introduce BitNet a4.8, enabling 4-bit activations for 1-bit LLMs. BitNet a4.8 employs a hybrid quantization and sparsification strategy to mitigate the quantization errors introduced by the outlier channels. Specifically, they utilize 4-bit activations for inputs to the attention and feed-forward network layers, while sparsifying intermediate states followed with 8-bit quantization. Extensive experiments demonstrate that BitNet a4.8 achieves performance comparable to BitNet b1.58 with equivalent training costs, while being faster in inference with enabling 4-bit (INT4/FP4) kernels. Additionally, BitNet a4.8 activates only 55% of parameters and supports 3-bit KV cache.
• MagR: Weight Magnitude Reduction for Enhancing Post-Training Quantization by Uniiversity at Albany and IBM (https://arxiv.org/pdf/2406.00800). MagR is an optimization-based preprocessing technique for improving post-training quantization. It solves an l_∞-regularized problem to reduce outlier weights and center them around zero, enabling smoother and more efficient quantization. Unlike linear transformations that require extra steps at inference, MagR is a non-linear transformation that adds no overhead. Experiments show state-of-the-art results, including a Wikitext2 perplexity of 6.7 on the LLaMA2-70B model using per-channel INT2 weight quantization.
• Cherry on Top: Parameter Heterogeneity and Quantization in Large Language Models by Shanghai University of Finance and Economics (https://arxiv.org/pdf/2404.02837). This paper identifies “cherry” parameters in large language models—those few parameters with a disproportionately large effect on performance—while most parameters matter far less. Building on this insight, the authors introduce CherryQ, a quantization technique that maintains these critical parameters in high precision and aggressively quantizes the rest. CherryQ delivers improved perplexity and downstream task results, enabling efficient LLM deployment. Remarkably, a 3-bit quantized Vicuna-1.5 model matches the performance of 16-bit models, illustrating the potential of leveraging parameter heterogeneity for more efficient inference.
• QTIP: Quantization with Trellises and Incoherence Processing by Cornell University (https://arxiv.org/pdf/2406.11235). QTIP is a new PTQ approach leveraging trellis-coded quantization (TCQ) for ultra-high-dimensional vector quantization of LLM weights. Unlike conventional VQ methods whose codebook size grows exponentially with dimension, TCQ uses a stateful decoder to maintain efficiency as dimensions scale. QTIP provides a hardware-friendly “bitshift” trellis structure and can be tuned for lookup-only or computed lookup-free decoding. This allows faster, more memory-efficient inference and achieves state-of-the-art quantization quality, outperforming previous VQ-based methods.
• ESPACE: Dimensionality Reduction of Activations for Model Compression by NVIDIA (https://arxiv.org/pdf/2410.05437). ESPACE introduces a new LLM compression method based on dimensionality reduction of activations rather than weight decomposition. By projecting activations onto pre-calibrated principal components, ESPACE retains model expressivity without retraining. It achieves weight compression indirectly through matrix multiplication associativity. Theoretically, it ensures optimal computational accuracy when constructing projection matrices. Experiments show up to 50% compression on GPT3, Llama2, and Nemotron4 with minimal accuracy loss, and in some cases, improved perplexity. ESPACE also speeds up inference. Compared to existing tensor decomposition methods, ESPACE advances state-of-the-art LLM compression.
• Delta-CoMe: Training-Free Delta-Compression with Mixed-Precision for Large Language Models by several Chinese universities (https://arxiv.org/pdf/2406.08903). This work addresses compressing delta weights for fine-tuned LLMs, where maintaining task-specific performance is challenging using low-rank or low-bit methods. Observing that delta weights’ singular values are long-tailed, the authors propose a mixed-precision delta quantization approach. By assigning higher-bit precision to more influential singular vectors, their method preserves accuracy. Experiments on diverse fine-tuned LLMs—including math, code, and chat models—show that this approach matches full-precision performance and significantly outperforms standard low-rank and low-bit baselines. It is also compatible with various backbone models, such as Llama-2, Llama-3, and Mistral.
• StepbaQ: Stepping backward as Correction for Quantized Diffusion Models by MediaTek and Purdue University (https://openreview.net/pdf?id=cEtExbAKYV). StepbaQ reframes quantization error in diffusion models as a “stepback” in their denoising process. By analyzing how this accumulated error distorts the sampling trajectory, StepbaQ introduces a correction mechanism that uses quantization error statistics from a small calibration dataset. Without altering quantization settings, it significantly improves model quality. For instance, StepbaQ boosts the FID score of quantized SD v1.5 by 7.30 under W8A8, and SDXL-Turbo by 17.31 under W4A8. This plug-and-play solution enhances performance on resource-constrained devices while maintaining broad applicability.
• LLMCBench: Benchmarking Large Language Model Compression for Efficient Deployment by Beihang University, ETH Zurich and Canerige Mellon University (https://arxiv.org/pdf/2410.21352). LLMCBench is a comprehensive benchmark designed to evaluate large language model compression techniques under realistic conditions. Moving beyond limited and specialized assessments, it tests various models, datasets, and metrics. LLMCBench establishes clearly defined evaluation tracks based on real production requirements and conducts extensive experiments with multiple mainstream compression methods. Through in-depth analysis, it offers insights into the strengths and weaknesses of these approaches. Ultimately, LLMCBench aims to guide the selection and design of effective compression algorithms, serving as a valuable resource for future research and development in LLM efficiency.
• DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (https://duquant.github.io/). Generalization of the SmoothQuant algorithm which allows to mitigate the massive outliers and quantize not just LLM weights but activations as well. Shows promising results for LLama2/3 -8B W6A6 and W4A4 quantization. The code is available at: https://github.com/Hsu1023/DuQuant.
• Efficient Multi-task LLM Quantization and Serving for Multiple LoRA Adapters (https://openreview.net/pdf?id=HfpV6u0kbX). Multi quantized Lora adapters quantization via techniques like Multi-Lora GPTQ and LoRa Inlaid. Technics to dynamically add a new task/dataset to existing quantized LLM are discussed in the paper, promising pipeline for quantized LLM serving / update is presented.
• PROGRESSIVE MIXED-PRECISION DECODING FOR EFFICIENT LLM INFERENCE. Samsung AI Center, Cambridge UK, Imperial College London UK (https://arxiv.org/abs/2410.13461). The authors propose a novel phase-aware method that selectively allocates precision during different phases of LLM inference, achieving both strong context extraction during prefill and efficient memory bandwidth utilization during decoding. To further address the memory-boundedness of the decoding phase, the authors introduce Progressive Mixed-Precision Decoding (PMPD), a technique that enables the gradual lowering of precision deeper in the generated sequence, together with a spectrum of precision-switching schedulers that dynamically drive the precision lowering decisions in either task-adaptive or prompt-adaptive manner. Extensive evaluation across diverse language tasks shows that when targeting Nvidia GPUs, PMPD achieves 1.4−12.2× speedup in LLM linear layers over fp16 models, while when targeting an LLM-optimized NPU, our approach delivers a throughput gain of 3.8−8.0× over fp16 models and up to 1.54× over uniform quantization approaches while preserving the output quality.
• AMXFP4: TAMING ACTIVATION OUTLIERS WITH ASYMMETRIC MICROSCALING FLOATING-POINT FOR 4-BIT LLM INFERENCE by Hanyang University and Rebellions Inc. (https://arxiv.org/pdf/2411.09909). Authors propose Asymmetric Microscaling 4-bit Floating-Point (AMXFP4) for efficient LLM inference. This data format leverages asymmetric shared scales to mitigate outliers while naturally capturing the asymmetry introduced by group-wise quantization. Unlike conventional 4-bit quantization methods that rely on data rotation and costly calibration, AMXFP4 uses asymmetric shared scales for direct 4-bit casting, achieving better quantization accuracy across various LLM tasks, including multi-turn conversations, long-context reasoning, and visual question answering The code is available at https://github.com/aiha-lab/MX-QLLM.git.
• SageAttention2 Technical Report: Accurate 4 Bit Attention for Plug-and-play Inference Acceleration by Tsinghua University (https://arxiv.org/pdf/2411.10958). Authors propose an improvement over the previous version of SageAttention method which utilizes 4-bit matrix multiplication (Matmul) alongside additional precision-enhancing techniques. First, they propose to quantize matrixes (Q, K) to INT4 in a warp-level granularity and quantize matrixes to FP8. Second, they propose a method to smooth Q and V, enhancing the accuracy of attention. Third, they propose an adaptive quantization method to ensure the end-to-end metrics over various models. Authors claim a good performance improvement at small drop of accuracy for large language processing, image generation, and video generation. The codes are available at https://github.com/thu-ml/SageAttention.
• CATASTROPHIC FAILURE OF LLM UNLEARNING VIA QUANTIZATION (https://openreview.net/pdf?id=lHSeDYamnz). The paper reveals that applying quantization to models that have undergone unlearning can restore the "forgotten" information. Authors conduct experiments using various quantization techniques across multiple precision levels to evaluate this phenomenon. They find that for unlearning methods with utility constraints, the unlearned model retains an average of 21% of the intended forgotten knowledge in full precision, which significantly increases to 83% after 4-bit quantization. They also provide a theoretical explanation for the observed phenomenon and propose a quantization-robust unlearning strategy aimed at mitigating this intricate issue. Results highlight a fundamental tension between preserving the utility of the unlearned model and preventing knowledge recovery through quantization, emphasizing the challenge of balancing these two objectives. The code is available at: https://anonymous.4open.science/r/FailureUnlearning-20DE.
• Llama Guard 3-1B-INT4: Compact and Efficient Safeguard for Human-AI Conversations by Meta (https://arxiv.org/pdf/2411.17713). Author used a complex approach to optimize Llama Guard 3-1B for mobile platforms. Namely, they reduce the number of decoder blocks and MLP width of Llama Guard 3-1B-INT4 using a block-level and neuron-level sensitivity analysis, respectively. They use quantization-aware training (QAT) to reduce the weight bitwidth to 4 and the activation bitwidth to 8, such that the model size is cut down by 4× and the model can be efficiently run via ExecuTorch’s XNNPACK backend. They make use of the fact that Llama Guard models only require a limited output vocabulary and reduce the unembedding layer output shape from 128k to 20. Finally, the authors fine-tune the model with distillation from a Llama Guard 2-8B teacher to recover any lost model quality resulting from the compression steps.
• MPQ-DM: Mixed Precision Quantization for Extremely Low Bit Diffusion Models by Institute of Computing Technology, University of Chinese Academy of Sciences, ETH Zurich, Beijing Jiaotong University (https://arxiv.org/pdf/2412.11549). The paper presents a Mixed-Precision Quantization method for Diffusion Models. It mainly relies on two techniques: (1) To mitigate the quantization error caused by outlier severe weight channels, authors propose an Outlier-Driven Mixed Quantization (OMQ) technique that uses Kurtosis to quantify outlier salient channels and apply optimized intra-layer mixed-precision bit-width allocation to recover accuracy performance within target efficiency. (2) To robustly learn representations crossing time steps, they construct a Time-Smoothed Relation Distillation (TRD) scheme between the quantized diffusion model and its full-precision counterpart, transferring discrete and continuous latent to a unified relation space to reduce the representation inconsistency. The method achieves good generation results on public benchmarks in low-bit quantization settings, e.g. W3A6, W3A4. Code is planned to be released here.
• Panacea: Novel DNN Accelerator using Accuracy-Preserving Asymmetric Quantization and Energy-Saving Bit-Slice Sparsity by POSTECH, University of Michigan (https://arxiv.org/pdf/2412.10059). The paper discloses how to build AI accelerator that leverages Bit-Slice Sparsity for the most prominent integer quantization scheme W-sym, A-asym. In contrast to the previous bit-slice computing, the accelerator compresses frequent nonzero slices, generated by asymmetric quantization, and skips their operations. To increase the slice level sparsity of activations, authors also introduce two algorithm hardware co-optimization methods: a zero-point manipulation and a distribution-based bit-slicing.
• Efficiency Meets Fidelity: A Novel Quantization Framework for Stable Diffusion by Zhejiang University and vivo Mobile Communication Co (https://arxiv.org/pdf/2412.06661). The paper introduces a mix-precision quantization strategy, multi-timestep activation quantization, and time information precalculation techniques to ensure high fidelity image generation of Stable Diffusion models in comparison to floating-point counterparts. The method achieves a good consistency of the image generation under the W8A8 and W4A8 settings.
• PREFIXQUANT: STATIC QUANTIZATION BEATS DYNAMIC THROUGH PREFIXED OUTLIERS IN LLMS by The University of Hong Kong, Shanghai AI Laboratory, Tongji University (https://arxiv.org/pdf/2410.05265). The paper proposes a technique that isolates outlier tokens offline without re-training. Specifically, it identifies high-frequency outlier tokens and prefixes them in the KV cache, preventing the generation of outlier tokens during inference and simplifying quantization. The method achieves very promising results in LLM static quantizaiton. For instance, in W4A4KV4 Llama-3-8B, with per-tensor static quantization it achieves a 7.43 WikiText2 perplexity and 71.08% average accuracy on 5 common-sense reasoning tasks. Additionally, the inference speed of W4A4 quantized models using PrefixQuant is 1.60× to 2.81× faster than FP16. The code is available at https://github.com/ChenMnZ/PrefixQuant.
• MixPE: Quantization and Hardware Co-design for Efficient LLM Inference by The Chinese University of Hong, Tsinghua University, Huawei Noah’s Ark Lab (https://arxiv.org/pdf/2411.16158). The paper proposes performing dequantization after per-group mixed-precision GEMM, significantly reducing dequantization overhead. Second, instead of relying on conventional multipliers, the method utilizes efficient shift&add operations for multiplication, optimizing both computation and energy efficiency. Experimental results demonstrate that the proposed design achieves better performance and energy trade-offs.
• “GIVE ME BF16 OR GIVE ME DEATH”? ACCURACY-PERFORMANCE TRADE-OFFS IN LLM QUANTIZATION by Neural Magic, Institute of Science and Technology Austria (https://arxiv.org/pdf/2411.02355). A thorough investigation, encompassing over 500,000 individual evaluations, yields several key findings: (1) FP8 weight and activation quantization (W8A8-FP) is lossless across all model scales, (2) INT8 weight and activation quantization (W8A8-INT) incurs surprisingly low 1-3% accuracy degradation, and (3) INT4 weight-only quantization (W4A16-INT) is competitive with 8-bit integer weight and activation quantization. They find that W4A16 offers the best cost-efficiency for synchronous deployments and for asynchronous deployment on mid-tier GPUs. At the same time, W8A8 formats excel in asynchronous “continuous batching” deployment of mid- and large-size models on high-end GPUs.
• GWQ: Gradient-Aware Weight Quantization for Large Language Models by PKU, CASIA, THU, USTB, UNITN, ETHz, PolyU, UCAS (https://arxiv.org/pdf/2411.00850). The authors propose gradient-aware weight quantization that leverages gradients to localize outliers, requiring only a minimal amount of calibration data for outlier detection. It retains the weights corresponding to the top 1% outliers preferentially at FP16 precision, while the remaining non-outlier weights are stored in a low-bit format. GWQ found experimentally that utilizing the sensitive weights in the gradient localization model is more scientific than utilizing the sensitive weights in the Hessian matrix localization model. The method shows accurate results for both LLM and VLM quantization.
• SDP4Bit: Toward 4-bit Communication Quantization in Sharded Data Parallelism for LLM Training by Indiana University, ByteDance, and University of Houston (https://arxiv.org/pdf/2410.15526). The paper proposes a method that reduces the communication of weights and gradients during the training to nearly 4 bits via two techniques: quantization on weight differences, and two-level gradient smooth quantization. Furthermore, the method presents an algorithm system co-design with runtime optimization to minimize the computation overhead of compression. Authors empirically evaluate the accuracy on the pre-training of GPT models with up to 6.7 billion parameters, and the results demonstrate a negligible impact on training loss. Furthermore, speed experiments show up to 4.08× speedup in end-to-end throughput on a scale of 128 GPUs.
• Quamba: A Post-Training Quantization Recipe for Selective State Space Models by University of Texas at Austin, National Yang Ming Chiao Tung University, and Cornell University (https://arxiv.org/pdf/2410.13229). Authors propose a static 8-bit per-tensor SSM quantization method which suppresses the maximum values of the input activations to the selective SSM for finer quantization precision and quantizes the output activations in an outlier-free space with Hadamard transform. 8-bit weight-activation quantized Mamba 2.8B SSM benefits from hardware acceleration and achieves a 1.72 × lower generation latency on an Nvidia Orin Nano 8G, with only a 0.9% drop in average accuracy on zero-shot tasks. Code is released at https://github.com/enyac-group/Quamba.
• RESTRUCTURING VECTOR QUANTIZATION WITH THE ROTATION TRICK by Stanford University and Google DeepMind (https://arxiv.org/pdf/2410.06424). The paper proposes a way to propagate gradients through the vector quantization layer of VQ-VAEs. The method smoothly transforms each encoder output into its corresponding codebook vector via a rotation and rescaling linear transformation that is treated as a constant during backpropagation. As a result, the relative magnitude and angle between encoder output and codebook vector becomes encoded into the gradient as it propagates through the vector quantization layer and back to the encoder. Еhis restructuring improves reconstruction metrics, codebook utilization, and quantization error. Code is available at https://github.com/cfifty/rotation_trick.

‍

Pruning / Sparsity

• MaskLLM: Learnable Semi-Structured Sparsity for Large Language Models by NVIDIA National University of Singapore (https://arxiv.org/pdf/2409.17481). The paper introduces several fundamental findings on applying N:M sparsity to LLM models. It explicitly models N:M patterns as a learnable distribution through Gumbel Softmax sampling. This approach facilitates end-to-end training on large-scale datasets and offers two notable advantages: 1) High-quality Masks - our method effectively scales to large datasets and learns accurate masks; 2) Transferability - the probabilistic modeling of mask distribution enables the transfer learning of sparsity across domains or tasks. The method achieves SOTA results on Wikitext and as well as shows lossless compression for many downstream language tasks. The code is available at https://github.com/NVlabs/MaskLLM.
• MRT5: DYNAMIC TOKEN MERGING FOR EFFICIENT BYTE-LEVEL LANGUAGE MODELS by Stanford University (https://arxiv.org/pdf/2410.20771). The paper introduces a more efficient variant of ByT5 that integrates a token deletion mechanism in its encoder to dynamically shorten the input sequence length. After processing through a fixed number of encoder layers, a learnt delete gate determines which tokens are to be removed and which are to be retained for subsequent layers. MrT5 effectively “merges” critical information from deleted tokens into a more compact sequence, leveraging contextual information from the remaining tokens. In continued pre-training experiments, we find that MrT5 can achieve significant gains in inference runtime with minimal effect on performance. Code is available here: https://github.com/jkallini/mrt5.
• SQFT: Low-cost Model Adaptation in Low-precision Sparse Foundation Models by Intel Labs (https://aclanthology.org/2024.findings-emnlp.749.pdf). The authors propose and end-to-end solution for low-precision sparse parameter-efficient fine-tuning of large pre-trained models, allowing for effective model adaptation in resource-constrained environments. Additionally, an innovative strategy enables the merging of sparse weights with low-rank adapters without losing sparsity and accuracy, overcoming the limitations of previous approaches. SQFT also addresses the challenge of having quantized weights and adapters with different numerical precisions, enabling merging in the desired numerical format without sacrificing accuracy. Multiple adaptation scenarios, models, and comprehensive sparsity levels demonstrate the effectiveness of SQFT. Models and code are available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.
• Post-Training Statistical Calibration for Higher Activation Sparsity by Intel Labs (https://arxiv.org/pdf/2412.07174). The paper presents a post-training activation pruning framework that (1) generalizes sparsification by input activations of Fully-Connected layers for generic and flexible application across Transformers, and (2) features a simple Mode-Centering technique to pre-calibrate activation distributions for maximizing post-training sparsity. The results demonstrate robust Pareto efficiency compared to prior methods, translating to a 1.5x additional LLM decoding speedup against] at iso model quality. The effectiveness of the method is empirically verified across a wide range of models, including recent Transformer Decoders, MoE, Mamba2, Encoding Transformer, and pre-quantized models. The code is available at: https://github.com/IntelLabs/SCAP.
• HashAttention: Semantic Sparsity for Faster Inference by UC Berkeley and ETH Zurich (https://arxiv.org/pdf/2412.14468). The paper proposes an approach that is casting pivotal token identification as a recommendation problem. Given a query, it encodes keys and queries in Hamming space capturing the required semantic similarity using learned mapping functions. The method identifies pivotal tokens for a given query in this Hamming space using bitwise operations, and only these pivotal tokens are used for attention computation. It can reduce the number of tokens used by a factor of 1/32× for the Llama-3.1-8B model with LongBench, keeping average quality loss within 0.6 points, while using only 32 bits per token auxiliary memory. Code is planned to be released.
• BEYOND 2:4: EXPLORING V:N:M SPARSITY FOR EFFICIENT TRANSFORMER INFERENCE ON GPUS by Tsinghua University, Huawei Noah’s Ark Lab, Beijing Jiaotong University (https://arxiv.org/pdf/2410.16135). Authors propose three approaches to enhance the applicability and accuracy of V:N:M-sparse Transformers, including heuristic V and M selection, V:N:M-specific channel permutation and three-staged LoRA training techniques. Experimental results show that, with with this, the DeiT-small achieves lossless accuracy at 64:2:5 sparsity, while the DeiT-base maintains accuracy even at 64:2:8 sparsity. In addition, the fine-tuned LLama2-7B at 64:2:5 sparsity performs comparably or better than training-free 2:4 sparse alternatives on downstream tasks.

‍

Other

• InfiniPot: Infinite Context Processing on Memory-Constrained LLMs from by Qualcomm AI Research , Qualcomm Korea YH (https://arxiv.org/pdf/2410.01518). The paper introduces a KV cache control framework designed to enable pre-trained LLMs to manage extensive sequences within fixed memory constraints efficiently, without requiring additional training. The method leverages Continual Context Distillation (CCD), an iterative process that compresses and retains essential information through novel importance metrics, maintaining critical data. This distillation process is based on the combination of CE-loss over the predicted tokens and Attention scores. Evaluations indicate that the method significantly outperforms models trained for long contexts in various NLP tasks.
• DEEP COMPRESSION AUTOENCODER FOR EFFICIENT HIGH-RESOLUTION DIFFUSION MODELS by MIT, Tsinghua University, and NVIDIA (https://arxiv.org/pdf/2410.10733). Authors highlight that existing autoencoders have demonstrated impressive results at a moderate spatial compression ratio (e.g., 8×) but fail to maintain satisfactory reconstruction accuracy for high spatial compression ratios (e.g., 64×). They address this by introducing two techniques: (1) Residual Autoencoding, where we design our models to learn residuals based on the space-to-channel transformed; (2) Decoupled High-Resolution Adaptation, a decoupled three-phase training strategy for mitigating the generalization penalty of high spatial-compression autoencoders. Authors improve the autoencoder’s spatial compression ratio up to 128 while maintaining the reconstruction quality achieving significant speedup without accuracy drop (19.1× inference speedup and 17.9× training speedup on H100 GPU). Code is available at https://github.com/mit-han-lab/efficientvit.
• EoRA: Training-free Compensation for Compressed LLM with Eigenspace Low-Rank Approximation by Nvidia (https://arxiv.org/pdf/2410.21271). The paper proposes a method that directly minimizes compression-induced errors without requiring gradient-based training small amount of calibration data. The method projects compression errors into the eigenspace of input activations, leveraging eigenvalues to effectively prioritize the reconstruction of high-importance error components. It shows good results for compressed LLaMA2/3 models on various tasks, such as language generation, commonsense reasoning, and math reasoning tasks (e.g., 31.31%/12.88% and 9.69% improvements on ARC-Easy/ARC-Challenge and MathQA when compensating LLaMA3-8B that is quantized to 4-bit and pruned to 2:4 sparsity).
• Eigen Attention: Attention in Low-Rank Space for KV Cache Compression by Purdue University (https://arxiv.org/pdf/2408.05646). Authors propose Eigen Attention, which performs the attention operation in a low-rank space, thereby reducing the KV cache memory overhead. The proposed approach is orthogonal to existing KV cache compression techniques and can be used synergistically with them. Experiments demonstrate that Eigen Attention results in up to 40% reduction in KV cache sizes and up to 60% reduction in attention operation latency with minimal drop in performance. Code is available at https://github.com/UtkarshSaxena1/EigenAttn.
• RAGCache: Efficient Knowledge Caching for Retrieval-Augmented Generation by Peking University and ByteDance (https://arxiv.org/pdf/2404.12457). Authors propose RAGCache, the system that caches the intermediate states of external knowledge and shares them across multiple queries to reduce the redundant computation. They design a prefix-aware GDSF replacement policy that leverages the characteristics of RAG to minimize the miss rate and a dynamic speculative pipelining approach to minimize the end-to-end latency. The experimental results show that RAGCache reduces the time to first token (TTFT) by up to 4× and improves the throughput by up to 2.1× compared to vLLM integrated with Faiss.
• STAR: Synthesis of Tailored Architectures by Liquid AI (https://arxiv.org/pdf/2411.17800). In this work, authors propose a NAS-based approach for the synthesis of LLM architectures. This approach combines a search space based on the theory of linear input-varying systems, supporting a hierarchical numerical encoding into architecture genomes. The genomes are automatically refined and recombined with gradient-free, evolutionary algorithms to optimize for multiple model quality and efficiency metrics. Using the method, authos optimize large populations of new architectures, leveraging diverse computational units and interconnection patterns, improving over highly-optimized Transformers and striped hybrid models on the frontier of quality, parameter size, and inference cache for autoregressive language modeling.
• SWITTI: Designing Scale-Wise Transformers for Text-to-Image Synthesis by Yandex Research, HSE University, MIPT, Skoltech, ITMO University (https://arxiv.org/pdf/2412.01819). The paper presents text-to-image transformer that employs architectural modifications to improve training stability and convergence and excludes explicit autoregression for more efficient sampling and better scalability. Compared to state-of-the-art text-to-image diffusion models, the model is up to 7× faster while demonstrating competitive performance. Additionally, the model reduces memory consumption during inference, previously needed for storing key-value (KV) cache, enabling better scaling to higher resolution image generation. The model has weaker reliance on the text at high-resolution scales. This observation allows to disable classifier-free guidance at the last two steps, resulting in further ∼20% acceleration and better generation of fine-grained details, as confirmed by human evaluation.
• SWIFTKV: FAST PREFILL-OPTIMIZED INFERENCE WITH KNOWLEDGE-PRESERVING MODEL TRANSFORMATION by Snowflake AI Research (https://arxiv.org/pdf/2410.03960).  The paper presents a model transformation and distillation procedure specifically designed to reduce the time and cost of processing prompt tokens while preserving the quality of generated tokens. The method combines three key mechanisms: i) SingleInputKV, which prefills later layers’ KV cache using a much earlier layer’s output, allowing prompt tokens to skip much of the model computation, ii) AcrossKV, which merges the KV caches of neighboring layers to reduce the memory footprint and support larger batch size for higher throughput, and iii) a knowledge-preserving distillation to recover the accuracy. For Llama-3.1-8B and 70B, the method reduces the compute requirement of prefill by 50% and the memory requirement of the KV cache by 62.5% while incurring minimum quality degradation across a wide range of tasks. Optimized models are available here.
• KV PREDICTION FOR IMPROVED TIME TO FIRST TOKEN by Apple (https://arxiv.org/pdf/2410.08391). In this method, a small auxiliary model is used to process the prompt and produce an approximation of the KV cache used by a base model. This approximated KV cache is then used with the base model for autoregressive generation without the need to query the auxiliary model again. Authors demonstrate that the method produces a pareto-optimal efficiency-accuracy trade-off when compared to baselines. On TriviaQA, they demonstrate relative accuracy improvements in the range of 15%−50% across a range of TTFT FLOPs budgets. They also demonstrate accuracy improvements of up to 30% on HumanEval python code completion at fixed TTFT FLOPs budgets. We release our code here.
• MAMBAEXTEND: A TRAINING-FREE APPROACH TO IMPROVE LONG-CONTEXT EXTENSION OF MAMBA (https://openreview.net/pdf?id=LgzRo1RpLS). The paper discloses the method that aims to extend the context length of SSM models, in particular Mamba family. The method leverages a training-free approach to calibrate only the scaling factors of discretization modules for different layers. Authors demonstrate both gradient-based and gradient-free zeroth-order optimization to learn the optimal scaling factors for each Mamba layer, requiring orders of magnitude fewer updates as opposed to the parameter fine-tuning-based alternatives. The method shows good accuracy on the Pile and Longbench benchmarks.
• Exploiting LLM Quantization by ETH Zurich (https://arxiv.org/pdf/2405.18137). A method which produces a malicious LLM from an original LLM. Malicious model performs similarly while in FP32 precision but malicious after the quantization. LLM -> malicious LLM -> Repairing malicious LLM via projected gradient descent subject to quantization blocks of the malicious LLM
• DEEP COMPRESSION AUTOENCODER FOR EFFICIENT HIGH-RESOLUTION DIFFUSION MODELS by MIT, Tsinghua University, and NVIDIA (https://arxiv.org/pdf/2410.10733). The proposed method is aimed to optimize image generation autoencoders by introducing two key techniques: (1) Residual Autoencoding, where authors design models to learn residuals based on the space-to-channel transformed features to alleviate the optimization difficulty of high spatial-compression autoencoders; (2) Decoupled High-Resolution Adaptation, a decoupled three-phase training strategy for mitigating the generalization penalty of high spatial-compression autoencoders. The method improves the autoencoder’s spatial compression ratio up to 128 while maintaining the reconstruction quality. Authors achieve significant speedup without accuracy drop. For example, on ImageNet 512 × 512, the model provides 19.1× inference speedup and 17.9× training speedup on H100 GPU for UViT-H while achieving a better FID. Code is available at: https://github.com/mit-han-lab/efficientvit.

• DUOATTENTION: EFFICIENT LONG-CONTEXT LLM INFERENCE WITH RETRIEVAL AND STREAMING HEADS by MIT, Tsinghua University, SJTU, University of Edinburgh, NVIDIA (https://arxiv.org/pdf/2410.10819). In this paper, authors identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens. In contrast, all other heads, which primarily focus on recent tokens and attention sinks–referred to as Streaming Heads–do not require full attention. They introduce a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM’s decoding and pre-filling memory and latency. DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately. The method reduces long-context inference memory by up to 2.55× for MHA and 1.67× for GQA models while speeding up decoding by up to 2.18× and 1.50× and accelerating pre-filling by up to 1.73× and 1.63× for MHA and GQA models, respectively. Code is available at: https://github.com/mit-han-lab/duo-attention.

‍

Software

• KV-COMPRESS: PAGED KV-CACHE COMPRESSION WITH VARIABLE COMPRESSION RATES PER ATTENTION HEAD by Cloudflare (https://arxiv.org/pdf/2410.00161). KV-Compress introduces a method to reduce the KV cache memory footprint by selectively compressing attention heads based on their importance. While early approaches measure KV importance by aggregating attention across all past queries, recent works show performance improvements by focusing on the final prompt tokens within a limited observation window. KV-Compress evicts contiguous KV blocks within a PagedAttention framework, reducing the memory footprint proportionally to the theoretical compression rate. Extending Ada-SnapKV, it supports per-layer and per-head variable compression rates, achieving state-of-the-art results on the LongBench suite. The "query-group-compression" technique further compresses the KV cache of GQA models without expanding it into the dimension of total query heads, achieving up to a 4x additional reduction. Integrated within vLLM, KV-Compress demonstrates the first end-to-end benchmarks of an eviction-based KV cache compression method within a paged-attention-enabled framework for efficient LLM inference. Code is available at https://github.com/IsaacRe/vllm-kvcompress.
• Introducing Machete, a Mixed-Input GEMM Kernel Optimized for NVIDIA Hopper GPUs.
• AMD released TensorCast, a casting/quantization PyTorch-based library to emulate various precisions: https://github.com/ROCm/tensorcast.
• MInference: Million-Tokens Prompt Inference for Long-context LLMs. A research project that is driven by Microsoft for a long-context text generation tasks. It contains implementation of several state-of-the-art methods.

‍