每日文章

Randomized sketching is a core primitive in randomized numerical linear algebra. On modern hardware architectures, in particular on GPUs, the performance of sparse sketches is limited by memory traffic and atomic accumulation rather than floating-point throughput. This makes sketching a natural target for mixed precision, provided that low-precision accumulation does not degrade the embedding quality. We study mixed-precision GPU implementations of sparse oblivious subspace embeddings, focusing on a SparseStack generalization of the GPU CountSketch kernel of Higgins et al. SparseStack improves embedding quality relative to CountSketch on coherent inputs, but its additional nonzeros per column increase atomic-update contention and reduce throughput. We therefore implement FP16 SparseStack variants using deterministic round-to-nearest, exact stochastic rounding, and dithered rounding, and compare them with FP32 SparseStack, CountSketch, mixed-precision CountSketch, and FlashSketch. Our main empirical finding is that, for the tested regimes, SparseStack embedding quality is insensitive to the FP16 rounding rule. Deterministic, stochastic, and dithered rounding FP16 SparseStack produce nearly identical subspace distortion and sketch-and-solve least-squares accuracy across incoherent, coherent, and adversarial test problems. The dominant accuracy factor is the sketch distribution rather than the quantization rule: SparseStack variants substantially improve distortion on coherent inputs, while all methods behave similarly on incoherent inputs. Since deterministic rounding has the lowest overhead, it provides the best performance--accuracy tradeoff among the FP16 SparseStack variants.

Context-heavy agents place unusual pressure on the key-value (KV) cache: long prefixes are reused across many short turns, while concurrency determines whether the serving system can keep GPUs utilized. We study 4-bit KV-cache compression for this setting, using TurboQuant-style rotation and codebook quantization as a quality anchor and vLLM FP8 KV caching as the deployment anchor. We report three contributions. First, we frame 4-bit KV caching around multi-round agent workloads where task quality, cache residency, and serving throughput must be measured jointly. Second, we describe the practical design choices needed to make the 4-bit path robust, including asymmetric K/V treatment, Walsh-Hadamard rotation, QJL removal, and block-scale variants. Third, we present serving optimizations on AMD GPUs, including optimized decode-attention kernels and UltraQuant, an FP4 approximation path that uses FP8 queries, FP4 KV tensors, UE8M0 group scales, and native scaled-MFMA support on CDNA4. On a long-context, multi-turn agentic workload, UltraQuant cuts P50 time-to-first-token by 3.47x in the cache-pressured late rounds (2.3x across all rounds) and raises output throughput by 1.63x over the FP8 KV baseline.

Flow-matching text-to-speech systems achieve remarkable zero-shot quality but remain static after deployment: pronunciation errors on out-of-vocabulary proper nouns persist unless the model is retrained. We introduce FlowEdit, a life-long adaptation framework for frozen flow-matching TTS that learns pronunciation corrections as latent conditioning edits rather than weight updates. When corrective feedback is provided, FlowEdit optimizes a token-level perturbation in the text embedding space, then stores the correction in a Modern Hopfield Network serving as content-addressable episodic memory. At inference, corrections are retrieved via soft attention with a similarity gate, enabling fuzzy morphological matching. On our curated benchmark of 312 multilingual proper nouns across 18 language families, FlowEdit reduces target-word Phoneme Error Rate by 92.7% relative to the zero-shot baseline while maintaining identical general-speech quality. Corrections complete in approximately 15 seconds on a single GPU.

Modern LLM training breaks a core assumption behind offline batch samplers: the true training cost of a sample is only observable after preprocessing, augmentation, templating, tokenization, and multimodal visual-token expansion. Unless one pays for a preprocessing- and augmentation-dependent length cache, batch construction is therefore blind to the quantity that determines padding, memory use, and GPU saturation. We introduce Online Dynamic Batching (ODB), a DataLoader-side drop-in system that moves batch formation to this point of accurate observability while preserving DDP step alignment. We formalize this synchronization requirement as the Distributed Group Alignment Problem and prove deadlock-free bounded termination with default join-mode identity coverage and opt-in non-join sample-quota closure. ODB requires no model, optimizer, or attention-kernel changes and is released as online-dynamic-batching with lightweight trainer adapters. Across public 2B/8B Qwen3-VL runs on UltraChat/LLaVA/ShareGPT4o, ODB improves literal emitted-sample throughput vs. fixed-batch Standard by 1.58-2.51x on single-node Full FT/LoRA and 1.71-3.78x on two-node Full FT, with Standard-comparable quality; production MM-Mix reaches 4.43x. Against GMT/BMT offline token-budget oracles, ODB is within 15% on UltraChat/LLaVA and faster on high-CV ShareGPT4o: 2.24-2.39x single-node Full FT/LoRA and 3.06-3.69x two-node Full FT. Together, ODB occupies the online/drop-in regime for high-heterogeneity LLM fine-tuning: large throughput gains at Standard-comparable quality, formal DGAP guarantees, and no length-cache precompute or kernel rewrites.

Attention distillation, which trains one attention distribution to match another by minimizing their Kullback-Leibler (KL) divergence, is widely used in knowledge distillation, model compression, continual learning, and sparse-attention LLM training. However, existing approaches materialize both attention distributions before computing the KL reduction, incurring \(O(N_QN_K)\) memory and IO costs that become prohibitive at long context lengths. We present StreamKL, the first fused GPU primitive for attention KL divergence that eliminates this quadratic materialization. StreamKL derives a novel online formulation for the coupled two-distribution KL reduction, enabling a single one-pass forward kernel that streams query-key tiles through on-chip SRAM. For the backward pass, StreamKL recomputes attention probabilities tile-by-tile, avoiding storage of quadratic intermediates. We further design and implement efficient GPU kernels with dedicated optimizations. Experiments show StreamKL delivers up to \(43\times\) and \(14\times\) speedups over baseline methods in the forward and backward passes, respectively. Most importantly, StreamKL reduces the extra HBM footprint of attention distillation from \(O(N_QN_K)\) to \(O(1)\), enabling long-context distillation on a single GPU.

Benchmarks for LLM-generated GPU kernels (KernelBench, TritonBench, GEAK) score correctness through fixed-shape, small-sample allclose-style checks. The number of inputs varies between benchmarks. The shape, dtype, and tolerance are fixed for each kernel. We test that oracle empirically. We construct a controlled corpus of 24 Triton and CPU stand-in kernels (15 correct controls and 9 LLM-style buggy variants seeded with documented transcription errors) and re-evaluate it under op-schema-aware seeded fuzzing with a high-precision (fp64) CPU reference and per-(op, dtype) absolute tolerances. The seeded oracle flags 9 of 9 buggy kernels and passes 15 of 15 correct controls, at zero precision cost on controls. We extend the corpus to 26 ops (adding a flash-attention pair) and re-run the same protocol on five GPU classes (RTX 3060, A10, L40S, A100 SXM4, H100 NVL). The verdicts are identical across all five GPUs: 10 of 10 illusions caught and 16 of 16 controls clean. The corpus result is about LLM-style transcription bugs that the allclose-on-one-shape oracle certifies as correct, not about the bug rate of any specific deployed LLM. Every flagged failure replays byte-for-byte from a stored seed.

Mainstream LLM serving systems reuse prefix work mainly through paged or radix key-value (KV) caches. This is highly effective for high-throughput, high-concurrency serving, but it manages only one positional fragment of execution state: the KV cache. We study the opposite regime: low-latency, small-batch, on-device physical-AI serving, where interactive LLM agents, speech systems, and robot policies repeatedly branch, reset, interrupt, and re-enter under tight responsiveness budgets. We introduce execution-state capsules, a graph-bound checkpoint and restore mechanism for the complete restorable state at a committed boundary. FlashRT is a white-box, backend-facing kernel runtime whose evaluated NVIDIA CUDA backend runs captured graph plans over contiguous static buffers with no block-table indirection. Because the live state is a closed set of named buffers, a capsule can snapshot, restore, fork, or roll back the whole execution boundary, including KV, recurrent state, convolution state, MTP state, and metadata. This moves reuse from token-addressed KV fragments to graph-bound execution-state boundaries. On an RTX 5090, capsule restore is byte-exact at the stored-state level and token-identical under greedy decode. A KV-only ablation diverges, showing that recurrent state is load-bearing. GPU-resident snapshot and restore are sub-millisecond, and TTFT speedup over cold prefill grows from 3.9x at 2k tokens to 27x at 16k tokens. On Jetson AGX Thor and DGX Spark, the same correctness and structural properties hold. Capsules are not a replacement for high-throughput KV-cache serving; they define a complementary latency-first serving point for explicit execution-state reuse.

Diffusion models are now a dominant approach for high-fidelity image and video generation, yet scaling their training across GPU clusters remains challenging. Unlike transformer-only architectures, diffusion backbones commonly adopt UNet-style encoder-decoder structures with heterogeneous layers and long-range skip connections. Under conventional pipeline parallelism, these non-local dependencies force large skip activations and their gradients to traverse multiple pipeline boundaries, making peer-to-peer (P2P) communication a dominant bottleneck and substantially reducing pipeline efficiency. In this paper, we present PULSE, an automatic pipeline-parallel training strategy that makes skip locality a first-class optimization objective. PULSE eliminates skip-induced communication by collocating skip-connected encoder-decoder layers on the same device and caching skip activations locally for later use in backpropagation. To realize this placement while maintaining high pipeline utilization, PULSE co-designs: (1) a skip-aware dynamic-programming partitioner that balances heterogeneous stage workloads under symmetric collocation constraints, (2) an ILP-based schedule synthesizer that generates bubble-efficient wave schedules for the resulting stage-to-device mapping, and (3) a hybrid parallelism tuner that selects pipeline/data-parallel degrees and microbatch sizes under memory and network constraints. Our extensive experiments show that the volume of communication can be reduced by 89 percent, and the training throughput can be increased by up to 2.3x on communication-bound hardware, compared with state-of-the-art parallelism strategies.

Pretraining language models with extended context windows enhances their ability to leverage rich information during generation. Existing methods split input sequences into chunks, broadcast them across multiple devices, and compute attention block by block which incurs significant communication overhead. While feasible in high-speed clusters, these methods are impractical for decentralized training over low-bandwidth connections. We propose a compression method for communication-efficient context parallelism in decentralized settings, achieving a remarkable compression rate of over 95\% with negligible overhead and no loss in convergence. Our key insight is to exploit the intrinsic low-rank structure of activation outputs by dynamically constraining them to learned mixtures of subspaces via efficient reparameterizations. We demonstrate scaling billion-parameter decentralized models to context lengths exceeding 100K tokens on networks as slow as 300Mbps, matching the wall-clock convergence speed of centralized models on 100Gbps interconnects.

Gaussian splatting methods have become increasingly popular for neural reconstruction of the real world. However, they are often limited in scale and resolution due to compute and memory constraints. We present a multi-GPU Gaussian splatting approach that scales reconstruction to higher resolutions and larger scenes while abstracting away the code complexity typically associated with distributing a model. To accomplish this, we propose a PyTorch backend that distributes the Gaussian parameters and splatting operators across GPUs via CUDA unified memory and NVLink. Because distribution occurs at the operator level, the model code requires no explicit cross-device communication. More broadly, the backend exposes multiple GPUs as an aggregate PyTorch device and supports other PyTorch operators. We demonstrate city-scale reconstructions with street-level detail consisting of over 1 billion Gaussian splats, more than 25 times as many as the current state of the art.

Formal neural network verification -- proving that a network satisfies safety properties for all inputs in a specified domain -- is bounded in practice by GPU memory: standard implementations of bound-propagation algorithms (IBP, CROWN, \(α\)-CROWN) require weight and relaxation-coefficient matrices to reside entirely on one accelerator. We adapt two parallelism techniques originally developed for large-scale model training to the auto_LiRPA / \(α,β\)-CROWN verification framework. Tensor Parallelism (TP) shards both weight and \(A\)-matrices across GPUs, achieving \({\approx}2\times\) peak-memory reduction at \(P{=}2\); soundness is confirmed on VNN-COMP 2022 MNIST-FC benchmarks, though bound tightness degrades with the number of sharded zones due to forced IBP substitution for intermediate bounds inside sharded zones. Fully Sharded Data Parallelism (FSDP) shards only weight matrices with a per-layer AllGather, producing bounds that are bitwise identical to the single-GPU baseline: baseline memory drops by 80--90%, peak memory by 34--39% on wide MLPs. FSDP integrates cleanly with complete verification (\(β\)-CROWN + Branch-and-Bound) and with convolutional layers (BoundConv); a complete unsat result is obtained for CIFAR-100 ResNet-large (VNN-COMP 2024) under FSDP. Across all experiments the memory bottleneck in \(α\)-CROWN+BaB mode proves to be per-neuron alpha tensors, not weight matrices, pointing to the key direction for future work.

Context parallelism (CP) is essential for training large-scale, long-context language models, as it partitions sequences to reduce memory overhead. However, existing CP methods suffer from workload imbalance, inefficient kernels, and redundant communication due to static sequence sharding and key-value (KV) tensor communication. We present FlashCP, a load-balanced and communication-efficient framework for CP training. FlashCP introduces a sharding-aware communication mechanism to eliminate redundant KV communication and proposes a novel Whole-Doc sharding strategy that maximizes communication savings while maintaining balanced workloads. To efficiently combine Whole-Doc and Per-Doc sharding, FlashCP further designs a heuristic algorithm to search for near-optimal sharding plans. Extensive experiments show that FlashCP achieves up to 1.63x speedup over state-of-the-art CP frameworks across diverse datasets.

Pipeline parallelism enables training of large language models that exceed single-device memory, yet inter-stage activation communication becomes the dominant bottleneck when trained on low-bandwidth networks. Recent work in this area has proposed using fixed orthogonal projections to compress activations. However, this still results in a significant performance degradation and requires a number of non-standard adaptations to constrain the optimization. A natural alternative is to learn a low rank projection for each pipeline stage, however maintaining the necessary orthogonality of these projectors during training remains a challenge. We present Manifold Aware Projection Learning (MAPL), a method that treats inter-stage compression as a learnable orthogonal projection under explicit Stiefel manifold (orthogonal matrices) constraints. Rather than prescribing a fixed global subspace, MAPL lets each pipeline stage discover and continuously adapt its own task-optimal compression subspace via manifold-constrained steepest descent. To recover token-specific signals at stage boundaries, we introduce per-stage factorized anchor embeddings that allow for full-rank activation reconstruction with negligible communication overhead. We further show that we can incorporate residual vector quantization after projection with a streaming codebook synchronization protocol that amortizes dictionary communication. Across LLaMA models from 150M to 1B parameters we show that MAPL can be easily applied to the existing pipeline and can achieve high compression with neglibile performance degradation with a drastically improved tradeoffs in performance vs. compression compared to Subspace Networks.

Training modern machine learning models increasingly requires computation to be distributed across many accelerators. Data parallelism remains the default choice and is often paired with tensor-parallel sharding, but model parallelism becomes unavoidable once parameters, activations, or optimizer states no longer fit on a single device. This paper studies pipeline model parallelism through the lens of PipeDream (PD) (Harlap et al., 2018). Our first contribution is theoretical: we introduce Randomized PipeDream (RPD), a stale block-SGD abstraction that yields, to our knowledge, the first clean nonconvex convergence guarantee for a PD-style method. Our second contribution is a scaling diagnosis: we prove that the delay induced by steady-state PD grows as \(S^2 - S/2 + O(1)\) for \(S\) stages, so the stale-read contribution in the convergence theorem scales as \(Θ(γ^2 S^4)\), equivalently as \(Θ(S^4/K)\) in the tuned-rate form. Our third contribution is a comparison with LocalSGD, whose periodic model averaging trades weight staleness for synchronization bubbles. In our reported simulated-time experiments, the better-performing method depends on the objective: PD performs better on the quadratic objective and on a small language-modeling training-loss task, while for logistic regression LocalSGD becomes superior as the number of stages increases.

Pipeline parallelism is essential for large-scale model training, but existing asynchronous approaches often degrade convergence due to parameter mismatch between forward and backward passes. We propose Asynchronous Multi-Directional Pipeline parallelism (AMDP) to mitigate this issue while sustaining high utilization. AMDP limits the first stage of each pipeline to process at most two minibatches before backpropagation, bounding the number of parameter updates between forward and backward passes. To alleviate the resulting pipeline bubbles, AMDP launches multiple concurrent pipelines and adapts their number according to pipeline depth. In addition, AMDP accumulates gradients across minibatches and applies them in a single update, ensuring that only a bounded number of minibatches experience parameter mismatch, limited to within one optimization step. Experiments on GPT- and BERT-style models demonstrate that AMDP significantly accelerates training while preserving convergence.

Diffusion Transformers achieve strong video generation quality, but the quadratic cost of full attention limits efficiency. We introduce OSP-Next, an efficient text-to-video generation model that integrates sparse attention, parallelism, quantization, and reinforcement learning. OSP-Next uses a hybrid full-sparse attention architecture, where the sparse component is implemented with Skiparse-2D Attention. This fixed-pattern mechanism applies token-wise and group-wise sparse attention along spatial dimensions, leveraging locality while maintaining native compatibility with FlashAttention kernels. Based on the local equivalence of rearrangement in Skiparse-2D Attention, we further propose Sparse Sequence Parallelism (SSP), which partitions subsequences across ranks and switches sparse patterns through a single All-to-All communication. Compared with Ulysses Sequence Parallelism (SP), SSP provides a native parallel strategy for sparse attention and reduces communication volume by 75%. OSP-Next also incorporates HiF8 quantization to enable stable joint training with 8-bit quantization and sparse fine-tuning, and applies Mix-GRPO post-training to improve the performance of the sparse model. Experiments show that OSP-Next achieves a VBench total score of 83.73%, surpassing the Wan2.1 baseline. Under the 5-second 720P and 5-second 768P settings, OSP-Next achieves up to 1.64\(\times\) single-GPU speedup and over 1.52\(\times\) eight-GPU speedup on NVIDIA H200 GPUs. In addition, with only a 0.4% drop in VBench total score, OSP-Next-HiF8 achieves 1.69\(\times\) and 2.27\(\times\) speedups under the two settings on a single Ascend 950PR, demonstrating the efficiency and performance of OSP-Next across hardware platforms.

We introduce Fuzzy PyTorch, a framework for rapid evaluation of numerical variability in deep learning (DL) models. As DL is increasingly applied to diverse tasks, understanding variability from floating-point arithmetic is essential to ensure robust and reliable performance. Tools assessing such variability must be scalable, efficient, and integrate seamlessly with existing frameworks while minimizing code modifications. Fuzzy PyTorch enables this by integrating stochastic arithmetic into PyTorch through Probabilistic Rounding with Instruction Set Management, a novel library interfacing with Verificarlo, a numerical analysis compiler. The library offers stochastic rounding mode and a novel mode; up-down rounding. Comparative evaluations show Fuzzy PyTorch maintains model performance and achieves runtime reductions of 5x to 60x versus Verrou, a state-of-the-art tool. We further demonstrate scalability by running models from 1 to 341 million parameters, confirming applicability across small and large DL architectures. Overall, Fuzzy PyTorch provides an efficient, scalable, and practical solution for assessing numerical variability in deep learning, enabling researchers and practitioners to quantify and manage floating-point uncertainty without compromising performance or computational efficiency.

The rapid evolution of large language models (LLMs) has made geographically distributed training necessary due to GPU scarcity within a single cloud region. In such cross-region settings, Pipeline Parallelism (PP) is communication-efficient, yet scheduling PP remains challenging under heterogeneous inter-region bandwidth and regional electricity prices. Existing schedulers are either delay-first, incurring high electricity cost, or cost-first, relying on rigid resource allocation that prolongs Job Completion Time (JCT). They are also ineffective at optimizing execution order in multi-tenant environments, where long-running and bandwidth-intensive jobs can cause head-of-line (HoL) blocking and degrade overall performance. To this end, we propose BACE-Pipe, a bandwidth-aware and cost-efficient pipeline scheduling framework for LLM training across geo-distributed clusters. BACE-Pipe first introduces a dynamic job prioritization mechanism that optimizes execution order by jointly considering job characteristics (e.g., computation time) and real-time network utilization. It then employs a bandwidth-aware pathfinder to identify feasible cross-region pipeline paths that satisfy communication constraints, thereby preventing communication from stalling the pipeline. Among all feasible paths, a cost-minimizing allocator determines the optimal GPU placement strategy by preferentially assigning resources to regions with lower electricity prices. Consequently, BACE-Pipe mitigates HoL blocking, improves resource utilization, and simultaneously reduces both JCT and total electricity cost. Extensive simulations show that BACE-Pipe reduces average JCT by 27.9%--64.7% and total electricity cost by 12.6%--30.6% compared with state-of-the-art baselines.

Diffusion-based generation is increasingly powering production content pipelines; however, deploying these models at scale remains a significant challenge. Model weights frequently exceed the memory capacity of commodity GPUs, while the encoder, diffusion transformer (DiT), and decoder stages exhibit highly imbalanced computational and memory footprints. A natural remedy is disaggregated serving-running stages as separate services on heterogeneous GPUs-yet this introduces new bottlenecks, including stage handoff overheads and fast-changing workloads that make cross-stage provisioning and scheduling brittle. This paper presents DisagFusion, enabling asynchronous pipeline parallelism and elastic scheduling for disaggregated diffusion serving. First, DisagFusion introduces asynchronous pipeline parallelism that overlaps computation and stage-to-stage communication to reduce pipeline bubbles and mitigate network jitter. Second, DisagFusion employs a hybrid instance scheduling strategy that combines lightweight performance prediction with runtime feedback to continuously rebalance instance ratio across stages under workload shifts. We implement DisagFusion and evaluate it with modern diffusion models. Compared to a monolithic baseline, DisagFusion improves throughput by 3.4x-20.5x and reduces end-to-end latency by 18.5x, while enabling flexible, cost-efficient deployment across heterogeneous GPUs.

Modern serving systems for Mixture-of-Experts (MoE) models adopt hybrid data-expert parallelism: expert parallelism (EP) shards experts across GPUs to scale capacity, while data parallelism (DP) replicates attention layers across instances to process independent requests. Existing systems bind each request's attention, MoE communication, and KV cache to a single instance. Because attention latency scales with KV cache size while MoE communication latency scales with batch size, this binding cannot balance both simultaneously, producing EP stragglers; it also fragments KV memory across instances, inflating tail latency under long contexts. While existing context parallelism (CP) mitigates these constraints, its uniform parallelism degree incurs prohibitive communication and attention-side overheads. We present \work, which decouples MoE communication from KV cache placement and achieves dual balance through dynamic context parallelism (DCP). DCP assigns each request a context-parallel degree sized to its KV footprint: long requests distribute attention across multiple instances; short requests remain local. This dynamic parallelism effectively liquefies the KV cache across the cluster, balancing both the per-instance KV cache occupancy and batch sizes without unnecessary load-balancing costs. To bridge DCP with static execution, \work introduces an ahead-of-time (AOT) graph engine paired with a custom routing-based communication backend. Experimental results show that \work maintains up to \(1.88\times\)--\(3.27\times\) higher request rates under strict time-per-output-token (TPOT) service level objectives (SLOs). Furthermore, \work significantly mitigates stragglers, reducing P99 tail latency by up to \(1.79\times\)--\(2.12\times\).