Hardware y chips
636
77
Bongard: Training Machine Intuition
Human intelligence relies heavily on learned intuition: recognising patterns and judging situations without explicitly unfolding every intermediate step. We introduce Bongard, an open-weight System One model that treats machine intuition as an independent capability to design and train. A T5Gemma 2 4B-4B encoder-decoder separates reading the evidence from making judgments. The encoder reads the state bidirectionally together with the question instructions, and separate decoder branches share this encoding, so many judgments about the same situation require only one reading of the state. A trained head returns probabilities over the supplied candidates without generating text. Training proceeds in three stages, from supervised judgments to semantic relationships to action outcomes, and each stage updates all 7.09 billion trainable parameters on one Blackwell GPU. Joint-embedding post-training raises accuracy on held-out rephrasings from 75.7% to 85.9%. A sandbox stage then learns outcome distributions from action rollouts and exact oracles, raising accuracy on a frozen sandbox panel from 50.6% to 64.8%. On DecisionBench, the final model reaches 78.05% accuracy over 23,900 decisions and ranks fourth of 61 systems in the public comparison. On one RTX PRO 6000, its median latency is 36 ms for short requests, and 32 questions about one state take 221 ms. Bongard demonstrates that machine intuition can be systematically trained via representation learning and outcome feedback, providing an open, efficient alternative for high-throughput decision workloads.
Shared Phase and Retention Control for Efficient Adaptive Spectral Recurrence
As new evidence arrives, a sequence model must update what it remembers and how memory influences predictions. While Transformers incur computation and cache costs scaling with context length, fixed-state recurrent models offer constant-memory inference. However, linear and spectral recurrences traditionally rely on static transitions, failing to dynamically revise how stored representations decay or rotate. While recent selective architectures introduce input-dependent transitions, they assign independent controls to every memory mode, coupling control cost to state capacity. We show that high-dimensional spectral memory does not require high-dimensional control, and introduce Shared Phase and Retention Control for Efficient Adaptive Spectral Recurrence (SPARC). SPARC employs just two input-dependent scalar signals to coordinate memory retention and phase rotation across heterogeneous complex modes, while preserving mode-specific baseline timescales and frequencies. Its diagonal affine recurrence supports parallel associative scans for sequence-level BPTT as well as exact structured Real-Time Recurrent Learning (RTRL) for online credit assignment. Across partially observable continuous control, POPGym, and sequence classification, SPARC achieves a 9.09% relative return improvement on Walker-P and a 1.36% relative accuracy gain on FordA over second-best methods. On an NVIDIA Blackwell GPU, our implementation reduces recurrent-mixer training latency by 18.2%-34.2% in fixed-token workloads and accelerates scans by 3.1x-4.7x over an optimized RG-LRU baseline. These results show that two shared control signals can efficiently govern adaptive spectral memory across online and full-sequence settings. Code is available at https://github.com/Botwwt/sparc.
Bring near-Astra intelligence to everyday work with GPT-6.1 Sol on Amazon Bedrock
GPT-6.1 Sol is now generally available on Amazon Bedrock, bringing stronger reasoning to coding, computer use, and professional workloads that run frequently.
AI Native by Design: Lessons Learned from Building NVIDIA TensorRT Model Connect
Parallel work, model-family isolation, reversible changes, and GPU-backed validation shaped an open source project designed around coding agents NVIDIA TensorRT...
STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization
Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Directly quantizing recurrent states to low precision often leads to severe accuracy degradation, as quantization errors propagate through successive state updates. We discover that the impact of these errors depends on two complementary dimensions: temporally, errors in long-lived memory can persist across many decoding steps; spatially, errors in different key rows affect model outputs differently, while state magnitudes vary substantially along both rows and columns. Motivated by these observations, we propose STEPQuant, a spatial-temporal post-training quantization framework for Delta-rule recurrent states. STEPQuant allocates precision according to error magnitude and memory lifetime, and jointly fits key-row and value-column scales based on state distributions and key-row impact on output error. Experiments on Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct across both long- and short-generation benchmarks show that STEPQuant closely matches FP32-state accuracy under a nominal 6-bit budget and outperforms uniform INT8 in its 4-bit configuration. Integrated into SGLang with optimized GPU kernels, 6-bit STEPQuant achieves over 5x recurrent-state compression and reduces total serving memory by up to 68.7%. Our code is available at https://github.com/Dreamer-Toby/STEPQuant.
Probe-Space Preconditioning for Fast and Stable Zero-Order Training
Backpropagation (BP) dominates deep learning but imposes a massive memory tax. For example, training OPT-30B with Adam requires 600GB of GPU memory (assuming batch size 8 and sequence length 2048). Alternatively, zero-order optimization (ZOO) trains in inference-mode (requiring only 60GB for the same model): no stored activations, no gradients, and no optimizer states. However, ZOO convergence has lagged behind BP. In this work, we evaluate two methods to close this gap. First, we show that reallocating training compute budget from many steps to large effective batch sizes with many perturbations (or probes) but fewer steps, allows 1SPSA (Spall, 1992) to outperform zero order methods like MeZO (Malladi et al., 2023) with less training compute. Next, we introduce 1.5-SPSA, adding a single "clean" forward-pass per step to 1SPSA to calculate a cheap diagonal preconditioner in probe-space, which improves convergence rate and convergence by down-weighting high curvature directions. Benchmarking on 6 post-training datasets on both Qwen3 and OPT model families, we show that 1.5-SPSA achieves State-of-the-Art results over previous ZOO solvers with much less optimization steps. For example, we train OPT-13B (for direct comparison to MeZO) and find 1.5-SPSA achieves +3.1% accuracy on SST-2 over both MeZO and BP in only 70 steps vs. MeZO's 100,000 steps. Finally, we combine an 8-bit-packing random generator, triton fused unpack/apply kernels, and distributed parallelism to achieve fast and stable training of models as large as OPT-30B in-place on commodity GPUs (e.g. A100).
Mira: Memory-Efficient MoE Inference Using Adaptive Caching and Predictive Expert Staging
Mixture-of-Experts (MoE) models are a compelling architecture for scaling model capacity, making them especially attractive for deployment on resource-constrained, single-GPU systems. However, this benefit is difficult to realize because expert parameters dominate memory, and token-level routing is dynamic, unpredictable, and skewed. Prior work using offloading and caching remains fundamentally reactive, as systems wait for router outputs before moving experts, leading to inefficient cache utilization and an inability to overlap transfers with compute under tight VRAM budgets. To address these challenges, we propose Mira, an algorithm-system co-design that enables high-capacity MoE inference on a single GPU. Mira shifts from a reactive to a proactive stance by coupling predictive expert management with a tailored quantization format. It introduces lightweight per-layer predictors that anticipate expert usage two layers ahead, enabling proactive prefetching. These predictions feed a two-tier HOT+STAGE GPU cache managed by token-level routing telemetry to retain frequently used experts while staging predicted ones. To minimize transfer overhead, Mira implements a custom compression for expert parameters, which reduces metadata and improves packing efficiency, while minimally degrading accuracy. Mira is implemented as a fully integrated runtime that coordinates predictors, caching policies, and quantized transfers to maximize overlap between communication and compute. Our experiments show that Mira reduces expert-induced stalls. Compared against state-of-the-art baselines, Mira achieves a 5.71x speedup in average throughput on a memory-constrained GPU. It accelerates Time-to-First-Token by 11.71x and achieves a 3.84$x average speedup in beam search inference, demonstrating its effectiveness across diverse inference scenarios.
Purlin: Separating Orchestration from the Datapath of Collectives
Distributed inference depends on GPU collective communication that must keep pace with evolving hardware and specialized workloads. However, existing collective implementations often couple semantics, orchestration (where and when data moves), and the datapath (how data moves). This coupling makes it costly to adopt new hardware mechanisms and customize communication for applications. We present Purlin, a scale-up communication framework that separates these concerns. At the top of Purlin, we specify collectives as a naming of an input and output layout and a copy or reduction operation. In the middle, we introduce a shared orchestration protocol, Stage, Notify, And Consume (SNAC), which derives coordination from these specifications. Below SNAC sits a hardware-specific datapath we call Atom, which implements two key data movement primitives for collectives: copy and reduce. This separation lets us customize collectives and adopt new hardware mechanisms while reusing orchestration via SNAC. We evaluate Purlin on A100, H200, and B200 GPUs. Across seven collectives, Purlin achieves latency speedups of up to 5.14x and bandwidth improvements of up to 4.50x over baselines. Integrated into SGLang, Purlin improves offline LLM serving throughput and interactivity by 1.13x on average and up to 1.37x over baselines. For online LLM inference, Purlin improves interactivity by 1.26x on average and up to 2.85x, with the largest gain occurring under overload. For diffusion image generation, Purlin reduces end-to-end latency by up to 1.13x.
AeroManip-VLA: Scalable Vision-Language-Action Learning for Aerial Manipulation with RL-Generated Demonstrations
Aerial manipulators extend robotic manipulation into 3D workspaces that are difficult for ground-based robots to access, creating new opportunities for general-purpose manipulation. However, extending Vision-Language-Action (VLA) models to aerial robots introduces distinct challenges due to the tight coupling between manipulation and flight, continuously changing observations, and safety-critical physical interactions. These challenges demand diverse training data and systematic policy evaluation, yet collecting demonstrations and evaluating policies directly on physical aerial platforms are costly, difficult to scale, and hard to repeat under controlled conditions. We present AeroManip-VLA, a scalable benchmark for aerial VLA data generation and policy evaluation. AeroManip-VLA provides a GPU-accelerated simulation framework with low-level payload-aware flight and manipulation control in massively parallel environments. Building on this framework, we combine reusable reinforcement learning policies with expert task rules to automatically generate demonstrations without human teleoperation across diverse objects, environments, and randomized initial conditions. The generated data include basic skills such as grasping and placing, as well as long-horizon tasks that require both navigation and manipulation. We further introduce automated event labeling and trajectory categorization to filter demonstrations. These mechanisms enable fine-grained analysis of task progress, behavioral outcomes, and safety-related failures. Finally, we evaluate a range of imitation learning and VLA baselines across different task settings, revealing their performance characteristics and failure modes. Together, AeroManip-VLA enables scalable aerial manipulation data generation, structured trajectory analysis, and systematic VLA evaluation in simulation prior to real-world deployment.
Telescopic Language Models
One deployed language model must often serve many compute budgets, yet serving each budget still means a separate training or compression run per point. We train a Telescopic Language Model (TLM) to be that continuum: a nested-capacity Transformer supervised by stochastic prefix supervision with a full anchor. At every step, one randomly truncated prefix of the capacity axis is trained against the full next-token target, alongside one full-capacity pass, so the trained artifact is a valid language model at every depth. Two forward-backward passes per step, no architectural change, nothing extra at inference. Fixed-exit suites such as Matryoshka Language Model Suites (MLMS) occupy one point in this design space, and the point has a cost: supervising only a few fixed exits leaves the nested model at chance level everywhere else (perplexity 10^2-10^5 in our baselines). On a 200M proxy suite (20B FineWeb-Edu tokens, identical data stream for all methods), a single TLM run is a valid language model at every one of its twenty layer prefixes, in perplexity and on perplexity-sensitive downstream tasks, reducing the area under the quality-budget curve by 43-44% relative to the fixed-exit suites while matching them at full capacity, at ~12% lower GPU cost per run. The prefix sampling density is a dial: concentrating it on a few depths recovers fixed-exit quality there at the price of the continuum, so the operating points become a training-time choice rather than an architectural one. These results indicate that the training objective, not the nesting itself, is what makes a model elastic.
KV-streams for Efficient Compaction in Agentic Reinforcement Learning
Scaling the horizon of agentic LLMs is bottlenecked by the need to fit ever longer context traces in GPU memory. Context compaction has been the most popular mechanism to alleviate this issue, keeping GPU memory constant for a given trace. Unfortunately, most compaction strategies rely on prefilling the LLM context many times over, hindering training throughput. To alleviate this bottleneck and enable efficient trainable compaction, we propose KV-streams, a plug-and-play strategy compatible with any compaction strategy that substantially increases throughput while showing no evidence of hindering performance. KV-streams enable scalable compaction by streaming the KV cache forward rather than flushing it after each compaction. We show that KV-streams enable three different compaction strategies, achieving a 2.6 to 5x wall-clock speedup in training. Beyond efficiency, we find that the streamed KV cache can act as a recurrent state, carrying forward information that has long since disappeared from the context. Specifically, in a controlled setting we show that, contrary to prior work, RL alone is all that is needed for this behavior to emerge. Overall, we show KV-streams to be an efficient and lightweight plug-and-play addition to any post-training pipeline.
Learned Preconditioning for a Primal-Dual Interior-Point Method
Interior-point methods (IPMs) are among the most widely used algorithms for constrained optimization, yet their Newton-based search directions require costly second-order information and large linear-system solves. Learning to optimize offers cheaper updates learned from data, but the singular behavior of logarithmic barriers near constraint boundaries makes IPMs highly sensitive to perturbations, complicating both warm starting and learning reliable updates. We introduce pdLIP, an IPM for smooth nonlinear programs that integrates learned preconditioning with pdProj, an all-shifted primal-dual projected-search IPM. A shared coordinate-wise recurrent network predicts a positive diagonal preconditioner that scales the right-hand side of the reduced Newton system for the primal step, and the remaining slack and multiplier directions are recovered analytically. The learned iterations avoid Hessian evaluations and Newton-system solves, using only first-order and coordinate-wise operations amenable to GPU parallelization. Training is self-supervised, with a loss based on a penalty-barrier merit function and the residual of perturbed optimality conditions, requiring neither target directions nor precomputed solutions. Primal and dual shifts mitigate the barrier's sensitivity to perturbations near constraint boundaries, enabling effective warm starting. Across four classes of 200-dimensional convex and nonconvex constrained problems, pdLIP warm starts reduce pdProj refinement iterations by 63-67% compared with cold starts at the same KKT residual tolerance of 10^{-8}, with negligible warm-start generation cost relative to the subsequent pdProj solve. Improvements persist on box-constrained QPs with 1000 variables and extend to applications including portfolio optimization, support vector machines, and a nonlinear control example.
GPUPhysBench: Benchmarking Coding Agents for Correct and Efficient GPU Physics Simulation
Writing fast GPU code for physical simulation is difficult: implementations must preserve numerical accuracy while handling irregular data access, synchronization, and iterative solvers. We introduce GPUPhysBench, a benchmark of 50 tasks testing whether coding agents can meet these demands. Tasks cover fluids, deformable solids, and granular materials, from individual simulation operators to complete simulators. Agents write, compile, test, and optimize GPU code with access to a NVIDIA GPU under fixed time budgets. We report pass rates and runtime performance relative to expert-optimized reference implementations. In a single-attempt evaluation of six frontier model-harness pairs, the two strongest pass all 50 tasks, but even the fastest reaches at least 0.9 the reference speed on only 22% of them, and no submission is more than 5% faster than the reference. The largest gaps arise in collision detection, constraint solving, and iterative solvers. GPUPhysBench brings physical simulation workloads to coding-agent evaluation, testing both the ability to implement numerical methods correctly and the ability to make them run efficiently.
SANTA++: Sampling Attention through Representative Keys
Attention often concentrates on a small subset of tokens in the context, but which subset matters changes from one query to the next. To exploit this changing structure, we introduce SANTA++, a training-free stochastic attention method that uses representative keys for memory-efficient selection without scanning the entire key-value (KV) cache. Cached keys are organized into teams, and the query scores one representative from each team to decide which teams to sample. We compute exact attention scores within the sampled teams and reweight each team's contribution by the inverse of its inclusion probability. This importance sampling correction estimates attention over the full cache, with a sampling budget that lets us trade memory reads for accuracy. Remarkably, with 32 or 64 sampled teams, SANTA++ uses 16% to 22% of dense attention's KV reads and retains 94% to 99% of the dense-attention baseline's scores on LongBench v2 and HELMET's retrieval-augmented generation subset, and 85% to 91% on RULER, with Qwen2.5-7B-Instruct at 32K context. With 31 sampled teams, our GPU implementation delivers a 1.69attention speedup over the dense FlashAttention baseline at 32K context. By reducing the number of cache entries read, SANTA++ in principle complements architectures with compressed KV representations, such as multi-head latent attention. Our kernels are available at: https://github.com/OPUSLab/santapp-kernel-demo.git.
Build real-time voice applications with vLLM-Omni on SageMaker AI – Part 1
Deploy a text-to-speech model on Amazon SageMaker AI with the AWS vLLM-Omni Deep Learning Container and stream generated speech over a persistent bidirectional connection. This Part 1 tutorial deploys Qwen3-TTS and streams speech through a Gradio application.
Generate images and video with vLLM-Omni on SageMaker AI – Part 2
Deploy two generative media models from one AWS vLLM-Omni Deep Learning Container on Amazon SageMaker AI. Generate an image with FLUX.2-klein through real-time inference, then animate it into video with Wan2.1-VACE through asynchronous inference, and retrieve the MP4 from Amazon S3.
How NVIDIA DSX MaxLPS Maximizes AI Factory Throughput and Efficiency
Every unused watt is capacity left on the table. AI factories are typically provisioned for the unlikely moment when every GPU reaches peak power, creating a...