The emergence of Large Language Models (LLMs) with strong reasoning capabilities marks a significant milestone, unlocking new frontiers in complex problem-solving. However, training these reasoning models, typically using Reinforcement Learning (RL), encounters critical efficiency bottlenecks: response generation during RL training exhibits a persistent long-tail distribution, where a few very long responses dominate execution time, wasting resources and inflating costs. To address this, we propose TLT, a system that accelerates reasoning RL training losslessly by integrating adaptive speculative decoding. Applying speculative decoding in RL is challenging due to the dynamic workloads, evolving target model, and draft model training overhead. TLT overcomes these obstacles with two synergistic components: (1) Adaptive Drafter, a lightweight draft model trained continuously on idle GPUs during long-tail generation to maintain alignment with the target model at no extra cost; and (2) Adaptive Rollout Engine, which maintains a memory-efficient pool of pre-captured CUDAGraphs and adaptively select suitable SD strategies for each input batch. Evaluations demonstrate that TLT achieves over 1.7x end-to-end RL training speedup over state-of-the-art systems, preserves the model accuracy, and yields a high-quality draft model as a free byproduct suitable for efficient deployment. Code is released at https://github.com/mit-han-lab/fastrl.

Training large language models (LLMs) with increasingly long and varying sequence lengths introduces severe load imbalance challenges in large-scale data-parallel training. Recent frameworks attempt to mitigate these issues through data reorganization or hybrid parallel strategies. However, they often overlook how computational and communication costs scale with sequence length, resulting in suboptimal performance. We identify three critical challenges: (1) varying computation-to-communication ratios across sequences of different lengths in distributed attention, (2) mismatch between static NIC-GPU affinity and dynamic parallel workloads, and (3) distinct optimal partitioning strategies required for quadratic attention versus linear components.

Fine-tuning large language models (LLMs) greatly improves model quality for downstream tasks. However, serving many fine-tuned LLMs concurrently is challenging due to the sporadic, bursty, and varying request patterns of different LLMs. To bridge this gap, we present DeltaZip, an LLM serving system that efficiently serves multiple full-parameter fine-tuned models concurrently by aggressively compressing model deltas by up to 10× while maintaining high model quality. The key insight behind this design is that fine-tuning results in small-magnitude changes to the pre-trained model. By co-designing the serving system with the compression algorithm, DeltaZip achieves 2× to 12× improvement in throughput compared to the state-of-the-art systems.

Long-context capability is critical for multi-modal foundation models, especially for long video understanding. We introduce LongVILA, a full-stack solution for long-context visual-language models by co-designing the algorithm and system. For model training, we upgrade existing VLMs to support long video understanding by incorporating two additional stages, i.e., long context extension and long video supervised fine-tuning. However, training on long video is computationally and memory intensive. We introduce the long-context Multi-Modal Sequence Parallelism (MM-SP) system that efficiently parallelizes long video training and inference, enabling 2M context length training on 256 GPUs without any gradient checkpointing. LongVILA efficiently extends the number of video frames of VILA from 8 to 2048, achieving 99.8% accuracy in 6,000-frame (more than 1 million tokens) video needle-in-a-haystack. LongVILA-7B demonstrates strong accuracy on 9 popular video benchmarks, e.g., 65.1% VideoMME with subtitle. Besides, MM-SP is 2.1x - 5.7x faster than ring style sequence parallelism and 1.1x - 1.4x faster than Megatron with a hybrid context and tensor parallelism. Moreover, it seamlessly integrates with Hugging Face Transformers.

We introduce a full-stack framework that scales up reasoning in vision-language models (VLMs) to long videos, leveraging reinforcement learning. We address the unique challenges of long video reasoning by integrating three critical components: (1) a large-scale dataset, LongVideo-Reason, comprising 104K long video QA pairs with high-quality reasoning annotations across diverse domains such as sports, games, and vlogs; (2) a two-stage training pipeline that extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL); and (3) a training infrastructure for long video RL, named Multi-modal Reinforcement Sequence Parallelism (MR-SP), which incorporates sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling. In our experiments, LongVILA-R1-7B achieves strong performance on video benchmarks, reaching 65.1% and 71.1% accuracy on VideoMME without and with subtitles, respectively, and consistently outperforming LongVILA-7B across multiple benchmarks. Moreover, LongVILA-R1-7B supports processing up to 8,192 video frames per video, and configurable FPS settings. Notably, our MR-SP system achieves up to 2.1x speedup on long video RL training. In addition, we release our training system for public availability that supports RL training on various modalities (video, text, and audio), various models (VILA and Qwen series), and even image and video generation models. On a single A100 node (8 GPUs), it supports RL training on hour-long videos (e.g., 3,600 frames). Code and models are available at https://github.com/NVlabs/Long-RL

We present Jet-Nemotron, a new family of hybrid-architecture language models, which matches or exceeds the accuracy of leading full-attention models while significantly improving generation throughput. Jet-Nemotron is developed using Post Neural Architecture Search (PostNAS), a novel neural architecture exploration pipeline that enables efficient model design. Unlike prior approaches, PostNAS begins with a pre-trained full-attention model and freezes its MLP weights, allowing efficient exploration of attention block designs. The pipeline includes four key components: (1) learning optimal full-attention layer placement and elimination, (2) linear attention block selection, (3) designing new attention blocks, and (4) performing hardware-aware hyperparameter search. Our Jet-Nemotron-2B model achieves comparable or superior accuracy to Qwen3, Qwen2.5, Gemma3, and Llama3.2 across a comprehensive suite of benchmarks while delivering up to 53.6× generation throughput speedup and 6.1× prefilling speedup. It also achieves higher accuracy on MMLU and MMLU-Pro than recent advanced MoE full-attention models, such as DeepSeek-V3-Small and Moonlight, despite their larger scale with 15B total and 2.2B activated parameters.

The high GPU demand of ML training makes it hard to allocate large homogeneous clusters of high-end GPUs in a single availability zone. Leveraging heterogeneous GPUs available within and across zones can improve throughput at a reasonable cost. However, training ML models on heterogeneous resources introduces significant challenges, such as stragglers and a large search space of possible job configurations. Current systems lack support for efficiently training models on heterogeneous resources. We present Sailor, a system that automates distributed training over heterogeneous, geo-distributed, and dynamically available resources. Sailor combines an efficient search space exploration algorithm, accurate runtime and memory footprint simulation, and a distributed training framework that supports different types of heterogeneity to optimize training throughput and cost.

Distributed full-graph training of Graph Neural Networks (GNNs) has been widely adopted to learn large-scale graphs. While recent system advancements can improve the training throughput of GNNs, their practical adoption is limited by the potential accuracy decline. This concern is particularly prominent in deeper and more intricate GNN architectures, where noticeable performance degradation becomes apparent. Moreover, existing works fail to comprehensively consider diverse opportunities for acceleration. Motivated by these deficiencies, we propose Sylvie,a full-graph training system that not only improves the training throughput substantially but also maintains the model quality for universal GNNs. By harnessing the inherent information embedded in the graph data and model structure, Sylvie intelligently optimizes GNN training across three key dimensions: data, time, and execution. It identifies performance-relevant features of the input graph offline as subsequent optimization guidance. Subsequently, Sylvie devises an online convergence-maintenance strategy that adaptively integrates and aligns GNN-specific quantization and inter-epoch asynchronous training with the real-time training characteristics. Extensive experiments demonstrate that Sylvie surpasses existing GNN training systems by up to 17.2× speedup for both shallow and deep GNNs, without compromising the model accuracy.

Training large language models (LLMs) encounters challenges in GPU memory consumption due to the high memory requirements of model states. The widely used Zero Redundancy Optimizer (ZeRO) addresses this issue through strategic sharding but introduces communication challenges at scale. To tackle this problem, we propose Lins, a system designed to optimize ZeRO for scalable LLM training. Lins incorporates three flexible sharding strategies: Full-Replica, Full-Sharding, and Partial-Sharding, and allows each component within the model states (Parameters, Gradients, Optimizer States) to independently choose a sharding strategy as well as the device mesh. We conduct a thorough analysis of communication costs, formulating an optimization problem to discover the optimal sharding strategy. Evaluations demonstrate up to 52% Model FLOPs Utilization (MFU) when training the LLaMA-based model on 1024 GPUs, resulting in a 1.56 times improvement in training throughput compared to newly proposed systems like MiCS and ZeRO++.

Graph Transformer is a new architecture that surpasses GNNs in graph learning. While there emerge inspiring algorithm advancements, their practical adoption is still limited, particularly on real-world graphs involving up to millions of nodes. We observe existing graph transformers fail on large-scale graphs mainly due to heavy computation, limited scalability and inferior model quality. Motivated by these observations, we propose TORCHGT, the first efficient, scalable, and accurate graph transformer training system. TORCHGT optimizes training at three different levels. At algorithm level, by harnessing the graph sparsity, TORCHGT introduces a Dual-interleaved Attention which is computation-efficient and accuracy-maintained. At runtime level, TORCHGT scales training across workers with a communicationlight Cluster-aware Graph Parallelism. At kernel level, an Elastic Computation Reformation further optimizes the computation by reducing memory access latency in a dynamic way. Extensive experiments demonstrate that TORCHGT boosts training by up to 62.7× and supports graph sequence lengths of up to 1M.

Sub-model extraction based federated learning has emerged as a popular strategy for training models on resource-constrained devices. However, existing methods treat all clients equally and extract sub-models using predetermined rules, which disregard the statistical heterogeneity across clients and may lead to fierce competition among them. Specifically, this paper identifies that when making predictions, different clients tend to activate different neurons of the entire model related to their respective distributions. If highly activated neurons from some clients with one distribution are incorporated into the sub-model allocated to other clients with different distributions, they will be forced to fit the new distributions, which can hinder their activation over the previous clients and result in a performance reduction. Motivated by this finding, we propose a novel method called FedDSE, which can reduce the conflicts among clients by extracting sub-models based on the data distribution of each client. The core idea of FedDSE is to empower each client to adaptively extract neurons from the entire model based on their activation over the local dataset. We theoretically show that FedDSE can achieve an improved classification score and convergence over general neural networks with the ReLU activation function. Experimental results on various datasets and models show that FedDSE outperforms all state-of-the-art baselines.

While recent deep learning workload schedulers exhibit excellent performance, it is arduous to deploy them in practice due to some substantial defects, including inflexible intrusive manner, exorbitant integration and maintenance cost, limited scalability, as well as opaque decision processes. Motivated by these issues, we design and implement Lucid, a non-intrusive deep learning workload scheduler based on interpretable models. It consists of three innovative modules. First, a two-dimensional optimized profiler is introduced for efficient job metric collection and timely debugging job feedback. Second, Lucid utilizes an indolent packing strategy to circumvent interference. Third, Lucid orchestrates resources based on estimated job priority values and sharing scores to achieve efficient scheduling. Additionally, Lucid promotes model performance maintenance and system transparent adjustment via a well-designed system optimizer. Our evaluation shows that Lucid reduces the average job completion time by up to 1.3× compared with state-of-the-art preemptive scheduler Tiresias. Furthermore, it provides explicit system interpretations and excellent scalability for practical deployment.

Modern GPU datacenters are critical for delivering Deep Learning (DL) models and services in both the research community and industry. When operating a datacenter, optimization of resource scheduling and management can bring significant financial benefits. Achieving this goal requires a deep understanding of the job features and user behaviors. We present a comprehensive study about the characteristics of DL jobs and resource management. First, we perform a large-scale analysis of real-world job traces from SenseTime. We uncover some interesting conclusions from the perspectives of clusters, jobs and users, which can facilitate the cluster system designs. Second, we introduce a general-purpose framework, which manages resources based on historical data. As case studies, we design (1) a Quasi-Shortest-Service-First scheduling service, which can minimize the cluster-wide average job completion time by up to 6.5×; (2) a Cluster Energy Saving service, which improves overall cluster utilization by up to 13%.