HierarchicalKV: A GPU Hash Table with Cache Semantics for Continuous Online Embedding Storage

Traditional GPU hash tables preserve every inserted key -- a dictionary assumption that wastes scarce High Bandwidth Memory (HBM) when embedding tables routinely exceed single-GPU capacity. We challenge this assumption with cache semantics, where policy-driven eviction is a first-class operation. We introduce HierarchicalKV (HKV), the first general-purpose GPU hash table library whose normal full-capacity operating contract is cache-semantic: each full-bucket upsert (update-or-insert) is resolved in place by eviction or admission rejection rather than by rehashing or capacity-induced failure. HKV co-designs four core mechanisms -- cache-line-aligned buckets, in-line score-driven upsert, score-based dynamic dual-bucket selection, and triple-group concurrency -- and uses tiered key-value separation as a scaling enabler beyond HBM. On an NVIDIA H100 NVL GPU, HKV achieves up to 3.9 billion key-value pairs per second (B-KV/s) find throughput, stable across load factors 0.50-1.00 (<5% variation), and delivers 1.4x higher find throughput than WarpCore (the strongest dictionary-semantic GPU baseline at lambda=0.50) and up to 2.6-9.4x over indirection-based GPU baselines. Since its open-source release in October 2022, HKV has been integrated into multiple open-source recommendation frameworks.

Joint Optimization of Storage and Loading for High-Performance 3D Point Cloud Data Processing

With the rapid development of computer vision and deep learning, significant advancements have been made in 3D vision, partic- ularly in autonomous driving, robotic perception, and augmented reality. 3D point cloud data, as a crucial representation of 3D information, has gained widespread attention. However, the vast scale and complexity of point cloud data present significant chal- lenges for loading and processing and traditional algorithms struggle to handle large-scale datasets.The diversity of storage formats for point cloud datasets (e.g., PLY, XYZ, BIN) adds complexity to data handling and results in inefficiencies in data preparation. Al- though binary formats like BIN and NPY have been used to speed up data access, they still do not fully address the time-consuming data loading and processing phase. To overcome these challenges, we propose the .PcRecord format, a unified data storage solution designed to reduce the storage occupation and accelerate the processing of point cloud data. We also introduce a high-performance data processing pipeline equipped with multiple modules. By leveraging a multi-stage parallel pipeline architecture, our system optimizes the use of computational resources, significantly improving processing speed and efficiency. This paper details the im- plementation of this system and demonstrates its effectiveness in addressing the challenges of handling large-scale point cloud datasets.On average, our system achieves performance improvements of 6.61x (ModelNet40), 2.69x (S3DIS), 2.23x (ShapeNet), 3.09x (Kitti), 8.07x (SUN RGB-D), and 5.67x (ScanNet) with GPU and 6.9x, 1.88x, 1.29x, 2.28x, 25.4x, and 19.3x with Ascend.

NCCL EP: Towards a Unified Expert Parallel Communication API for NCCL

Mixture-of-Experts (MoE) architectures have become essential for scaling large language models, driving the development of specialized device-initiated communication libraries such as DeepEP, Hybrid-EP, and others. These libraries demonstrate the performance benefits of GPU-initiated RDMA for MoE dispatch and combine operations. This paper presents NCCL EP (Expert Parallelism), a ground-up MoE communication library built entirely on NCCL's Device API. NCCL EP provides unified ncclEpDispatch and ncclEpCombine primitives with both C and Python interfaces, supporting Low-Latency (LL) mode for inference decoding and High-Throughput (HT) mode for training and inference prefill. LL targets small batch sizes (1-128 tokens) using direct all-to-all RDMA+NVLink mesh connectivity with double-buffered communication for overlapping dispatch and combine phases. HT targets large batches (4096+ tokens) using hierarchical communication that aggregates tokens within NVLink domains before inter-node RDMA transmission. Both modes leverage Device API for both intra- and inter-node communications, taking advantage of its topology awareness and optimized GPU-initiated implementation. We evaluate NCCL EP on an H100-based cluster across multi-node configurations, demonstrating competitive LL kernel performance and presenting end-to-end results with vLLM integration. By building MoE communication natively within NCCL, NCCL EP provides a supported path for expert parallelism on current and emerging NVIDIA platforms.

Thinking Traps in Long Chain-of-Thought: A Measurable Study and Trap-Aware Adaptive Restart

Scaling test-time compute via Long Chain-of-Thought (Long-CoT) significantly enhances reasoning capabilities, yet extended generation does not guarantee correctness: after an early wrong commitment, models may keep elaborating a self-consistent but incorrect prefix. Through fine-grained trajectory analysis, we identify Thinking Traps, prefix-dominant deadlocks where later reflection, alternative attempts, or verification fails to revise the root error. On a curated subset of DAPO-MATH, 89\% of failures exhibit such traps. To solve this problem, we introduce TAAR (Trap-Aware Adaptive Restart), a test-time control framework that trains a diagnostic policy to predict two signals from partial trajectories: a trap index for where to truncate and an escape probability for whether and how strongly to intervene. At inference time, TAAR truncates the trajectory before the predicted trap segment and adaptively restarts decoding; for severely trapped cases, it applies stronger perturbations, including higher-temperature resampling and an optional structured reboot suffix. Experiments on challenging mathematical and scientific reasoning benchmarks (AIME24, AIME25, GPQA-Diamond, HMMT25, BRUMO25) show that TAAR improves reasoning performance without fine-tuning base model parameters.

JoyVoice: Long-Context Conditioning for Anthropomorphic Multi-Speaker Conversational Synthesis

Large speech generation models are evolving from single-speaker, short sentence synthesis to multi-speaker, long conversation geneartion. Current long-form speech generation models are predominately constrained to dyadic, turn-based interactions. To address this, we introduce JoyVoice, a novel anthropomorphic foundation model designed for flexible, boundary-free synthesis of up to eight speakers. Unlike conventional cascaded systems, JoyVoice employs a unified E2E-Transformer-DiT architecture that utilizes autoregressive hidden representations directly for diffusion inputs, enabling holistic end-to-end optimization. We further propose a MM-Tokenizer operating at a low bitrate of 12.5 Hz, which integrates multitask semantic and MMSE losses to effectively model both semantic and acoustic information. Additionally, the model incorporates robust text front-end processing via large-scale data perturbation. Experiments show that JoyVoice achieves state-of-the-art results in multilingual generation (Chinese, English, Japanese, Korean) and zero-shot voice cloning. JoyVoice achieves top-tier results on both the Seed-TTS-Eval Benchmark and multi-speaker long-form conversational voice cloning tasks, demonstrating superior audio quality and generalization. It achieves significant improvements in prosodic continuity for long-form speech, rhythm richness in multi-speaker conversations, paralinguistic naturalness, besides superior intelligibility. We encourage readers to listen to the demo at https://jea-speech.github.io/JoyVoice

SlotPi: Physics-informed Object-centric Reasoning Models

Understanding and reasoning about dynamics governed by physical laws through visual observation, akin to human capabilities in the real world, poses significant challenges. Currently, object-centric dynamic simulation methods, which emulate human behavior, have achieved notable progress but overlook two critical aspects: 1) the integration of physical knowledge into models. Humans gain physical insights by observing the world and apply this knowledge to accurately reason about various dynamic scenarios; 2) the validation of model adaptability across diverse scenarios. Real-world dynamics, especially those involving fluids and objects, demand models that not only capture object interactions but also simulate fluid flow characteristics. To address these gaps, we introduce SlotPi, a slot-based physics-informed object-centric reasoning model. SlotPi integrates a physical module based on Hamiltonian principles with a spatio-temporal prediction module for dynamic forecasting. Our experiments highlight the model's strengths in tasks such as prediction and Visual Question Answering (VQA) on benchmark and fluid datasets. Furthermore, we have created a real-world dataset encompassing object interactions, fluid dynamics, and fluid-object interactions, on which we validated our model's capabilities. The model's robust performance across all datasets underscores its strong adaptability, laying a foundation for developing more advanced world models.

CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training

In our prior works, we introduced a scalable streaming speech synthesis model, CosyVoice 2, which integrates a large language model (LLM) and a chunk-aware flow matching (FM) model, and achieves low-latency bi-streaming speech synthesis and human-parity quality. Despite these advancements, CosyVoice 2 exhibits limitations in language coverage, domain diversity, data volume, text formats, and post-training techniques. In this paper, we present CosyVoice 3, an improved model designed for zero-shot multilingual speech synthesis in the wild, surpassing its predecessor in content consistency, speaker similarity, and prosody naturalness. Key features of CosyVoice 3 include: 1) A novel speech tokenizer to improve prosody naturalness, developed via supervised multi-task training, including automatic speech recognition, speech emotion recognition, language identification, audio event detection, and speaker analysis. 2) A new differentiable reward model for post-training applicable not only to CosyVoice 3 but also to other LLM-based speech synthesis models. 3) Dataset Size Scaling: Training data is expanded from ten thousand hours to one million hours, encompassing 9 languages and 18 Chinese dialects across various domains and text formats. 4) Model Size Scaling: Model parameters are increased from 0.5 billion to 1.5 billion, resulting in enhanced performance on our multilingual benchmark due to the larger model capacity. These advancements contribute significantly to the progress of speech synthesis in the wild. We encourage readers to listen to the demo at https://funaudiollm.github.io/cosyvoice3.

EmoVoice: LLM-based Emotional Text-To-Speech Model with Freestyle Text Prompting

Human speech goes beyond the mere transfer of information; it is a profound exchange of emotions and a connection between individuals. While Text-to-Speech (TTS) models have made huge progress, they still face challenges in controlling the emotional expression in the generated speech. In this work, we propose EmoVoice, a novel emotion-controllable TTS model that exploits large language models (LLMs) to enable fine-grained freestyle natural language emotion control, and a phoneme boost variant design that makes the model output phoneme tokens and audio tokens in parallel to enhance content consistency, inspired by chain-of-thought (CoT) and chain-of-modality (CoM) techniques. Besides, we introduce EmoVoice-DB, a high-quality 40-hour English emotion dataset featuring expressive speech and fine-grained emotion labels with natural language descriptions. EmoVoice achieves state-of-the-art performance on the English EmoVoice-DB test set using only synthetic training data, and on the Chinese Secap test set using our in-house data. We further investigate the reliability of existing emotion evaluation metrics and their alignment with human perceptual preferences, and explore using SOTA multimodal LLMs GPT-4o-audio and Gemini to assess emotional speech. Demo samples are available at https://yanghaha0908.github.io/EmoVoice/. Dataset, code, and checkpoints will be released.

Accurate Expert Predictions in MoE Inference via Cross-Layer Gate

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.

Klotski: Efficient Mixture-of-Expert Inference via Expert-Aware Multi-Batch Pipeline

Mixture of Experts (MoE), with its distinctive sparse structure, enables the scaling of language models up to trillions of parameters without significantly increasing computational costs. However, the substantial parameter size presents a challenge for inference, as the expansion in GPU memory cannot keep pace with the growth in parameters. Although offloading techniques utilise memory from the CPU and disk and parallelise the I/O and computation for efficiency, the computation for each expert in MoE models is often less than the I/O, resulting in numerous bubbles in the pipeline. Therefore, we propose Klotski, an efficient MoE inference engine that significantly reduces pipeline bubbles through a novel expert-aware multi-batch pipeline paradigm. The proposed paradigm uses batch processing to extend the computation time of the current layer to overlap with the loading time of the next layer. Although this idea has been effectively applied to dense models, more batches may activate more experts in the MoE, leading to longer loading times and more bubbles. Thus, unlike traditional approaches, we balance computation and I/O time and minimise bubbles by orchestrating their inference orders based on their heterogeneous computation and I/O requirements and activation patterns under different batch numbers. Moreover, to adapt to different hardware environments and models, we design a constraint-sensitive I/O-compute planner and a correlation-aware expert prefetcher for a schedule that minimises pipeline bubbles. Experimental results demonstrate that Klotski achieves a superior throughput-latency trade-off compared to state-of-the-art techniques, with throughput improvements of up to 85.12x.