PyTorchKR Papers

최신 논문

총 2,230편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2405.02267공식 코드 · Papers with Code

Structural Pruning of Pre-trained Language Models via Neural Architecture Search

Pre-trained language models (PLM), for example BERT or RoBERTa, mark the state-of-the-art for natural language understanding task when fine-tuned on labeled data. However, their large size poses challenges in deploying them for inference in real-world applications, due to significant GPU memory requirements and high inference latency. This paper explores neural architecture search (NAS) for structural pruning to find sub-parts of the fine-tuned network that optimally trade-off efficiency, for example in terms of model size or latency, and generalization performance. We also show how we can utilize more recently developed two-stage weight-sharing NAS approaches in this setting to accelerate the search process. Unlike traditional pruning methods with fixed thresholds, we propose to adopt a multi-objective approach that identifies the Pareto optimal set of sub-networks, allowing for a more flexible and automated compression process.

arXiv:2405.00738공식 코드 · Papers with Code

HLSTransform: Energy-Efficient Llama 2 Inference on FPGAs Via High Level Synthesis

Graphics Processing Units (GPUs) have become the leading hardware accelerator for deep learning applications and are used widely in training and inference of transformers; transformers have achieved state-of-the-art performance in many areas of machine learning and are especially used in most modern Large Language Models (LLMs). However, GPUs require large amounts of energy, which poses environmental concerns, demands high operational costs, and causes GPUs to be unsuitable for edge computing. We develop an accelerator for transformers, namely, Llama 2, an open-source state-of-the-art LLM, using high level synthesis (HLS) on Field Programmable Gate Arrays (FPGAs). HLS allows us to rapidly prototype FPGA designs without writing code at the register-transfer level (RTL). We name our method HLSTransform, and the FPGA designs we synthesize with HLS achieve up to a 12.75x reduction and 8.25x reduction in energy used per token on the Xilinx Virtex UltraScale+ VU9P FPGA compared to an Intel Xeon Broadwell E5-2686 v4 CPU and NVIDIA RTX 3090 GPU respectively, while increasing inference speeds by up to 2.46x compared to CPU and maintaining 0.53x the speed of an RTX 3090 GPU despite the GPU's 4 times higher base clock rate. With the lack of existing open-source FPGA accelerators for transformers, we open-source our code and document our steps for synthesis. We hope this work will serve as a step in democratizing the use of FPGAs in transformer inference and inspire research into energy-efficient inference methods as a whole. The code can be found on https://github.com/HLSTransform/submission.

arXiv:2405.01491공식 코드 · Papers with Code

FeNNol: an Efficient and Flexible Library for Building Force-field-enhanced Neural Network Potentials

Neural network interatomic potentials (NNPs) have recently proven to be powerful tools to accurately model complex molecular systems while bypassing the high numerical cost of ab-initio molecular dynamics simulations. In recent years, numerous advances in model architectures as well as the development of hybrid models combining machine-learning (ML) with more traditional, physically-motivated, force-field interactions have considerably increased the design space of ML potentials. In this paper, we present FeNNol, a new library for building, training and running force-field-enhanced neural network potentials. It provides a flexible and modular system for building hybrid models, allowing to easily combine state-of-the-art embeddings with ML-parameterized physical interaction terms without the need for explicit programming. Furthermore, FeNNol leverages the automatic differentiation and just-in-time compilation features of the Jax Python library to enable fast evaluation of NNPs, shrinking the performance gap between ML potentials and standard force-fields. This is demonstrated with the popular ANI-2x model reaching simulation speeds nearly on par with the AMOEBA polarizable force-field on commodity GPUs (GPU=Graphics processing unit). We hope that FeNNol will facilitate the development and application of new hybrid NNP architectures for a wide range of molecular simulation problems.

arXiv:2405.00732공식 코드 · Papers with Code

LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4, A Technical Report

Low Rank Adaptation (LoRA) has emerged as one of the most widely adopted methods for Parameter Efficient Fine-Tuning (PEFT) of Large Language Models (LLMs). LoRA reduces the number of trainable parameters and memory usage while achieving comparable performance to full fine-tuning. We aim to assess the viability of training and serving LLMs fine-tuned with LoRA in real-world applications. First, we measure the quality of LLMs fine-tuned with quantized low rank adapters across 10 base models and 31 tasks for a total of 310 models. We find that 4-bit LoRA fine-tuned models outperform base models by 34 points and GPT-4 by 10 points on average. Second, we investigate the most effective base models for fine-tuning and assess the correlative and predictive capacities of task complexity heuristics in forecasting the outcomes of fine-tuning. Finally, we evaluate the latency and concurrency capabilities of LoRAX, an open-source Multi-LoRA inference server that facilitates the deployment of multiple LoRA fine-tuned models on a single GPU using shared base model weights and dynamic adapter loading. LoRAX powers LoRA Land, a web application that hosts 25 LoRA fine-tuned Mistral-7B LLMs on a single NVIDIA A100 GPU with 80GB memory. LoRA Land highlights the quality and cost-effectiveness of employing multiple specialized LLMs over a single, general-purpose LLM.

arXiv:2405.01413공식 코드 · Papers with Code

MiniGPT-3D: Efficiently Aligning 3D Point Clouds with Large Language Models using 2D Priors

Large 2D vision-language models (2D-LLMs) have gained significant attention by bridging Large Language Models (LLMs) with images using a simple projector. Inspired by their success, large 3D point cloud-language models (3D-LLMs) also integrate point clouds into LLMs. However, directly aligning point clouds with LLM requires expensive training costs, typically in hundreds of GPU-hours on A100, which hinders the development of 3D-LLMs. In this paper, we introduce MiniGPT-3D, an efficient and powerful 3D-LLM that achieves multiple SOTA results while training for only 27 hours on one RTX 3090. Specifically, we propose to align 3D point clouds with LLMs using 2D priors from 2D-LLMs, which can leverage the similarity between 2D and 3D visual information. We introduce a novel four-stage training strategy for modality alignment in a cascaded way, and a mixture of query experts module to adaptively aggregate features with high efficiency. Moreover, we utilize parameter-efficient fine-tuning methods LoRA and Norm fine-tuning, resulting in only 47.8M learnable parameters, which is up to 260x fewer than existing methods. Extensive experiments show that MiniGPT-3D achieves SOTA on 3D object classification and captioning tasks, with significantly cheaper training costs. Notably, MiniGPT-3D gains an 8.12 increase on GPT-4 evaluation score for the challenging object captioning task compared to ShapeLLM-13B, while the latter costs 160 total GPU-hours on 8 A800. We are the first to explore the efficient 3D-LLM, offering new insights to the community. Code and weights are available at https://github.com/TangYuan96/MiniGPT-3D.

arXiv:2405.00263공식 코드 · Papers with Code

Clover: Regressive Lightweight Speculative Decoding with Sequential Knowledge

Large language models (LLMs) suffer from low efficiency as the mismatch between the requirement of auto-regressive decoding and the design of most contemporary GPUs. Specifically, billions to trillions of parameters must be loaded to the GPU cache through its limited memory bandwidth for computation, but only a small batch of tokens is actually computed. Consequently, the GPU spends most of its time on memory transfer instead of computation. Recently, parallel decoding, a type of speculative decoding algorithms, is becoming more popular and has demonstrated impressive efficiency improvement in generation. It introduces extra decoding heads to large models, enabling them to predict multiple subsequent tokens simultaneously and verify these candidate continuations in a single decoding step. However, this approach deviates from the training objective of next token prediction used during pre-training, resulting in a low hit rate for candidate tokens. In this paper, we propose a new speculative decoding algorithm, Clover, which integrates sequential knowledge into the parallel decoding process. This enhancement improves the hit rate of speculators and thus boosts the overall efficiency. Clover transmits the sequential knowledge from pre-speculated tokens via the Regressive Connection, then employs an Attention Decoder to integrate these speculated tokens. Additionally, Clover incorporates an Augmenting Block that modifies the hidden states to better align with the purpose of speculative generation rather than next token prediction. The experiment results demonstrate that Clover outperforms the baseline by up to 91% on Baichuan-Small and 146% on Baichuan-Large, respectively, and exceeds the performance of the previously top-performing method, Medusa, by up to 37% on Baichuan-Small and 57% on Baichuan-Large, respectively.

arXiv:2404.19553공식 코드 · Papers with Code

Extending Llama-3's Context Ten-Fold Overnight

We extend the context length of Llama-3-8B-Instruct from 8K to 80K via QLoRA fine-tuning. The entire training cycle is super efficient, which takes 8 hours on one 8xA800 (80G) GPU machine. The resulted model exhibits superior performances across a broad range of evaluation tasks, such as NIHS, topic retrieval, and long-context language understanding; meanwhile, it also well preserves the original capability over short contexts. The dramatic context extension is mainly attributed to merely 3.5K synthetic training samples generated by GPT-4 , which indicates the LLMs' inherent (yet largely underestimated) potential to extend its original context length. In fact, the context length could be extended far beyond 80K with more computation resources. Therefore, the team will publicly release the entire resources (including data, model, data generation pipeline, training code) so as to facilitate the future research from the community: \url{https://github.com/FlagOpen/FlagEmbedding}.

arXiv:2404.19024공식 코드 · Papers with Code

Multi-Page Document Visual Question Answering using Self-Attention Scoring Mechanism

Documents are 2-dimensional carriers of written communication, and as such their interpretation requires a multi-modal approach where textual and visual information are efficiently combined. Document Visual Question Answering (Document VQA), due to this multi-modal nature, has garnered significant interest from both the document understanding and natural language processing communities. The state-of-the-art single-page Document VQA methods show impressive performance, yet in multi-page scenarios, these methods struggle. They have to concatenate all pages into one large page for processing, demanding substantial GPU resources, even for evaluation. In this work, we propose a novel method and efficient training strategy for multi-page Document VQA tasks. In particular, we employ a visual-only document representation, leveraging the encoder from a document understanding model, Pix2Struct. Our approach utilizes a self-attention scoring mechanism to generate relevance scores for each document page, enabling the retrieval of pertinent pages. This adaptation allows us to extend single-page Document VQA models to multi-page scenarios without constraints on the number of pages during evaluation, all with minimal demand for GPU resources. Our extensive experiments demonstrate not only achieving state-of-the-art performance without the need for Optical Character Recognition (OCR), but also sustained performance in scenarios extending to documents of nearly 800 pages compared to a maximum of 20 pages in the MP-DocVQA dataset. Our code is publicly available at \url{https://github.com/leitro/SelfAttnScoring-MPDocVQA}.

arXiv:2404.19525공식 코드 · Papers with Code

MicroDreamer: Efficient 3D Generation in $\sim$20 Seconds by Score-based Iterative Reconstruction

Optimization-based approaches, such as score distillation sampling (SDS), show promise in zero-shot 3D generation but suffer from low efficiency, primarily due to the high number of function evaluations (NFEs) required for each sample and the limitation of optimization confined to latent space. This paper introduces score-based iterative reconstruction (SIR), an efficient and general algorithm mimicking a differentiable 3D reconstruction process to reduce the NFEs and enable optimization in pixel space. Given a single set of images sampled from a multi-view score-based diffusion model, SIR repeatedly optimizes 3D parameters, unlike the single-step optimization in SDS. With other improvements in training, we present an efficient approach called MicroDreamer that generally applies to various 3D representations and 3D generation tasks. In particular, MicroDreamer is 5-20 times faster than SDS in generating neural radiance field while retaining a comparable performance and takes about 20 seconds to create meshes from 3D Gaussian splatting on a single A100 GPU, halving the time of the fastest optimization-based baseline DreamGaussian with significantly superior performance compared to the measurement standard deviation. Our code is available at https://github.com/ML-GSAI/MicroDreamer.

arXiv:2404.18617공식 코드 · Papers with Code

CoSense3D: an Agent-based Efficient Learning Framework for Collective Perception

Collective Perception has attracted significant attention in recent years due to its advantage for mitigating occlusion and expanding the field-of-view, thereby enhancing reliability, efficiency, and, most crucially, decision-making safety. However, developing collective perception models is highly resource demanding due to extensive requirements of processing input data for many agents, usually dozens of images and point clouds for a single frame. This not only slows down the model development process for collective perception but also impedes the utilization of larger models. In this paper, we propose an agent-based training framework that handles the deep learning modules and agent data separately to have a cleaner data flow structure. This framework not only provides an API for flexibly prototyping the data processing pipeline and defining the gradient calculation for each agent, but also provides the user interface for interactive training, testing and data visualization. Training experiment results of four collective object detection models on the prominent collective perception benchmark OPV2V show that the agent-based training can significantly reduce the GPU memory consumption and training time while retaining inference performance. The framework and model implementations are available at \url{https://github.com/YuanYunshuang/CoSense3D}

arXiv:2404.18174공식 코드 · Papers with Code

Mamba-FETrack: Frame-Event Tracking via State Space Model

RGB-Event based tracking is an emerging research topic, focusing on how to effectively integrate heterogeneous multi-modal data (synchronized exposure video frames and asynchronous pulse Event stream). Existing works typically employ Transformer based networks to handle these modalities and achieve decent accuracy through input-level or feature-level fusion on multiple datasets. However, these trackers require significant memory consumption and computational complexity due to the use of self-attention mechanism. This paper proposes a novel RGB-Event tracking framework, Mamba-FETrack, based on the State Space Model (SSM) to achieve high-performance tracking while effectively reducing computational costs and realizing more efficient tracking. Specifically, we adopt two modality-specific Mamba backbone networks to extract the features of RGB frames and Event streams. Then, we also propose to boost the interactive learning between the RGB and Event features using the Mamba network. The fused features will be fed into the tracking head for target object localization. Extensive experiments on FELT and FE108 datasets fully validated the efficiency and effectiveness of our proposed tracker. Specifically, our Mamba-based tracker achieves 43.5/55.6 on the SR/PR metric, while the ViT-S based tracker (OSTrack) obtains 40.0/50.9. The GPU memory cost of ours and ViT-S based tracker is 13.98GB and 15.44GB, which decreased about $9.5\%$. The FLOPs and parameters of ours/ViT-S based OSTrack are 59GB/1076GB and 7MB/60MB, which decreased about $94.5\%$ and $88.3\%$, respectively. We hope this work can bring some new insights to the tracking field and greatly promote the application of the Mamba architecture in tracking. The source code of this work will be released on \url{https://github.com/Event-AHU/Mamba_FETrack}.

arXiv:2404.17107공식 코드 · Papers with Code

Exploring Pre-trained General-purpose Audio Representations for Heart Murmur Detection

To reduce the need for skilled clinicians in heart sound interpretation, recent studies on automating cardiac auscultation have explored deep learning approaches. However, despite the demands for large data for deep learning, the size of the heart sound datasets is limited, and no pre-trained model is available. On the contrary, many pre-trained models for general audio tasks are available as general-purpose audio representations. This study explores the potential of general-purpose audio representations pre-trained on large-scale datasets for transfer learning in heart murmur detection. Experiments on the CirCor DigiScope heart sound dataset show that the recent self-supervised learning Masked Modeling Duo (M2D) outperforms previous methods with the results of a weighted accuracy of 0.832 and an unweighted average recall of 0.713. Experiments further confirm improved performance by ensembling M2D with other models. These results demonstrate the effectiveness of general-purpose audio representation in processing heart sounds and open the way for further applications. Our code is available online which runs on a 24 GB consumer GPU at https://github.com/nttcslab/m2d/tree/master/app/circor

arXiv:2404.14527공식 코드 · Papers with Code

Mélange: Cost Efficient Large Language Model Serving by Exploiting GPU Heterogeneity

Large language models (LLMs) are increasingly integrated into many online services, yet they remain cost-prohibitive to deploy due to the requirement of expensive GPU instances. Prior work has addressed the high cost of LLM serving by improving the inference engine, but less attention has been given to selecting the most cost-efficient GPU type(s) for a specific LLM service. There is a large and growing landscape of GPU types and, within these options, higher cost does not always lead to increased performance. Instead, through a comprehensive investigation, we find that three key LLM service characteristics (request size, request rate, SLO) strongly influence GPU cost efficiency, and differing GPU types are most cost efficient for differing LLM service settings. As a result, the most cost-efficient allocation for a given service is typically a mix of heterogeneous GPU types. Based on this analysis, we introduce M\'elange, a GPU allocation framework that navigates these diverse LLM service characteristics and heterogeneous GPU option space to automatically and efficiently derive the minimal-cost GPU allocation for a given LLM service. We formulate the GPU allocation task as a cost-aware bin packing problem where GPUs are bins and items are slices of the service workload. Our formulation's constraints account for a service's unique characteristics, allowing M\'elange to be flexible to support diverse service settings and heterogeneity-aware to adapt the GPU allocation to a specific service. Compared to using only a single GPU type, M\'elange reduces deployment costs by up to 77% in conversational settings, 33% in document-based settings, and 51% in a mixed setting.

arXiv:2404.14469공식 코드 · Papers with Code

SnapKV: LLM Knows What You are Looking for Before Generation

Large Language Models (LLMs) have made remarkable progress in processing extensive contexts, with the Key-Value (KV) cache playing a vital role in enhancing their performance. However, the growth of the KV cache in response to increasing input length poses challenges to memory and time efficiency. To address this problem, this paper introduces SnapKV, an innovative and fine-tuning-free approach that efficiently minimizes KV cache size while still delivering comparable performance in real-world applications. We discover that each attention head in the model consistently focuses on specific prompt attention features during generation. Meanwhile, this robust pattern can be obtained from an 'observation' window located at the end of the prompts. Drawing on this insight, SnapKV automatically compresses KV caches by selecting clustered important KV positions for each attention head. Our approach significantly reduces the growing computational overhead and memory footprint when processing long input sequences. Specifically, SnapKV achieves a consistent decoding speed with a 3.6x increase in generation speed and an 8.2x enhancement in memory efficiency compared to the baseline when processing inputs of 16K tokens. At the same time, it maintains comparable performance to the baseline models across 16 long sequence datasets. Moreover, SnapKV can process up to 380K context tokens on a single A100-80GB GPU using HuggingFace implementation with minor changes, exhibiting only a negligible accuracy drop in the Needle-in-a-Haystack test. Further comprehensive studies suggest SnapKV's potential for practical applications.

arXiv:2404.15159공식 코드 · Papers with Code

MixLoRA: Enhancing Large Language Models Fine-Tuning with LoRA-based Mixture of Experts

Fine-tuning Large Language Models (LLMs) is a common practice to adapt pre-trained models for specific applications. While methods like LoRA have effectively addressed GPU memory constraints during fine-tuning, their performance often falls short, especially in multi-task scenarios. In contrast, Mixture-of-Expert (MoE) models, such as Mixtral 8x7B, demonstrate remarkable performance in multi-task learning scenarios while maintaining a reduced parameter count. However, the resource requirements of these MoEs remain challenging, particularly for consumer-grade GPUs with less than 24GB memory. To tackle these challenges, we propose MixLoRA, an approach to construct a resource-efficient sparse MoE model based on LoRA. MixLoRA inserts multiple LoRA-based experts within the feed-forward network block of a frozen pre-trained dense model and employs a commonly used top-k router. Unlike other LoRA-based MoE methods, MixLoRA enhances model performance by utilizing independent attention-layer LoRA adapters. Additionally, an auxiliary load balance loss is employed to address the imbalance problem of the router. Our evaluations show that MixLoRA improves about 9% accuracy compared to state-of-the-art PEFT methods in multi-task learning scenarios. We also propose a new high-throughput framework to alleviate the computation and memory bottlenecks during the training and inference of MOE models. This framework reduces GPU memory consumption by 40% and token computation latency by 30% during both training and inference.

arXiv:2404.15141공식 코드 · Papers with Code

CutDiffusion: A Simple, Fast, Cheap, and Strong Diffusion Extrapolation Method

Transforming large pre-trained low-resolution diffusion models to cater to higher-resolution demands, i.e., diffusion extrapolation, significantly improves diffusion adaptability. We propose tuning-free CutDiffusion, aimed at simplifying and accelerating the diffusion extrapolation process, making it more affordable and improving performance. CutDiffusion abides by the existing patch-wise extrapolation but cuts a standard patch diffusion process into an initial phase focused on comprehensive structure denoising and a subsequent phase dedicated to specific detail refinement. Comprehensive experiments highlight the numerous almighty advantages of CutDiffusion: (1) simple method construction that enables a concise higher-resolution diffusion process without third-party engagement; (2) fast inference speed achieved through a single-step higher-resolution diffusion process, and fewer inference patches required; (3) cheap GPU cost resulting from patch-wise inference and fewer patches during the comprehensive structure denoising; (4) strong generation performance, stemming from the emphasis on specific detail refinement.

arXiv:2404.13781공식 코드 · Papers with Code

Evaluating Retrieval Quality in Retrieval-Augmented Generation

Evaluating retrieval-augmented generation (RAG) presents challenges, particularly for retrieval models within these systems. Traditional end-to-end evaluation methods are computationally expensive. Furthermore, evaluation of the retrieval model's performance based on query-document relevance labels shows a small correlation with the RAG system's downstream performance. We propose a novel evaluation approach, eRAG, where each document in the retrieval list is individually utilized by the large language model within the RAG system. The output generated for each document is then evaluated based on the downstream task ground truth labels. In this manner, the downstream performance for each document serves as its relevance label. We employ various downstream task metrics to obtain document-level annotations and aggregate them using set-based or ranking metrics. Extensive experiments on a wide range of datasets demonstrate that eRAG achieves a higher correlation with downstream RAG performance compared to baseline methods, with improvements in Kendall's $\tau$ correlation ranging from 0.168 to 0.494. Additionally, eRAG offers significant computational advantages, improving runtime and consuming up to 50 times less GPU memory than end-to-end evaluation.

arXiv:2404.14243공식 코드 · Papers with Code

Turbo-CF: Matrix Decomposition-Free Graph Filtering for Fast Recommendation

A series of graph filtering (GF)-based collaborative filtering (CF) showcases state-of-the-art performance on the recommendation accuracy by using a low-pass filter (LPF) without a training process. However, conventional GF-based CF approaches mostly perform matrix decomposition on the item-item similarity graph to realize the ideal LPF, which results in a non-trivial computational cost and thus makes them less practical in scenarios where rapid recommendations are essential. In this paper, we propose Turbo-CF, a GF-based CF method that is both training-free and matrix decomposition-free. Turbo-CF employs a polynomial graph filter to circumvent the issue of expensive matrix decompositions, enabling us to make full use of modern computer hardware components (i.e., GPU). Specifically, Turbo-CF first constructs an item-item similarity graph whose edge weights are effectively regulated. Then, our own polynomial LPFs are designed to retain only low-frequency signals without explicit matrix decompositions. We demonstrate that Turbo-CF is extremely fast yet accurate, achieving a runtime of less than 1 second on real-world benchmark datasets while achieving recommendation accuracies comparable to best competitors.

arXiv:2404.12968공식 코드 · Papers with Code

Scalable Data Assimilation with Message Passing

Data assimilation is a core component of numerical weather prediction systems. The large quantity of data processed during assimilation requires the computation to be distributed across increasingly many compute nodes, yet existing approaches suffer from synchronisation overhead in this setting. In this paper, we exploit the formulation of data assimilation as a Bayesian inference problem and apply a message-passing algorithm to solve the spatial inference problem. Since message passing is inherently based on local computations, this approach lends itself to parallel and distributed computation. In combination with a GPU-accelerated implementation, we can scale the algorithm to very large grid sizes while retaining good accuracy and compute and memory requirements.

arXiv:2404.11912공식 코드 · Papers with Code

TriForce: Lossless Acceleration of Long Sequence Generation with Hierarchical Speculative Decoding

With large language models (LLMs) widely deployed in long content generation recently, there has emerged an increasing demand for efficient long-sequence inference support. However, key-value (KV) cache, which is stored to avoid re-computation, has emerged as a critical bottleneck by growing linearly in size with the sequence length. Due to the auto-regressive nature of LLMs, the entire KV cache will be loaded for every generated token, resulting in low utilization of computational cores and high latency. While various compression methods for KV cache have been proposed to alleviate this issue, they suffer from degradation in generation quality. We introduce TriForce, a hierarchical speculative decoding system that is scalable for long sequence generation. This approach leverages the original model weights and dynamic sparse KV cache via retrieval as a draft model, which serves as an intermediate layer in the hierarchy and is further speculated by a smaller model to reduce its drafting latency. TriForce not only facilitates impressive speedups for Llama2-7B-128K, achieving up to 2.31$\times$ on an A100 GPU but also showcases scalability in handling even longer contexts. For the offloading setting on two RTX 4090 GPUs, TriForce achieves 0.108s/token$\unicode{x2014}$only half as slow as the auto-regressive baseline on an A100, which attains 7.78$\times$ on our optimized offloading system. Additionally, TriForce performs 4.86$\times$ than DeepSpeed-Zero-Inference on a single RTX 4090 GPU. TriForce's robustness is highlighted by its consistently outstanding performance across various temperatures. The code is available at https://github.com/Infini-AI-Lab/TriForce.