Shuaiwen Leon Song

Shuaiwen Leon Song, Bonnie Kruft, Minjia Zhang et al. · microsoft-research

In the upcoming decade, deep learning may revolutionize the natural sciences, enhancing our capacity to model and predict natural occurrences. This could herald a new era of scientific exploration, bringing significant advancements across sectors from drug development to renewable energy. To answer this call, we present DeepSpeed4Science initiative (deepspeed4science.ai) which aims to build unique capabilities through AI system technology innovations to help domain experts to unlock today's biggest science mysteries. By leveraging DeepSpeed's current technology pillars (training, inference and compression) as base technology enablers, DeepSpeed4Science will create a new set of AI system technologies tailored for accelerating scientific discoveries by addressing their unique complexity beyond the common technical approaches used for accelerating generic large language models (LLMs). In this paper, we showcase the early progress we made with DeepSpeed4Science in addressing two of the critical system challenges in structural biology research.

Zhewei Yao, Reza Yazdani Aminabadi, Olatunji Ruwase et al.

ChatGPT-like models have revolutionized various applications in artificial intelligence, from summarization and coding to translation, matching or even surpassing human performance. However, the current landscape lacks an accessible, efficient, and cost-effective end-to-end RLHF (Reinforcement Learning with Human Feedback) training pipeline for these powerful models, particularly when training at the scale of billions of parameters. This paper introduces DeepSpeed-Chat, a novel system that democratizes RLHF training, making it accessible to the AI community. DeepSpeed-Chat offers three key capabilities: an easy-to-use training and inference experience for ChatGPT-like models, a DeepSpeed-RLHF pipeline that replicates the training pipeline from InstructGPT, and a robust DeepSpeed-RLHF system that combines various optimizations for training and inference in a unified way. The system delivers unparalleled efficiency and scalability, enabling training of models with hundreds of billions of parameters in record time and at a fraction of the cost. With this development, DeepSpeed-Chat paves the way for broader access to advanced RLHF training, even for data scientists with limited resources, thereby fostering innovation and further development in the field of AI.

Sam Ade Jacobs, Masahiro Tanaka, Chengming Zhang et al.

Computation in a typical Transformer-based large language model (LLM) can be characterized by batch size, hidden dimension, number of layers, and sequence length. Until now, system works for accelerating LLM training have focused on the first three dimensions: data parallelism for batch size, tensor parallelism for hidden size and pipeline parallelism for model depth or layers. These widely studied forms of parallelism are not targeted or optimized for long sequence Transformer models. Given practical application needs for long sequence LLM, renewed attentions are being drawn to sequence parallelism. However, existing works in sequence parallelism are constrained by memory-communication inefficiency, limiting their scalability to long sequence large models. In this work, we introduce DeepSpeed-Ulysses, a novel, portable and effective methodology for enabling highly efficient and scalable LLM training with extremely long sequence length. DeepSpeed-Ulysses at its core partitions input data along the sequence dimension and employs an efficient all-to-all collective communication for attention computation. Theoretical communication analysis shows that whereas other methods incur communication overhead as sequence length increases, DeepSpeed-Ulysses maintains constant communication volume when sequence length and compute devices are increased proportionally. Furthermore, experimental evaluations show that DeepSpeed-Ulysses trains 2.5x faster with 4x longer sequence length than the existing method SOTA baseline.

Haojun Xia, Zhen Zheng, Yuchao Li et al.

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

Chengming Zhang, Shaden Smith, Baixi Sun et al.

Collaborative filtering (CF) has been proven to be one of the most effective techniques for recommendation. Among all CF approaches, SimpleX is the state-of-the-art method that adopts a novel loss function and a proper number of negative samples. However, there is no work that optimizes SimpleX on multi-core CPUs, leading to limited performance. To this end, we perform an in-depth profiling and analysis of existing SimpleX implementations and identify their performance bottlenecks including (1) irregular memory accesses, (2) unnecessary memory copies, and (3) redundant computations. To address these issues, we propose an efficient CF training system (called HEAT) that fully enables the multi-level caching and multi-threading capabilities of modern CPUs. Specifically, the optimization of HEAT is threefold: (1) It tiles the embedding matrix to increase data locality and reduce cache misses (thus reduces read latency); (2) It optimizes stochastic gradient descent (SGD) with sampling by parallelizing vector products instead of matrix-matrix multiplications, in particular the similarity computation therein, to avoid memory copies for matrix data preparation; and (3) It aggressively reuses intermediate results from the forward phase in the backward phase to alleviate redundant computation. Evaluation on five widely used datasets with both x86- and ARM-architecture processors shows that HEAT achieves up to 45.2X speedup over existing CPU solution and 4.5X speedup and 7.9X cost reduction in Cloud over existing GPU solution with NVIDIA V100 GPU.

Yifan Yu, Yuqing Jian, Junxiong Wang et al.

Diffusion language models promise parallel generation, yet still lag behind autoregressive (AR) models in quality. We stem this gap to a failure of introspective consistency: AR models agree with their own generations, while DLMs often do not. We define the introspective acceptance rate, which measures whether a model accepts its previously generated tokens. This reveals why AR training has a structural advantage: causal masking and logit shifting implicitly enforce introspective consistency. Motivated by this observation, we introduce Introspective Diffusion Language Model (I-DLM), a paradigm that retains diffusion-style parallel decoding while inheriting the introspective consistency of AR training. I-DLM uses a novel introspective strided decoding (ISD) algorithm, which enables the model to verify previously generated tokens while advancing new ones in the same forward pass. From a systems standpoint, we build I-DLM inference engine on AR-inherited optimizations and further customize it with a stationary-batch scheduler. To the best of our knowledge, I-DLM is the first DLM to match the quality of its same-scale AR counterpart while outperforming prior DLMs in both model quality and practical serving efficiency across 15 benchmarks. It reaches 69.6 on AIME-24 and 45.7 on LiveCodeBench-v6, exceeding LLaDA-2.1-mini (16B) by more than 26 and 15 points, respectively. Beyond quality, I-DLM is designed for the growing demand of large-concurrency serving, delivering about 3x higher throughput than prior state-of-the-art DLMs.

Jinda Jia, Jisen Li, Zhongzhu Zhou et al.

KV-cache memory is a major bottleneck in real-world LLM serving, where systems must simultaneously support latency-sensitive small-batch requests and high-throughput concurrent workloads. Although many KV-cache compression methods improve offline accuracy or compression ratio, they often violate practical serving constraints such as paged memory layouts, regular memory access, and fused attention execution, limiting their effectiveness in deployment. In this work, we identify the minimal set of 4-bit KV-cache quantization methods that remain viable under these constraints. Our central finding is that a simple design--token-wise INT4 quantization with block-diagonal Hadamard rotation--consistently achieves the best accuracy-efficiency trade-off. Across multiple models and benchmarks, this approach recovers nearly all of the accuracy lost by naive INT4, while more complex methods such as vector quantization and Hessian-aware quantization provide only marginal additional gains once serving compatibility is taken into account. To make this practical, we implement a fused rotation-quantization kernel that integrates directly into paged KV-cache layouts and introduces zero measurable end-to-end overhead, matching plain INT4 throughput across concurrency levels. Our results show that effective KV-cache compression is fundamentally a systems co-design problem: under real serving constraints, lightweight block-diagonal Hadamard rotation is a viable method that delivers near-lossless accuracy without sacrificing serving efficiency.

Zhongzhu Zhou, Fengxiang Bie, Ziyan Chen et al.

Converting pretrained attention modules such as grouped-query attention (GQA) into multi-head latent attention (MLA) can improve expressivity without increasing KV-cache cost, making it attractive for efficient inference. However, many practical conversion baselines rely on weight-only low-rank approximations (e.g., SVD-style initializations) and uniform rank allocation. They focus on minimizing the difference between weight matrices rather than on how those weights affect input activations, ignore the covariance structure of activations, and enforce uniform rank across layers, causing activation drift and degraded attention fidelity. To address these issues, we propose CARE, a Covariance-Aware, Rank-Enhanced MLA conversion pipeline under a fixed KV width. CARE introduces three key steps: (i) activation-preserving factorization, which aligns the approximation with the actual input activations rather than just the weights; (ii) adjusted-rank allocation, which spreads a fixed KV budget across layers by giving more capacity to layers that need it most; and (iii) KV-parity mapping, which reparameterizes the converted K and V to fit the MLA format while keeping the KV-cache size unchanged. Our method outperforms a uniform-rank SVD baseline on Qwen3-4B/30B-A3B-Instruct-2507 and Llama-3.1-8B/70B-Instruct, reducing one-shot perplexity by up to 215x and improving mean accuracy by up to 1.70x at matched KV budgets. With a brief post-SVD healing fine-tune, we fully recover the original model's accuracy.

Zhongzhu Zhou, Donglin Zhuang, Jisen Li et al.

INT2 KV-cache quantization is attractive for long-context LLM serving, but it remains difficult to make both accurate and deployable. Simple rotations such as Hadamard transforms reduce outliers, but still degrade at INT2 because they are not aligned with downstream attention. We propose OSCAR, an Ultra-low-bit KV Cache quantization method that estimates attention-aware covariance structures offline and uses them to derive fixed rotations and clipping thresholds for quantization. In this way, it aligns KV quantization with the covariance structures that attention actually consumes. More importantly, we not only provide theoretical justification but also develop a fully deployable OSCAR system with a custom INT2 attention kernel that remains compatible with paged KV-cache serving and fused kernel pipelines, enabling seamless integration into modern LLM serving frameworks such as SGLang and vLLM. We evaluate our methods on recent reasoning models with reasoning traces of up to 32k tokens across 5 tasks. On Qwen3-4B-Thinking-2507 and Qwen3-8B, OSCAR reduces the BF16 accuracy gap to 3.78 and 1.42 points, respectively, while naive rotation INT2 collapses to nearly zero. We further scale OSCAR to Qwen3-32B and GLM-4.7 (358B params), where it remains effectively on par with BF16. On long context - RULER-NIAH up to 128K, OSCAR remains robust on both Qwen3 models, while naive rotation INT2 collapses. System-wise, OSCAR reduces KV-cache memory by approximately 8x, improves throughput by up to 7x at large batch sizes under the same memory budget, and accelerates batch-size-1 decoding by up to 3x over BF16 due to reduced memory bandwidth overhead.

Junlin Wang, Roy Xie, Shang Zhu et al.

Building helpful and harmless large language models (LLMs) requires effective model alignment approach based on human instructions and feedback, which necessitates high-quality human-labeled data. Constructing such datasets is often expensive and hard to scale, and may face potential limitations on diversity and generalization. To address these challenges, we introduce Mixture of Agents Alignment (MoAA), that leverages the collective strengths of various language models to provide high-quality data for model alignment. By employing MoAA, we enhance both supervised fine-tuning and preference optimization, leading to improved performance compared to using a single model alone to generate alignment data (e.g. using GPT-4o alone). Evaluation results show that our approach can improve win rate of LLaMA-3.1-8B-Instruct from 19.5 to 48.3 on Arena-Hard and from 22.33 to 57.23 on AlpacaEval2, highlighting a promising direction for model alignment through this new scalable and diverse synthetic data recipe. Furthermore, we demonstrate that MoAA enables a self-improvement pipeline, where models finetuned on MoA-generated data surpass their own initial capabilities, providing evidence that our approach can push the frontier of open-source LLMs without reliance on stronger external supervision. Data and code will be released.

Zhenyu Zhang, Xiaoxia Wu, Zhongzhu Zhou et al.

Large Language Models (LLMs) often rely on long chain-of-thought (CoT) reasoning to solve complex tasks. While effective, these trajectories are frequently inefficient, leading to high latency from excessive token generation, or unstable reasoning that alternates between underthinking (shallow, inconsistent steps) and overthinking (repetitive, verbose reasoning). In this work, we study the structure of reasoning trajectories and uncover specialized attention heads that correlate with distinct cognitive behaviors such as verification and backtracking. By lightly intervening on these heads at inference time, we can steer the model away from inefficient modes. Building on this insight, we propose CREST, a training-free method for Cognitive REasoning Steering at Test-time. CREST has two components: (1) an offline calibration step that identifies cognitive heads and derives head-specific steering vectors, and (2) an inference-time procedure that rotates hidden representations to suppress components along those vectors. CREST adaptively suppresses unproductive reasoning behaviors, yielding both higher accuracy and lower computational cost. Across diverse reasoning benchmarks and models, CREST improves accuracy by up to 17.5% while reducing token usage by 37.6%, offering a simple and effective pathway to faster, more reliable LLM reasoning.

Junxiong Wang, Fengxiang Bie, Jisen Li et al.

Speculative decoding can significantly accelerate LLM serving, yet most deployments today disentangle speculator training from serving, treating speculator training as a standalone offline modeling problem. We show that this decoupled formulation introduces substantial deployment and adaptation lag: (1) high time-to-serve, since a speculator must be trained offline for a considerable period before deployment; (2) delayed utility feedback, since the true end-to-end decoding speedup is only known after training and cannot be inferred reliably from acceptance rate alone due to model-architecture and system-level overheads; and (3) domain-drift degradation, as the target model is repurposed to new domains and the speculator becomes stale and less effective. To address these issues, we present Aurora, a unified training-serving system that closes the loop by continuously learning a speculator directly from live inference traces. Aurora reframes online speculator learning as an asynchronous reinforcement-learning problem: accepted tokens provide positive feedback, while rejected speculator proposals provide implicit negative feedback that we exploit to improve sample efficiency. Our design integrates an SGLang-based inference server with an asynchronous training server, enabling hot-swapped speculator updates without service interruption. Crucially, Aurora supports day-0 deployment: a speculator can be served immediately and rapidly adapted to live traffic, improving system performance while providing immediate utility feedback. Across experiments, Aurora achieves a 1.5x day-0 speedup on recently released frontier models (e.g., MiniMax M2.1 229B and Qwen3-Coder-Next 80B). Aurora also adapts effectively to distribution shifts in user traffic, delivering an additional 1.25x speedup over a well-trained but static speculator on widely used models (e.g., Qwen3 and Llama3).

Muru Zhang, Mayank Mishra, Zhongzhu Zhou et al.

Large language model inference is both memory-intensive and time-consuming, often requiring distributed algorithms to efficiently scale. Various model parallelism strategies are used in multi-gpu training and inference to partition computation across multiple devices, reducing memory load and computation time. However, using model parallelism necessitates communication of information between GPUs, which has been a major bottleneck and limits the gains obtained by scaling up the number of devices. We introduce Ladder Residual, a simple architectural modification applicable to all residual-based models that enables straightforward overlapping that effectively hides the latency of communication. Our insight is that in addition to systems optimization, one can also redesign the model architecture to decouple communication from computation. While Ladder Residual can allow communication-computation decoupling in conventional parallelism patterns, we focus on Tensor Parallelism in this paper, which is particularly bottlenecked by its heavy communication. For a Transformer model with 70B parameters, applying Ladder Residual to all its layers can achieve 29% end-to-end wall clock speed up at inference time with TP sharding over 8 devices. We refer the resulting Transformer model as the Ladder Transformer. We train a 1B and 3B Ladder Transformer from scratch and observe comparable performance to a standard dense transformer baseline. We also show that it is possible to convert parts of the Llama-3.1 8B model to our Ladder Residual architecture with minimal accuracy degradation by only retraining for 3B tokens. We release our code for training and inference for easier replication of experiments.

Haojun Xia, Xiaoxia Wu, Jisen Li et al.

The KV cache is a dominant memory bottleneck for LLM inference. While 4-bit KV quantization preserves accuracy, 2-bit often degrades it, especially on long-context reasoning. We close this gap via an algorithm-system co-design for mixed-precision KV caching: Kitty. On the algorithm side, extensive experiments show that Dynamic Channel-wise Precision Boost -- which ranks Key-cache channels by sensitivity and keeps only a small fraction at higher precision -- maintains near-zero loss in accuracy drop while approaching 2-bit memory. The main challenge is handling dynamic 4-bit channel boosts while keeping the page layout coalesced and the dequantization uniform, with no scattered reads or hard-coded masks. Kitty addresses these issues by decompose each mixed-precision Key page into two tensors with unified 2-bit precision. Based on this, Kitty provides a page-centric KV layout, Triton-compatible page dequantization kernels, and a lightweight runtime pipeline that preserves coalescing and avoids divergence. Across seven tasks and two model families (Qwen3, LLaMA3), Kitty cuts KV memory by nearly 8x with negligible accuracy loss, enabling up to 8x larger batches and 2.1x-4.1x higher throughput under the same memory budget. We release the full implementation of Kitty at https://github.com/Summer-Summer/Kitty.

Haojun Xia, Zhen Zheng, Xiaoxia Wu et al.

Six-bit quantization (FP6) can effectively reduce the size of large language models (LLMs) and preserve the model quality consistently across varied applications. However, existing systems do not provide Tensor Core support for FP6 quantization and struggle to achieve practical performance improvements during LLM inference. It is challenging to support FP6 quantization on GPUs due to (1) unfriendly memory access of model weights with irregular bit-width and (2) high runtime overhead of weight de-quantization. To address these problems, we propose TC-FPx, the first full-stack GPU kernel design scheme with unified Tensor Core support of float-point weights for various quantization bit-width. We integrate TC-FPx kernel into an existing inference system, providing new end-to-end support (called FP6-LLM) for quantized LLM inference, where better trade-offs between inference cost and model quality are achieved. Experiments show that FP6-LLM enables the inference of LLaMA-70b using only a single GPU, achieving 1.69x-2.65x higher normalized inference throughput than the FP16 baseline. The source code is publicly available at https://github.com/usyd-fsalab/fp6_llm.

Donglin Zhuang, Xingyao Zhang, Shuaiwen Leon Song et al.

The quest for determinism in machine learning has disproportionately focused on characterizing the impact of noise introduced by algorithmic design choices. In this work, we address a less well understood and studied question: how does our choice of tooling introduce randomness to deep neural network training. We conduct large scale experiments across different types of hardware, accelerators, state of art networks, and open-source datasets, to characterize how tooling choices contribute to the level of non-determinism in a system, the impact of said non-determinism, and the cost of eliminating different sources of noise. Our findings are surprising, and suggest that the impact of non-determinism in nuanced. While top-line metrics such as top-1 accuracy are not noticeably impacted, model performance on certain parts of the data distribution is far more sensitive to the introduction of randomness. Our results suggest that deterministic tooling is critical for AI safety. However, we also find that the cost of ensuring determinism varies dramatically between neural network architectures and hardware types, e.g., with overhead up to $746\%$, $241\%$, and $196\%$ on a spectrum of widely used GPU accelerator architectures, relative to non-deterministic training. The source code used in this paper is available at https://github.com/usyd-fsalab/NeuralNetworkRandomness.

Junlin Wang, Shang Zhu, Jon Saad-Falcon et al.

There is intense interest in investigating how inference time compute (ITC) (e.g. repeated sampling, refinements, etc) can improve large language model (LLM) capabilities. At the same time, recent breakthroughs in reasoning models, such as Deepseek-R1, unlock the opportunity for reinforcement learning to improve LLM reasoning skills. An in-depth understanding of how ITC interacts with reasoning across different models could provide important guidance on how to further advance the LLM frontier. This work conducts a comprehensive analysis of inference-time scaling methods for both reasoning and non-reasoning models on challenging reasoning tasks. Specifically, we focus our research on verifier-free inference time-scaling methods due to its generalizability without needing a reward model. We construct the Pareto frontier of quality and efficiency. We find that non-reasoning models, even with an extremely high inference budget, still fall substantially behind reasoning models. For reasoning models, majority voting proves to be a robust inference strategy, generally competitive or outperforming other more sophisticated ITC methods like best-of-N and sequential revisions, while the additional inference compute offers minimal improvements. We further perform in-depth analyses of the association of key response features (length and linguistic markers) with response quality, with which we can improve the existing ITC methods. We find that correct responses from reasoning models are typically shorter and have fewer hedging and thinking markers (but more discourse markers) than the incorrect responses.

Rahul Thapa, Andrew Li, Qingyang Wu et al.

Publicly available biomedical videos, such as those on YouTube, serve as valuable educational resources for medical students. Unlike standard machine learning datasets, these videos are designed for human learners, often mixing medical imagery with narration, explanatory diagrams, and contextual framing. In this work, we investigate whether such pedagogically rich, yet non-standardized and heterogeneous videos can effectively teach general-domain vision-language models biomedical knowledge. To this end, we introduce OpenBiomedVi, a biomedical video instruction tuning dataset comprising 1031 hours of video-caption and Q/A pairs, curated through a multi-step human-in-the-loop pipeline. Diverse biomedical video datasets are rare, and OpenBiomedVid fills an important gap by providing instruction-style supervision grounded in real-world educational content. Surprisingly, despite the informal and heterogeneous nature of these videos, the fine-tuned Qwen-2-VL models exhibit substantial performance improvements across most benchmarks. The 2B model achieves gains of 98.7% on video tasks, 71.2% on image tasks, and 0.2% on text tasks. The 7B model shows improvements of 37.09% on video and 11.2% on image tasks, with a slight degradation of 2.7% on text tasks compared to their respective base models. To address the lack of standardized biomedical video evaluation datasets, we also introduce two new expert curated benchmarks, MIMICEchoQA and SurgeryVideoQA. On these benchmarks, the 2B model achieves gains of 99.1% and 98.1%, while the 7B model shows gains of 22.5% and 52.1%, respectively, demonstrating the models' ability to generalize and perform biomedical video understanding on cleaner and more standardized datasets than those seen during training. These results suggest that educational videos created for human learning offer a surprisingly effective training signal for biomedical VLMs.

Zhongzhu Zhou, Yibo Yang, Ziyan Chen et al.

Policy gradient (PG) methods in reinforcement learning frequently utilize deep neural networks (DNNs) to learn a shared backbone of feature representations used to compute likelihoods in an action selection layer. Numerous studies have been conducted on the convergence and global optima of policy networks, but few have analyzed representational structures of those underlying networks. While training an optimal policy DNN, we observed that under certain constraints, a gentle structure resembling neural collapse, which we refer to as Action Collapse (AC), emerges. This suggests that 1) the state-action activations (i.e. last-layer features) sharing the same optimal actions collapse towards those optimal actions respective mean activations; 2) the variability of activations sharing the same optimal actions converges to zero; 3) the weights of action selection layer and the mean activations collapse to a simplex equiangular tight frame (ETF). Our early work showed those aforementioned constraints to be necessary for these observations. Since the collapsed ETF of optimal policy DNNs maximally separates the pair-wise angles of all actions in the state-action space, we naturally raise a question: can we learn an optimal policy using an ETF structure as a (fixed) target configuration in the action selection layer? Our analytical proof shows that learning activations with a fixed ETF as action selection layer naturally leads to the AC. We thus propose the Action Collapse Policy Gradient (ACPG) method, which accordingly affixes a synthetic ETF as our action selection layer. ACPG induces the policy DNN to produce such an ideal configuration in the action selection layer while remaining optimal. Our experiments across various OpenAI Gym environments demonstrate that our technique can be integrated into any discrete PG methods and lead to favorable reward improvements more quickly and robustly.

Berkan Dokmeci, Qingyang Wu, Ben Athiwaratkun et al.

While recent advances in preference learning have enhanced alignment in human feedback, mathematical reasoning remains a persistent challenge. We investigate how data diversification strategies in preference optimization can improve the mathematical reasoning abilities of large language models (LLMs). We evaluate three common data generation methods: temperature sampling, Chain-of-Thought prompting, and Monte Carlo Tree Search (MCTS), and introduce Diversified-ThinkSolve (DTS), a novel structured approach that systematically decomposes problems into diverse reasoning paths. Our results show that with strategically diversified preference data, models can substantially improve mathematical reasoning performance, with the best approach yielding gains of 7.1% on GSM8K and 4.2% on MATH over the base model. Despite its strong performance, DTS incurs only a marginal computational overhead (1.03x) compared to the baseline, while MCTS is nearly five times more costly with lower returns. These findings demonstrate that structured exploration of diverse problem-solving methods creates more effective preference data for mathematical alignment than traditional approaches.

Yibo Yang, Xiaojie Li, Zhongzhu Zhou et al.

Current parameter-efficient fine-tuning (PEFT) methods build adapters widely agnostic of the context of downstream task to learn, or the context of important knowledge to maintain. As a result, there is often a performance gap compared to full-parameter fine-tuning, and meanwhile the fine-tuned model suffers from catastrophic forgetting of the pre-trained world knowledge. In this paper, we propose CorDA, a Context-oriented Decomposition Adaptation method that builds learnable task-aware adapters from weight decomposition oriented by the context of downstream task or the world knowledge to maintain. Concretely, we collect a few data samples, and perform singular value decomposition for each linear layer of a pre-trained LLM multiplied by the covariance matrix of the input activation using these samples. The inverse of the covariance matrix is multiplied with the decomposed components to reconstruct the original weights. By doing so, the context of the representative samples is captured through deciding the factorizing orientation. Our method enables two options, the knowledge-preserved adaptation and the instruction-previewed adaptation. For the former, we use question-answering samples to obtain the covariance matrices, and use the decomposed components with the smallest $r$ singular values to initialize a learnable adapter, with the others frozen such that the world knowledge is better preserved. For the latter, we use the instruction data from the fine-tuning task, such as math or coding, to orientate the decomposition and train the largest $r$ components that most correspond to the task to learn. We conduct extensive experiments on Math, Code, and Instruction Following tasks.

Rahul Thapa, Kezhen Chen, Ian Covert et al.

Recent advances in vision-language models (VLMs) have demonstrated the advantages of processing images at higher resolutions and utilizing multi-crop features to preserve native resolution details. However, despite these improvements, existing vision transformers (ViTs) still struggle to capture fine-grained details from less prominent objects, charts, and embedded text, limiting their effectiveness in certain tasks. In this paper, we extend recent high-resolution and multi-crop techniques by not only preserving the native resolution, but zooming in beyond it and extracting features from a large number of image sub-crops. This enhancement allows our model to better capture fine-grained details, overcoming the limitations of current ViTs. To manage the increased token count and computational complexity, we demonstrate that a simple mean-pooling aggregation over tokens is effective. Our model, Dragonfly, achieves competitive performance on general-domain tasks such as ScienceQA and AI2D, and excels in tasks requiring fine-grained image understanding, including TextVQA and ChartQA. Among models in the 7-8B parameter range, Dragonfly consistently ranks at the top across ten general-domain benchmarks, achieving the highest or second-highest scores in most cases, outperforming models that are significantly larger or trained on larger datasets. Our biomedical model, Dragonfly-Med, sets new benchmarks on several medical tasks, achieving 91.6% accuracy on SLAKE (compared to 84.8% for Med-Gemini), a 67.1% token F1 score on Path-VQA (compared to 62.7% for Med-PaLM M), and state-of-the-art results across the majority of image captioning tasks. Overall, our work highlights the persistent challenge of engineering visual representations with fixed-resolution ViTs, and proposes a simple yet effective solution to address this issue and boost performance in both general and specialized domains.

Fengxiang Bie, Yibo Yang, Zhongzhu Zhou et al.

Text-to-image generation (TTI) refers to the usage of models that could process text input and generate high fidelity images based on text descriptions. Text-to-image generation using neural networks could be traced back to the emergence of Generative Adversial Network (GAN), followed by the autoregressive Transformer. Diffusion models are one prominent type of generative model used for the generation of images through the systematic introduction of noises with repeating steps. As an effect of the impressive results of diffusion models on image synthesis, it has been cemented as the major image decoder used by text-to-image models and brought text-to-image generation to the forefront of machine-learning (ML) research. In the era of large models, scaling up model size and the integration with large language models have further improved the performance of TTI models, resulting the generation result nearly indistinguishable from real-world images, revolutionizing the way we retrieval images. Our explorative study has incentivised us to think that there are further ways of scaling text-to-image models with the combination of innovative model architectures and prediction enhancement techniques. We have divided the work of this survey into five main sections wherein we detail the frameworks of major literature in order to delve into the different types of text-to-image generation methods. Following this we provide a detailed comparison and critique of these methods and offer possible pathways of improvement for future work. In the future work, we argue that TTI development could yield impressive productivity improvements for creation, particularly in the context of the AIGC era, and could be extended to more complex tasks such as video generation and 3D generation.

Sian Jin, Chengming Zhang, Xintong Jiang et al.

Training wide and deep neural networks (DNNs) require large amounts of storage resources such as memory because the intermediate activation data must be saved in the memory during forward propagation and then restored for backward propagation. However, state-of-the-art accelerators such as GPUs are only equipped with very limited memory capacities due to hardware design constraints, which significantly limits the maximum batch size and hence performance speedup when training large-scale DNNs. Traditional memory saving techniques either suffer from performance overhead or are constrained by limited interconnect bandwidth or specific interconnect technology. In this paper, we propose a novel memory-efficient CNN training framework (called COMET) that leverages error-bounded lossy compression to significantly reduce the memory requirement for training, to allow training larger models or to accelerate training. Different from the state-of-the-art solutions that adopt image-based lossy compressors (such as JPEG) to compress the activation data, our framework purposely adopts error-bounded lossy compression with a strict error-controlling mechanism. Specifically, we perform a theoretical analysis on the compression error propagation from the altered activation data to the gradients, and empirically investigate the impact of altered gradients over the training process. Based on these analyses, we optimize the error-bounded lossy compression and propose an adaptive error-bound control scheme for activation data compression. We evaluate our design against state-of-the-art solutions with five widely-adopted CNNs and ImageNet dataset. Experiments demonstrate that our proposed framework can significantly reduce the training memory consumption by up to 13.5X over the baseline training and 1.8X over another state-of-the-art compression-based framework, respectively, with little or no accuracy loss.

Qiyu Wan, Haojun Xia, Xingyao Zhang et al.

Bayesian Neural Networks (BNNs) that possess a property of uncertainty estimation have been increasingly adopted in a wide range of safety-critical AI applications which demand reliable and robust decision making, e.g., self-driving, rescue robots, medical image diagnosis. The training procedure of a probabilistic BNN model involves training an ensemble of sampled DNN models, which induces orders of magnitude larger volume of data movement than training a single DNN model. In this paper, we reveal that the root cause for BNN training inefficiency originates from the massive off-chip data transfer by Gaussian Random Variables (GRVs). To tackle this challenge, we propose a novel design that eliminates all the off-chip data transfer by GRVs through the reversed shifting of Linear Feedback Shift Registers (LFSRs) without incurring any training accuracy loss. To efficiently support our LFSR reversion strategy at the hardware level, we explore the design space of the current DNN accelerators and identify the optimal computation mapping scheme to best accommodate our strategy. By leveraging this finding, we design and prototype the first highly efficient BNN training accelerator, named Shift-BNN, that is low-cost and scalable. Extensive evaluation on five representative BNN models demonstrates that Shift-BNN achieves an average of 4.9x (up to 10.8x) boost in energy efficiency and 1.6x (up to 2.8x) speedup over the baseline DNN training accelerator.

Anil Gaihre, Da Zheng, Scott Weitze et al.

Recent top-$k$ computation efforts explore the possibility of revising various sorting algorithms to answer top-$k$ queries on GPUs. These endeavors, unfortunately, perform significantly more work than needed. This paper introduces Dr. Top-k, a Delegate-centric top-$k$ system on GPUs that can reduce the top-$k$ workloads significantly. Particularly, it contains three major contributions: First, we introduce a comprehensive design of the delegate-centric concept, including maximum delegate, delegate-based filtering, and $β$ delegate mechanisms to help reduce the workload for top-$k$ up to more than 99%. Second, due to the difficulty and importance of deriving a proper subrange size, we perform a rigorous theoretical analysis, coupled with thorough experimental validations to identify the desirable subrange size. Third, we introduce four key system optimizations to enable fast multi-GPU top-$k$ computation. Taken together, this work constantly outperforms the state-of-the-art.

Chengming Zhang, Geng Yuan, Wei Niu et al.

Convolutional neural networks (CNNs) are becoming increasingly deeper, wider, and non-linear because of the growing demand on prediction accuracy and analysis quality. The wide and deep CNNs, however, require a large amount of computing resources and processing time. Many previous works have studied model pruning to improve inference performance, but little work has been done for effectively reducing training cost. In this paper, we propose ClickTrain: an efficient and accurate end-to-end training and pruning framework for CNNs. Different from the existing pruning-during-training work, ClickTrain provides higher model accuracy and compression ratio via fine-grained architecture-preserving pruning. By leveraging pattern-based pruning with our proposed novel accurate weight importance estimation, dynamic pattern generation and selection, and compiler-assisted computation optimizations, ClickTrain generates highly accurate and fast pruned CNN models for direct deployment without any extra time overhead, compared with the baseline training. ClickTrain also reduces the end-to-end time cost of the pruning-after-training method by up to 2.3X with comparable accuracy and compression ratio. Moreover, compared with the state-of-the-art pruning-during-training approach, ClickTrain provides significant improvements both accuracy and compression ratio on the tested CNN models and datasets, under similar limited training time.

Sian Jin, Guanpeng Li, Shuaiwen Leon Song et al.

Deep neural networks (DNNs) are becoming increasingly deeper, wider, and non-linear due to the growing demands on prediction accuracy and analysis quality. When training a DNN model, the intermediate activation data must be saved in the memory during forward propagation and then restored for backward propagation. However, state-of-the-art accelerators such as GPUs are only equipped with very limited memory capacities due to hardware design constraints, which significantly limits the maximum batch size and hence performance speedup when training large-scale DNNs. In this paper, we propose a novel memory-driven high performance DNN training framework that leverages error-bounded lossy compression to significantly reduce the memory requirement for training in order to allow training larger networks. Different from the state-of-the-art solutions that adopt image-based lossy compressors such as JPEG to compress the activation data, our framework purposely designs error-bounded lossy compression with a strict error-controlling mechanism. Specifically, we provide theoretical analysis on the compression error propagation from the altered activation data to the gradients, and then empirically investigate the impact of altered gradients over the entire training process. Based on these analyses, we then propose an improved lossy compressor and an adaptive scheme to dynamically configure the lossy compression error-bound and adjust the training batch size to further utilize the saved memory space for additional speedup. We evaluate our design against state-of-the-art solutions with four popular DNNs and the ImageNet dataset. Results demonstrate that our proposed framework can significantly reduce the training memory consumption by up to 13.5x and 1.8x over the baseline training and state-of-the-art framework with compression, respectively, with little or no accuracy loss.

Xingyao Zhang, Shuaiwen Leon Song, Chenhao Xie et al.

In recent years, the CNNs have achieved great successes in the image processing tasks, e.g., image recognition and object detection. Unfortunately, traditional CNN's classification is found to be easily misled by increasingly complex image features due to the usage of pooling operations, hence unable to preserve accurate position and pose information of the objects. To address this challenge, a novel neural network structure called Capsule Network has been proposed, which introduces equivariance through capsules to significantly enhance the learning ability for image segmentation and object detection. Due to its requirement of performing a high volume of matrix operations, CapsNets have been generally accelerated on modern GPU platforms that provide highly optimized software library for common deep learning tasks. However, based on our performance characterization on modern GPUs, CapsNets exhibit low efficiency due to the special program and execution features of their routing procedure, including massive unshareable intermediate variables and intensive synchronizations, which are very difficult to optimize at software level. To address these challenges, we propose a hybrid computing architecture design named \textit{PIM-CapsNet}. It preserves GPU's on-chip computing capability for accelerating CNN types of layers in CapsNet, while pipelining with an off-chip in-memory acceleration solution that effectively tackles routing procedure's inefficiency by leveraging the processing-in-memory capability of today's 3D stacked memory. Using routing procedure's inherent parallellization feature, our design enables hierarchical improvements on CapsNet inference efficiency through minimizing data movement and maximizing parallel processing in memory.

Linnan Wang, Jinmian Ye, Yiyang Zhao et al.

Going deeper and wider in neural architectures improves the accuracy, while the limited GPU DRAM places an undesired restriction on the network design domain. Deep Learning (DL) practitioners either need change to less desired network architectures, or nontrivially dissect a network across multiGPUs. These distract DL practitioners from concentrating on their original machine learning tasks. We present SuperNeurons: a dynamic GPU memory scheduling runtime to enable the network training far beyond the GPU DRAM capacity. SuperNeurons features 3 memory optimizations, \textit{Liveness Analysis}, \textit{Unified Tensor Pool}, and \textit{Cost-Aware Recomputation}, all together they effectively reduce the network-wide peak memory usage down to the maximal memory usage among layers. We also address the performance issues in those memory saving techniques. Given the limited GPU DRAM, SuperNeurons not only provisions the necessary memory for the training, but also dynamically allocates the memory for convolution workspaces to achieve the high performance. Evaluations against Caffe, Torch, MXNet and TensorFlow have demonstrated that SuperNeurons trains at least 3.2432 deeper network than current ones with the leading performance. Particularly, SuperNeurons can train ResNet2500 that has $10^4$ basic network layers on a 12GB K40c.