This machine learning study investigates a lowcost edge device integrated with an embedded system having computer vision and resulting in an improved performance in inferencing time and precision of object detection and classification. A primary aim of this study focused on reducing inferencing time and low-power consumption and to enable an embedded device of a competition-ready autonomous humanoid robot and to support real-time object recognition, scene understanding, visual navigation, motion planning, and autonomous navigation of the robot. This study compares processors for inferencing time performance between a central processing unit (CPU), a graphical processing unit (GPU), and a tensor processing unit (TPU). CPUs, GPUs, and TPUs are all processors that can be used for machine learning tasks. Related to the aim of supporting an autonomous humanoid robot, there was an additional effort to observe whether or not there was a significant difference in using a camera having monocular vision versus stereo vision capability. TPU inference time results for this study reflect a 25% reduction in time over the GPU, and a whopping 87.5% reduction in inference time compared to the CPU. Much information in this paper is contributed to the final selection of Google's Coral brand, Edge TPU device. The Arduino Nano 33 BLE Sense Tiny ML Kit was also considered for comparison but due to initial incompatibilities and in the interest of time to complete this study, a decision was made to review the kit in a future experiment.
We present an accurate and GPU-accelerated Stereo Visual SLAM design called Jetson-SLAM. It exhibits frame-processing rates above 60FPS on NVIDIA's low-powered 10W Jetson-NX embedded computer and above 200FPS on desktop-grade 200W GPUs, even in stereo configuration and in the multiscale setting. Our contributions are threefold: (i) a Bounded Rectification technique to prevent tagging many non-corner points as a corner in FAST detection, improving SLAM accuracy. (ii) A novel Pyramidal Culling and Aggregation (PyCA) technique that yields robust features while suppressing redundant ones at high speeds by harnessing a GPU device. PyCA uses our new Multi-Location Per Thread culling strategy (MLPT) and Thread-Efficient Warp-Allocation (TEWA) scheme for GPU to enable Jetson-SLAM achieving high accuracy and speed on embedded devices. (iii) Jetson-SLAM library achieves resource efficiency by having a data-sharing mechanism. Our experiments on three challenging datasets: KITTI, EuRoC, and KAIST-VIO, and two highly accurate SLAM backends: Full-BA and ICE-BA show that Jetson-SLAM is the fastest available accurate and GPU-accelerated SLAM system (Fig. 1).
LLM inference for popular enterprise use cases, such as summarization, RAG, and code-generation, typically observes orders of magnitude longer prompt lengths than generation lengths. This characteristic leads to high cost of prefill and increased response latency. In this paper, we present SwiftKV, a novel model transformation and distillation procedure specifically designed to reduce the time and cost of processing prompt tokens while preserving high quality of generated tokens. SwiftKV 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 procedure that can adapt existing LLMs for SwiftKV with minimal accuracy impact and low compute and data requirement. For Llama-3.1-8B and 70B, SwiftKV 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. In the end-to-end inference serving using an optimized vLLM implementation, SwiftKV realizes up to 2x higher aggregate throughput and 60% lower time per output token. It can achieve a staggering 560 TFlops/GPU of normalized inference throughput, which translates to 16K tokens/s for Llama-3.1-70B in 16-bit precision on 4x H100 GPUs. Our training, inference, and model implementations are open-sourced and can be found through https://huggingface.co/collections/Snowflake/swiftkv-models-674f7d7474eb789e185d31cb.
Models for streaming speech translation (ST) can achieve high accuracy and low latency if they're developed with vast amounts of paired audio in the source language and written text in the target language. Yet, these text labels for the target language are often pseudo labels due to the prohibitive cost of manual ST data labeling. In this paper, we introduce a methodology named Connectionist Temporal Classification guided modality matching (CTC-GMM) that enhances the streaming ST model by leveraging extensive machine translation (MT) text data. This technique employs CTC to compress the speech sequence into a compact embedding sequence that matches the corresponding text sequence, allowing us to utilize matched {source-target} language text pairs from the MT corpora to refine the streaming ST model further. Our evaluations with FLEURS and CoVoST2 show that the CTC-GMM approach can increase translation accuracy relatively by 13.9% and 6.4% respectively, while also boosting decoding speed by 59.7% on GPU.
In the field of image editing, three core challenges persist: controllability, background preservation, and efficiency. Inversion-based methods rely on time-consuming optimization to preserve the features of the initial images, which results in low efficiency due to the requirement for extensive network inference. Conversely, inversion-free methods lack theoretical support for background similarity, as they circumvent the issue of maintaining initial features to achieve efficiency. As a consequence, none of these methods can achieve both high efficiency and background consistency. To tackle the challenges and the aforementioned disadvantages, we introduce PostEdit, a method that incorporates a posterior scheme to govern the diffusion sampling process. Specifically, a corresponding measurement term related to both the initial features and Langevin dynamics is introduced to optimize the estimated image generated by the given target prompt. Extensive experimental results indicate that the proposed PostEdit achieves state-of-the-art editing performance while accurately preserving unedited regions. Furthermore, the method is both inversion- and training-free, necessitating approximately 1.5 seconds and 18 GB of GPU memory to generate high-quality results.
Neurosymbolic learning enables the integration of symbolic reasoning with deep learning but faces significant challenges in scaling to complex symbolic programs, large datasets, or both. We introduce Dolphin, a framework that tackles these challenges by supporting neurosymbolic programs in Python, executing complex symbolic reasoning on the CPU while vectorizing probabilistic computations and gradient propagation on the GPU. Across 13 benchmarks spanning tasks over text, image, and video data, with symbolic reasoning features like recursion and black-box functions, Dolphin converges to state-of-the-art accuracies on the more complex benchmarks while existing frameworks such as Scallop, ISED, and IndeCateR+ fail to converge within the time limit. On simpler benchmarks, Dolphin matches their performance, while achieving these results 1.71x to 62x faster than the baselines. Overall, Dolphin advances the scalability of neurosymbolic frameworks, achieving state-of-the-art efficiency and convergence on difficult benchmarks where existing frameworks struggle.
Recent advancements in Large Language Models (LLMs) have significantly increased context window sizes, enabling sophisticated applications but also introducing substantial computational overheads, particularly computing key-value (KV) cache in the prefill stage. Prefix caching has emerged to save GPU power in this scenario, which saves KV cache at disks and reuse them across multiple queries. However, traditional prefix caching mechanisms often suffer from substantial latency because the speed of loading KV cache from disks to GPU memory is bottlenecked by the throughput of I/O devices. To optimize the latency of long-context prefill, we propose Cake, a novel KV cache loader, which employs a bidirectional parallelized KV cache generation strategy. Upon receiving a prefill task, Cake simultaneously and dynamically loads saved KV cache from prefix cache locations and computes KV cache on local GPUs, maximizing the utilization of available computation and I/O bandwidth resources. Additionally, Cake automatically adapts to diverse system statuses without manual parameter. tuning. In experiments on various prompt datasets, GPUs, and I/O devices, Cake offers up to 68.1% Time To First Token (TTFT) reduction compare with compute-only method and 94.6% TTFT reduction compare with I/O-only method.
The Key-Value (KV) cache is a crucial component in serving transformer-based autoregressive large language models (LLMs), enabling faster inference by storing previously computed KV vectors. However, its memory consumption scales linearly with sequence length and batch size, posing a significant bottleneck in LLM deployment. Existing approaches to mitigate this issue include: (1) efficient attention variants integrated in upcycling stages, which requires extensive parameter tuning thus unsuitable for pre-trained LLMs; (2) KV cache compression at test time, primarily through token eviction policies, which often overlook inter-layer dependencies and can be task-specific. This paper introduces an orthogonal approach to KV cache compression. We propose a low-rank approximation of KV weight matrices, allowing for plug-in integration with existing transformer-based LLMs without model retraining. To effectively compress KV cache at the weight level, we adjust for layerwise sensitivity and introduce a progressive compression strategy, which is supported by our theoretical analysis on how compression errors accumulate in deep networks. Our method is designed to function without model tuning in upcycling stages or task-specific profiling in test stages. Extensive experiments with LLaMA models ranging from 8B to 70B parameters across various tasks show that our approach significantly reduces the GPU memory footprint while maintaining performance.
Throughout its lifecycle, a large language model (LLM) generates a substantially larger carbon footprint during inference than training. LLM inference requests vary in batch size, prompt length, and token generation number, while cloud providers employ different GPU types and quantities to meet diverse service-level objectives for accuracy and latency. It is crucial for both users and cloud providers to have a tool that quickly and accurately estimates the carbon impact of LLM inferences based on a combination of inference request and hardware configurations before execution. Estimating the carbon footprint of LLM inferences is more complex than training due to lower and highly variable model FLOPS utilization, rendering previous equation-based models inaccurate. Additionally, existing machine learning (ML) prediction methods either lack accuracy or demand extensive training data, as they inadequately handle the distinct prefill and decode phases, overlook hardware-specific features, and inefficiently sample uncommon inference configurations. We introduce \coo, a graph neural network (GNN)-based model that greatly improves the accuracy of LLM inference carbon footprint predictions compared to previous methods.
Universal dexterous grasping across diverse objects presents a fundamental yet formidable challenge in robot learning. Existing approaches using reinforcement learning (RL) to develop policies on extensive object datasets face critical limitations, including complex curriculum design for multi-task learning and limited generalization to unseen objects. To overcome these challenges, we introduce ResDex, a novel approach that integrates residual policy learning with a mixture-of-experts (MoE) framework. ResDex is distinguished by its use of geometry-unaware base policies that are efficiently acquired on individual objects and capable of generalizing across a wide range of unseen objects. Our MoE framework incorporates several base policies to facilitate diverse grasping styles suitable for various objects. By learning residual actions alongside weights that combine these base policies, ResDex enables efficient multi-task RL for universal dexterous grasping. ResDex achieves state-of-the-art performance on the DexGraspNet dataset comprising 3,200 objects with an 88.8% success rate. It exhibits no generalization gap with unseen objects and demonstrates superior training efficiency, mastering all tasks within only 12 hours on a single GPU.
As Large Language Models (LLMs) are rapidly growing in popularity, LLM inference services must be able to serve requests from thousands of users while satisfying performance requirements. The performance of an LLM inference service is largely determined by the hardware onto which it is deployed, but understanding of which hardware will deliver on performance requirements remains challenging. In this work we present LLM-Pilot - a first-of-its-kind system for characterizing and predicting performance of LLM inference services. LLM-Pilot performs benchmarking of LLM inference services, under a realistic workload, across a variety of GPUs, and optimizes the service configuration for each considered GPU to maximize performance. Finally, using this characterization data, LLM-Pilot learns a predictive model, which can be used to recommend the most cost-effective hardware for a previously unseen LLM. Compared to existing methods, LLM-Pilot can deliver on performance requirements 33% more frequently, whilst reducing costs by 60% on average.
Recently, integrating the local modeling capabilities of Convolutional Neural Networks (CNNs) with the global dependency strengths of Transformers has created a sensation in the semantic segmentation community. However, substantial computational workloads and high hardware memory demands remain major obstacles to their further application in real-time scenarios. In this work, we propose a Lightweight Multiple-Information Interaction Network (LMIINet) for real-time semantic segmentation, which effectively combines CNNs and Transformers while reducing redundant computations and memory footprints. It features Lightweight Feature Interaction Bottleneck (LFIB) modules comprising efficient convolutions that enhance context integration. Additionally, improvements are made to the Flatten Transformer by enhancing local and global feature interaction to capture detailed semantic information. Incorporating a combination coefficient learning scheme in both LFIB and Transformer blocks facilitates improved feature interaction. Extensive experiments demonstrate that LMIINet excels in balancing accuracy and efficiency. With only 0.72M parameters and 11.74G FLOPs (Floating Point Operations Per Second), LMIINet achieves 72.0\% mIoU at 100 FPS (Frames Per Second) on the Cityscapes test set and 69.94\% mIoU (mean Intersection over Union) at 160 FPS on the CamVid test dataset using a single RTX2080Ti GPU.
We present a foundation model for zero-shot metric monocular depth estimation. Our model, Depth Pro, synthesizes high-resolution depth maps with unparalleled sharpness and high-frequency details. The predictions are metric, with absolute scale, without relying on the availability of metadata such as camera intrinsics. And the model is fast, producing a 2.25-megapixel depth map in 0.3 seconds on a standard GPU. These characteristics are enabled by a number of technical contributions, including an efficient multi-scale vision transformer for dense prediction, a training protocol that combines real and synthetic datasets to achieve high metric accuracy alongside fine boundary tracing, dedicated evaluation metrics for boundary accuracy in estimated depth maps, and state-of-the-art focal length estimation from a single image. Extensive experiments analyze specific design choices and demonstrate that Depth Pro outperforms prior work along multiple dimensions. We release code and weights at https://github.com/apple/ml-depth-pro
We introduce Learning from Offline Foundation Features with Tensor Augmentations (LOFF-TA), an efficient training scheme designed to harness the capabilities of foundation models in limited resource settings where their direct development is not feasible. LOFF-TA involves training a compact classifier on cached feature embeddings from a frozen foundation model, resulting in up to $37\times$ faster training and up to $26\times$ reduced GPU memory usage. Because the embeddings of augmented images would be too numerous to store, yet the augmentation process is essential for training, we propose to apply tensor augmentations to the cached embeddings of the original non-augmented images. LOFF-TA makes it possible to leverage the power of foundation models, regardless of their size, in settings with limited computational capacity. Moreover, LOFF-TA can be used to apply foundation models to high-resolution images without increasing compute. In certain scenarios, we find that training with LOFF-TA yields better results than directly fine-tuning the foundation model.
Dense document embeddings are central to neural retrieval. The dominant paradigm is to train and construct embeddings by running encoders directly on individual documents. In this work, we argue that these embeddings, while effective, are implicitly out-of-context for targeted use cases of retrieval, and that a contextualized document embedding should take into account both the document and neighboring documents in context - analogous to contextualized word embeddings. We propose two complementary methods for contextualized document embeddings: first, an alternative contrastive learning objective that explicitly incorporates the document neighbors into the intra-batch contextual loss; second, a new contextual architecture that explicitly encodes neighbor document information into the encoded representation. Results show that both methods achieve better performance than biencoders in several settings, with differences especially pronounced out-of-domain. We achieve state-of-the-art results on the MTEB benchmark with no hard negative mining, score distillation, dataset-specific instructions, intra-GPU example-sharing, or extremely large batch sizes. Our method can be applied to improve performance on any contrastive learning dataset and any biencoder.
Single-molecule localization microscopy (SMLM) surpasses the diffraction limit, achieving subcellular resolution. Traditional SMLM analysis methods often rely on point spread function (PSF) model fitting, limiting the application of complex PSF models. In recent years, deep learning approaches have significantly improved SMLM algorithms, yielding promising results. However, limitations in inference speed and model size have restricted the widespread adoption of deep learning in practical applications. To address these challenges, this paper proposes an efficient model deployment framework and introduces a lightweight neural network, DilatedLoc, aimed at enhancing both image reconstruction quality and inference speed. Compared to leading network models, DilatedLoc reduces network parameters to under 100 MB and achieves a 50% improvement in inference speed, with superior GPU utilization through a novel deployment architecture compatible with various network models.
This work addresses the persistent challenges of substantial training time and GPU resource requirements even when training lightweight encoder-vocoder models for Textless NLP. We reduce training steps significantly while improving performance by a) leveraging learning rate schedulers for efficient and faster convergence b) optimizing hop length and c) tuning the interpolation scale factors for better audio quality. Additionally, we explore the latent space representation for Indian languages such as Tamil and Bengali for the acoustic unit discovery and voice conversion task. Our approach leverages a quantized encoder architecture, in conjunction with a vocoder which utilizes the proposed mixture of optimized hop length, tuned interpolation scale factors and a cyclic learning rate scheduler. We obtain consistently good results across English, Tamil and Bengali datasets. The proposed method excels in capturing complex linguistic patterns, resulting in clear reconstructed audio during voice conversion with significantly reduced training time.
Scaling the input context length of a large language model (LLM) incurs a significant increase in computation cost and memory footprint to maintain the attention key-value (KV) cache. Existing KV cache compression methods suffer from inefficient compression strategies and limited memory reduction effects, making it difficult for LLMs to conduct long-context inference on consumer-grade devices, especially when inferring long-context stream input. Such obstacles prevent consumer-grade devices from supporting more complex applications, creating challenges for the democratization of LLMs. To overcome this, we propose Locret, the first framework to create an eviction policy compatible with chunked prefill. By evaluating the causal importance of KV cache units by learnable retaining heads, Locret enables precise eviction of cache units, facilitating efficient long-context inference. In our extensive empirical studies, Locret outperforms the recent popular and competitive approaches in terms of memory efficiency and generation quality -- Locret achieves up to 20x of KV cache compression ratio within less than 10% performance loss. Furthermore, Locret achieves 128K+ long-context inference on a single NVIDIA 4090 GPU without compromising generation quality and only costs <1 GPU hour of additional training.
Training and serving long-context large language models (LLMs) incurs substantial overhead. To address this, two critical steps are often required: a pretrained LLM typically undergoes a separate stage for context length extension by training on long-context data, followed by architectural modifications to reduce the overhead of KV cache during serving. This paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training overhead during length extension, but also achieves better long-context performance. This leads to our proposed LongGen, which finetunes a pretrained LLM into an efficient architecture during length extension. LongGen builds on three key insights: (1) Sparse attention patterns, such as window attention (attending to recent tokens), attention sink (initial ones), and blockwise sparse attention (strided token blocks) are well-suited for building efficient long-context models, primarily due to their GPU-friendly memory access patterns, enabling efficiency gains not just theoretically but in practice as well. (2) It is essential for the model to have direct access to all tokens. A hybrid architecture with 1/3 full attention layers and 2/3 efficient ones achieves a balanced trade-off between efficiency and long-context performance. (3) Lightweight training on 5B long-context data is sufficient to extend the hybrid model's context length from 4K to 128K. We evaluate LongGen on both Llama-2 7B and Llama-2 70B, demonstrating its effectiveness across different scales. During training with 128K-long contexts, LongGen achieves 1.55x training speedup and reduces wall-clock time by 36%, compared to a full-attention baseline. During inference, LongGen reduces KV cache memory by 62%, achieving 1.67x prefilling speedup and 1.41x decoding speedup.
The rapid scaling of large language models necessitates more lightweight finetuning methods to reduce the explosive GPU memory overhead when numerous customized models are served simultaneously. Targeting more parameter-efficient low-rank adaptation (LoRA), parameter sharing presents a promising solution. Empirically, our research into high-level sharing principles highlights the indispensable role of differentiation in reversing the detrimental effects of pure sharing. Guided by this finding, we propose Mixture of Shards (MoS), incorporating both inter-layer and intra-layer sharing schemes, and integrating four nearly cost-free differentiation strategies, namely subset selection, pair dissociation, vector sharding, and shard privatization. Briefly, it selects a designated number of shards from global pools with a Mixture-of-Experts (MoE)-like routing mechanism before sequentially concatenating them to low-rank matrices. Hence, it retains all the advantages of LoRA while offering enhanced parameter efficiency, and effectively circumvents the drawbacks of peer parameter-sharing methods. Our empirical experiments demonstrate approximately 8x parameter savings in a standard LoRA setting. The ablation study confirms the significance of each component. Our insights into parameter sharing and MoS method may illuminate future developments of more parameter-efficient finetuning methods.