Papers for
cloud ai infrastructure teams
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Factored optimizer speeds up large language model training efficiently
Factored One-sided Adam-Moment for Practical and Scalable SOAP
Abstract: Second-order optimizers (e.g., Shampoo, SOAP) have emerged as compelling alternatives to Adam/AdamW and promise faster per-step convergence than AdamW in transformer training. However, they still remain costly in practice, and even SOAP still pays substantial optimizer-state memory and eigendecomposition overhead. We ask whether the benefits of SOAP can be retained at nearly the efficiency of AdamW. In this paper, we answer yes with a lightweight SOAP variant, Factored One-sided Adam-Moment (FOAM). The path to practical impact is twofold: first, FOAM employs input-only preconditioning, dropping the output-side Kronecker factor to reduce memory and remove one eigendecomposition. The Kronecker structure of the Hessian shows that this is exact for a diagonal output-side curvature and nearly lossless for a near-diagonal one, as we observe empirically. Second, FOAM replaces SOAP's dense rotated second moment with a factored Adam-moment estimator based on row-by-column statistics. Under the K-FAC Fisher assumption with a diagonal output curvature, this estimator exactly recovers SOAP's dense second moment, and we measure only a 15-20% K-FAC violation, with negligible effect on the update. Our empirical results support the claim that FOAM matches strong SOAP and Muon baselines in validation loss across scales of large language model (LLM) pre-training, uses half of SOAP's optimizer state, runs at Muon-class wall-clock time, and retains robustness across a three-fold learning-rate range. The results hold from 124M to 760M parameters, demonstrating that SOAP's gains need not require SOAP's cost. Code: https://github.com/naver-ai/foam
Modality aware pruning saves memory in multimodal large language models
Modality-Aware Expert Pruning for MoE-Based Multimodal Large Language Models
Abstract: Mixture-of-Experts (MoE) enables multimodal large language models (MLLMs) to scale capacity efficiently, but expert parameters dominate memory even though only a few experts are activated per token: they account for over 90% of model size in modern MoE-MLLMs (e.g., 54 GB out of about 58 GB in BF16 for InternVL3.5-30B-A3B and Qwen3-VL-30B-A3B). Expert pruning has emerged as an effective compression strategy, but existing methods aggregate importance uniformly across tokens, ignoring that the same expert may serve different roles for vision versus language processing. To address this, we propose MAEP (Modality-Aware Expert Pruning), a training-free framework that decomposes expert importance by modality in a single forward pass. MAEP introduces (1) Delta-H Output ($ΔH$), a redistribution-aware metric measuring the MoE output change when an expert is removed, and (2) low-attention vision tokens as a vision-side pruning signal that is dataset-stable: its expert-importance ranking remains stable across different datasets. Text-side importance serves as a preservation signal while vision-side importance serves as a pruning signal, and experts are ranked globally across layers. Experiments on three MoE-based MLLMs across multimodal and text benchmarks show that MAEP outperforms modality-agnostic baselines on the majority of multimodal benchmarks while remaining competitive on text-only tasks, reducing the BF16 expert weight footprint by about 27 GB (about 47% of total memory) so that the full weight footprint of an approximately 58 GB model fits within a single 40 GB GPU. Code is available at https://github.com/nota-github/MAEP.
MoE language models gain accuracy with symmetric FP4 reinforcement learning
Aligning Training and Rollout Policies for End-to-End FP4 RL on Mixture-of-Experts LLMs
Abstract: In this work, we study end-to-end FP4 reinforcement learning (RL) post-training for Mixture-of-Experts (MoE) large language models. Recent work on FP4 RL has introduced effective recovery recipes that restore accuracy under aggressive low-precision quantization for dense models, but we find that the same approach yields no improvement over the naive FP4 baseline in the MoE setting. We conduct an in-depth study to understand why. Our analysis identifies the root cause in the routing mechanism unique to MoE: because the router makes discrete expert-selection decisions, training-inference policy alignment becomes acutely sensitive to whether the same recovery technique is applied symmetrically on the training and inference sides. When an asymmetric recipe is used-including the dense-model approach, which corrects only the rollout-training-inference alignment deteriorates sharply, as measured by the Pearson correlation coefficient between training and inference probabilities, and router selection matching rates. Building on this finding, we propose MoeResQ, a quantization-recovery recipe for MoE RL post-training that enforces symmetry between the training- and inference-side corrections. MoeResQ restores training-inference alignment and substantially closes the accuracy gap to full precision, establishing symmetric recovery as the key principle in FP4 RL for MoE models.
Mixed ADC resolution improves energy efficiency and accuracy in analog AI computing
SMART-AIMC: Sensitivity-Aware Mixed ADC Resolution with Fine-Grained Tunability for Analog In-Memory Computing
Abstract: Analog In-Memory Computing (IMC) is a promising paradigm for accelerating GEMM-based AI workloads by executing computations directly within the memory array in the analog domain. While IMC delivers substantial energy and throughput improvements, it can suffer from significant task accuracy degradation due to analog-to-digital converter (ADC) quantization noise, which occurs at the GEMM output and is thus different from weight/activation quantization, which occurs at the GEMM input. Mixed ADC resolution allows IMC modules to choose between high- and low-resolution ADC modes with different energy costs, offering a compelling accuracy-energy trade-off. However, selecting ADC resolutions across IMC units while respecting hardware granularity constraints remains challenging. In this paper, we propose SMART-AIMC, a Sensitivity-aware Mixed ADC Resolution framework with fine-grained Tunability for Analog In-Memory Computing. SMART-AIMC correlates the algorithmic task loss to hardware ADC quantization noise in the form of an ADC Channel Sensitivity (ACS) metric. Using this metric, SMART-AIMC systematically, and in a disciplined manner, allocates ADC resources to critical computations. Furthermore, we introduce a sensitivity-based channel reordering scheme that overcomes the hardware granularity constraints with no runtime overhead and approaches the optimal ADC quantization possible for the model. Evaluated across diverse workloads, SMART-AIMC unlocks a superior accuracy-energy Pareto frontier, providing (1) higher task accuracy at matched energy budgets and (2) fine-grained tuning across efficiency targets, from aggressive energy savings to near-digital inference accuracy. At the latter operating point, SMART-AIMC achieves $6.99\times$ higher modeled system-level energy efficiency for linear projections than the INT8 digital baseline with negligible accuracy degradation.
RankLane speeds up mixture-of-experts model prefill on AI hardware
RankLane: Efficient MoE Prefill on Accelerators with Ranked Expert Capacities
Abstract: Mixture-of-experts (MoE) models expand model capacity by activating only a subset of experts for each token. On accelerators with static execution plans, expert capacities and placement are fixed before routing reveals each expert's load. Uniform worst-case capacities waste work on padding; smaller capacities tied to expert identities can overflow. We present RankLane, which ranks each device's experts by current load and assigns them to fixed-capacity lanes. Runtime indices select experts and token rows while tier populations, shapes and device placement remain fixed. An aggregation graph preserves device ownership and reduces each token's local contributions on one core. Routing calibration and measured execution costs guide tier selection; overflowing chunks rerun at full capacity. We evaluate full-model prefill of Qwen3-30B-A3B and Gemma 4 26B-A4B with MXFP6 weights on four Cloud AI 100 Ultra SoCs, using 128-2048-token chunks and including overflow detection and in-program recovery support. Against a dense baseline, RankLane achieves end-to-end prefill speedups of 1.27-4.18x on Qwen3 and 1.93-3.46x on Gemma.
TranSAGE boosts transformer model speed and energy use with spatial pipelines
TranSAGE: Automating Spatial Accelerator Generation for Large-Scale Transformer Models
Abstract: Large-scale Transformers have achieved remarkable success across NLP, computer vision, and robotics, but their growing computational and memory demands challenge efficient inference. Spatial accelerators offer advantages in on-chip bandwidth and energy efficiency, yet designing them for large models remains costly and time-consuming. Existing accelerator generators typically focus on single-operation support, overlooking the benefits of multi-operation pipelining, which limits throughput and energy efficiency. We propose TranSAGE, a compiler-based framework that jointly optimizes scheduling for large-scale Transformers under multi-operation pipelining and generates corresponding accelerator implementations. Our approach integrates a design space exploration (DSE) engine with multi-operation dataflow optimization and a dynamic programming-based graph partitioning algorithm to efficiently identify pipeline-friendly operation clusters and their optimal dataflow parameters. We further design spatial architecture primitives and lowering passes to automatically generate efficient RTL implementations. Evaluations on four representative LLMs demonstrate that TranSAGE achieves up to 2.41x higher throughput and 13.04x higher energy efficiency than commercial GPU/NPU baselines. Compared with state-of-the-art accelerator generators, TranSAGE achieves up to 11.33x higher throughput and 6.07x higher energy efficiency. Our code is available at https://github.com/sjtu-zhao-lab/TranSAGE.
Vision language model reduces parameters by looping layers with adapters
Retrofitting Recurrent Vision-Language Models from Pretrained Backbones
Abstract: Recent work shows that depth recurrence can be retrofitted into pretrained language models. In this work, we adapt this paradigm for visual understanding, using depth recurrence to reduce the parameter footprint of a pretrained vision-language model (VLM). To this end, we prune layers from the language backbone of the VLM and loop over a subset of the remaining layers. At each step of the loop, a linear adapter merges the block's output with the token representations from the preceding layers and projects the fused representation back into the block. As the pruning impacts the original model performance, we propose a three-stage recovery training: first, we leverage a language-only pretraining to recover the performance of the language backbone, followed by an alignment of the vision-language projector, and a full fine-tuning of the model via visual instruction tuning. In the first and the last stage, we distill the outputs of the pretrained non-recurrent model into the recurrent model, which further improves the recovery. We evaluate the proposed approach on the TinyLLaVA-1.4B backbone without introducing any new vision-language data during retrofitting. Our experiments demonstrate that the retrofitted recurrent model outperforms the unpruned baseline trained with the same recovery data while storing 23.3% fewer language parameters. We further show that the proposed retrofitting recurrence outperforms the setup of converting a recurrent LM into a VLM on the same data. Additionally, our ablations show how design choices, such as the adapter between recurrences and a larger recurrent block, lead to the improved performance of our approach. Finally, we find that our proposed model gains on reasoning-oriented benchmarks, whereas its performance on basic visual understanding benchmarks remains on par with the baseline.
Static sparse memory boosts pretrained model performance efficiently
STAPLE: Static Sparse Upcycling of Pretrained Checkpoints via Embedding Modules
Abstract: Sparse computation can scale the parametric capacity of LLMs without proportionally increasing active computation, but pretraining sparse architectures remains expensive and data-hungry. Existing sparse upcycling methods mainly target Mixture-of-Experts models, where expert homogeneity, sample inefficiency, and system overheads can limit practical scalability. Static sparse memory offers predictable access and token-specific updates that can help address these challenges, but naively integrating it into pretrained computation yields poor returns. In this work, we introduce STAPLE, a recipe built upon STEM for integrating static sparse memory into pretrained models. Whereas STEM replaces FFN up-projections with token-indexed embeddings, STAPLE retains them and adds the embeddings as a residual pathway, preserving pretrained computation while learning complementary token-specific information in a scalable manner. On OLMo-2 1B, STAPLE improves over continued dense training by 4.27 points on HumanEval, 2.80 on MBPP, 2.27 on GSM8K, and 1.65 on DROP. Average performance increases with static capacity, and improvements persist on the larger OLMo-2 7B backbone. STAPLE also outperforms static sparse baselines such as STEM and an Engram-style model, and achieves a higher average score than an MoE upcycling baseline while decoding 2x faster with less than half its GPU model memory and measured GPU energy per token in our throughput-oriented decoding evaluation. Importantly, despite introducing billions of parameters late in training, our host-offloaded implementation keeps generation throughput within 1.5% and training throughput within 3% of the dense baseline, while caching only the ~5% most frequent embeddings on GPU. These results show that static sparse memory can provide a sample-efficient, scalable, and sustainable path for expanding pretrained models without sparse pretraining from scratch.
Looped transformers cut memory use with low-bit residual quantization
ResidualQuant: KV Cache Quantization for Looped Transformers with 2-Bit Residuals
Abstract: Looped Transformers improve parameter efficiency by repeatedly applying shared Transformer blocks over multiple recurrent loops, increasing computational depth without increasing the parameter count. However, KV cache memory still scales with the number of loops, becoming a key memory bottleneck that limits batch size and inference throughput. KV cache quantization can alleviate this bottleneck, but existing methods often suffer substantial accuracy degradation at aggressive low-precision regimes. We observe that looped Transformers offer a unique opportunity: KV states across loops are highly similar. Based on this observation, we propose ResidualQuant, which uses the final-loop KV states as a reference and represents the remaining loops with low-precision residuals. Our method further combines least-square scaling and rotations applied to the residuals, as well as loop-wise mixed precision, to enable accurate quantization down to INT2 while retaining efficient reconstruction. Across multiple looped Transformer models and mathematical reasoning and code generation benchmarks, ResidualQuant consistently improves the accuracy-memory tradeoff over state-of-the-art rotation-based KV quantization. In particular, our method retains accuracy close to BF16 under mixed-precision settings while reducing theoretical KV storage by 80.7%, achieving up to 13.0% higher accuracy than the rotation-based baseline at the same memory budget. On an RTX 5090, the reduced KV memory traffic improves fixed-batch decode throughput by up to 2.73x, while the smaller memory footprint enables up to 2x larger batches, improving peak throughput by up to 4.15x.
Looped language models save memory and computation on easy tokens
Enabling Dynamic Computation in Looped LMs
Abstract: Looped LMs are parameter efficient and promise dynamic computation (saving memory and FLOPs on easy tokens). However, state-of-the-art open Looped LMs trained with this dynamic computation capability (Ouro models) do not realize it in practice as each loop iteration (depth) requires its own level of KV-cache, necessitating all loop computations. Moreover, Ouro's early-exit prior is enforced on each token equally, which results in static lower-depth like processing of all tokens regardless of difficulty. In this work, we propose a simple "best-available" KV caching strategy that works out-of-the-box, creating a new frontier in the performance vs depth space. Our approach enables up to 30% reduction in FLOPs and KV memory while retaining full-depth performance, showing the true flexibility of Looped LMs. Furthermore, training looped LMs with awareness about this KV caching strategy improves performance and efficiency. Finally, we apply a small but effective fix to the early-exit prior enforcement objective that makes tokens exit at truly heterogeneous depths based on effort. Our findings are validated on Ouro models as well as smaller looped LMs pre-trained from scratch.
Activation denoising improves large language model quantization speed and accuracy
Activation Denoising: A Robustness View on Parallel vs Sequential LLM Quantization
Abstract: Post-training quantization is a powerful tool for compressing large language models. The most scalable methods quantize every layer in parallel, but quantization errors then compound through the residual stream, as no layer corrects for the errors of the layers before it. Sequential quantization accounts for this error compounding by re-calibrating each layer on the already-quantized outputs of its predecessors, yielding stronger results but at the cost of a serial schedule that becomes a bottleneck at scale. As a solution, we propose parallel quantization with activation denoising, which recovers much of the sequential benefit while keeping quantization fully parallel. Rather than re-calibrating layer-by-layer, we take a robustness perspective and model the upstream error as noise, regularizing to be robust to it through a preprocessing step followed by metric-weighted rounding. Applied at every layer, this regularization forms a depth-compounding smoothness penalty that dampens how strongly quantization errors amplify through the model. Unlike orthogonal rotations commonly used in quantization, which must preserve the model's function, we multiply the weights by a more general linear transformation. We find that the two are complementary and their effects compound. Empirically, our robustness regularization recovers a significant part of sequential quantization's benefit in a single parallel pass, at a fraction of its time. Overall, by treating compounding quantization errors as a robustness problem, we offer a principled foundation for more efficient and accurate LLM quantization at scale.
Scaling up machine learning models boosts jet classification performance
How to scale your HEP ML models: A recipe for robust architecture comparisons at scale
Abstract: Much of the recent progress in machine learning domains such as language models has come from scaling laws that predict performance as a function of training effort. In high-energy physics (HEP) similar behavior has now been observed. To aid further study, we present a systematic procedure to derive robust scaling laws and compare design choices on the relevant budget axes for HEP tasks. We first validate the full scaling trajectory on toy problems and then apply the procedure to multi-task transformers on the ~11 billion-jet ATLAS JetSet2 dataset, in both the compute- and data-constrained regimes. For the latter, we predict, to the best of our knowledge for the first time, the jointly optimal model size, training horizon, learning rate and batch size under early stopping. At compute-optimal scaling, we recover a near-equal $\sqrt{C}$ dependence of model and dataset size, and find that auxiliary objectives lower the primary jet-classification loss at equal compute budget. Expanding the inputs toward lower-level data systematically lowers the loss while leaving the scaling exponent nearly unchanged. The onset of the power-law regime is itself set by scale: below a threshold in dataset size the loss carries little information about high-compute scaling, underscoring the value of large, high-quality full-simulation datasets as a foundation for scaling studies and the development of foundation models in HEP.
QATFactory improves quantized large language model accuracy for production use
QATFactory: A Versatile, Deployment-Aligned Framework for Quantization-aware Training and Distillation of LLMs
Abstract: Large language model (LLM) inference is increasingly moving toward lower precision to realize the throughput of hardware accelerators, but aggressive post-training quantization (PTQ) can degrade model quality. We present QATFactory, an open-source framework for deployment-aligned quantization-aware distillation (QAD) and reinforcement learning (QARL). QATFactory simulates deployment-time quantization while performing matrix multiplications in BF16, allowing models to adapt to quantization noise without requiring training hardware that natively supports the target format; for example, it supports NVFP4 training on H100 GPUs, which lack FP4 Tensor Cores. The framework supports NVFP4, MXFP4, and llama.cpp's Q4_K format; dense and mixture-of-experts models; and both full-parameter and LoRA-based training. It exports checkpoints directly to vLLM and llama.cpp without an additional lossy conversion step or added inference overhead. With QATFactory, we conduct extensive experiments on models ranging from 8B to 230B parameters and evaluate exported checkpoints in production inference engines. Across models and formats, QAD consistently improves deployed-model quality over strong PTQ baselines. On Qwen3.5-9B, QAD achieves average benchmark accuracies of 68.9% under NVFP4 and 66.0% under MXFP4, outperforming the best PTQ results of 65.4% and 56.4%, respectively. Through our experiments, we found that although both FP4 formats quantize weights and activations at deployment, the best training strategy is format-dependent: NVFP4 generally performs better when only weights are quantized during training, whereas MXFP4 benefits from quantizing both weights and activations. At a fixed training token budget, training on fewer 32K sequences improves average accuracy by 1.9 points over training on more 4K sequences. We release the complete QATFactory training code and the resulting checkpoints.
LLM inference system ensures consistent outputs across runs
Vosti: Specifying, Implementing, and Verifying Deterministic LLM Inference
Abstract: LLM inference systems may vary batch composition, prompt chunking, prefill/decode execution, and KV-cache reuse, eviction, or recomputation. These optimizations should not affect system outputs. Production systems, including vLLM's batch-invariant mode and SGLang's deterministic mode, target this goal but lack a formal system-level specification. We formalize deterministic LLM inference: under a fixed model and deployment configuration, requests with the same prompt and initial sampler state produce bitwise-identical logits at corresponding output positions across executions. Our tests find that these production modes produce different logits under some execution variations. To address this limitation, we present Vosti, an inference engine designed and verified against this specification. Vosti chooses kernels independently of runtime engine state and ties cached KV values to their logical token prefixes. Its proof decomposes at the engine/GPU kernel boundary: a Verus inductive proof establishes that scheduling and the paged, prefix-sharing KV-cache preserve output logits, while a Triton analyzer proves bitwise-equal selected kernel outputs across batches, query lengths, and paged KV-cache layouts. Vosti produces bitwise-identical logits across every tested execution variation and achieves performance comparable to vLLM's batch-invariant mode on decode-heavy workloads, while providing a stronger, formally verified determinism guarantee.
Delta-matching enables reliable native 8-bit training for large language models
Delta-Matching: Closing the Final Gap of Native 8-bit Training for LLMs
Abstract: Reliable FP8 attention remains a barrier to fully native 8-bit large language model training. We derive how forward-backward inconsistencies produce stale delta and empirically show how it distorts training dynamics. Our stale-delta hybrid runs show a modest loss gap at 569M parameters but substantial loss increases and downstream degradation at 1.67B and 5.29B. QK normalization, NoPE (no positional encoding), and lower-learning-rate context extension mitigate or delay degradation without eliminating it. This pattern suggests accumulated optimization error that smaller models and short runs can conceal. We propose Delta-Matching, proving that it restores the softmax gradient's zero-row-sum invariant under the stated numerical assumptions. It enables native block-scaled FP8 in every forward and backward attention-core matmul without architectural changes, smaller global batches, or auxiliary forward outputs. Across tested architectures, scales, and training stages, Delta-Matching matches BF16/FP32 mixed-precision training loss and overall downstream performance. We will release our implementation, trained models, and data recipes.
Looped mixture of experts improves model efficiency by reshaping experts
How to Loop MoE: Flatten the Experts, Untie the Attention
Abstract: Looped Transformers reuse one block of layers several times: by spending extra computation they push a model of fixed size further, and so use its parameters more fully; while sparse mixture-of-experts (MoE) models activate only a few of many experts for each token. Looped MoE bridges these two design philosophies and gives MoE models new potential for better expert usage, but it raises a question: how to loop a MoE? We answer it with Foil. With the expert parameters and the expert compute per token held fixed, Foil (1) flattens the experts, halving the expert layers, doubling the experts per layer and doubling the passes, so that every routing decision chooses from a larger pool, and (2) unties the attention, giving each pass its own attention parameters while the experts and routers stay shared. Experiments show that Foil clearly outperforms the unflattened looped baseline: at 20B tokens every Foil model has lower pretraining loss than the baseline; at 100B tokens the loss improves monotonically with the degree of flattening, the most flattened Foil ending 0.012 nat below the baseline at equal parameters and compute, with downstream accuracy on par or better; untying the attention also yields more balanced and more confident routing at equal shape. Our ablations analyse why Foil works and turn the findings into design guidance for looped MoE: the returns of looping and of widening the expert layers amplify each other, routing confidence tracks healthy expert use better than load balance, and a sparse looped MoE should therefore use more experts per layer and more passes. Code and configurations are available at https://github.com/SR-A-W/how-to-loop-moe.
Kv cache compression method retains long and off context query strength
Cartridges++: KV Cache Compression without Off-Context Derailment
Abstract: Serving long documents to a Large Language Model (LLM) repeatedly is expensive: computations grow with context length, and the memory footprint of the key-value (KV) cache balloons. Compressed KV (CKV) representations aim to mimic the cache of a document and are typically computed once and for all, ahead of inference time. Methods to obtain CKVs range from drop mechanisms that reduce their number of columns, to learned approaches. Among the latter, Cartridges have emerged as a leading compression method, learning compact KV representations through distillation on relevant Q/A pairs. While existing evaluations focus primarily on whether Cartridges and other CKVs yield approximately similar responses to document-related, on-context queries, we investigate the crucial deployment question of whether they can handle off-context queries, something the native KV representation is particularly good at, thanks to the mechanics of attention. We observe a fundamental trade-off: while Cartridges perform better for on-context queries, heuristic-variants preserve better the original LLM's ability to operate off-context. We measure this through their capability to avoid context contamination in their response, retain general knowledge, and follow instructions. We propose Cartridges++, simple modifications to cartridges that retain off-context abilities at small or negligible cost. The router variant decides at inference time whether the query should use the learned long-context memory, while the data-mixing variant allocates a small fraction of training Q/As to queries outside the reference long document. Our study shows that assessing CKVs on document utility alone can mask substantial degradation in broader model capabilities, yet those issues can be fixed with benign changes to CKV inference or training.
Efficiently select experts to speed up large AI models
EAT: Expert Account Tracker for Efficient MoE Inference
Abstract: Mixture-of-Experts (MoE) models have emerged as a revolutionary method to scale Transformer models. However, traditional MoE architecture still suffers from inefficiency since a large number of experts are unnecessarily activated. Existing approaches for reducing the number of activated experts often overlook the historical performance of each expert. In this paper, we propose EAT, a novel method called Expert Account Tracker (EAT), which utilizes history-awareness metrics and adaptive thresholding to dynamically select the most important experts, thereby reducing the activated expert number while effectively maintaining the model performance. Experiments show that EAT outperforms the existing baseline Top-P method across multiple models and datasets, achieving over 25% an average reduction compared to the vanilla method in the number of activated experts and performing better token generation speed compared to the baseline. Furthermore, the performance of pruned models can be efficiently recovered via OPD using only 9K data. Additionally, through ablation studies, we find that excessively reducing the number of activated experts can significantly harm model performance, and the importance of experts varies across layers, with higher-level experts being generally more critical.
Oled moe speeds up large language model inference with smarter expert memory use
OLED-MoE: Accelerating MoE-Based dLLM Inference via Inter-Iteration Locality-Aware Expert Offloading
Abstract: Semi-autoregressive diffusion large language models (dLLMs) improve decoding parallelism through iterative block-wise denoising, but scaling them with mixture-of-experts (MoE) layers introduces a large expert parameter footprint that exceeds memory-constrained GPU capacity. Expert offloading is a natural remedy, yet existing MoE serving systems target autoregressive decoding and rely on intra-iteration layer-wise prefetching: while computing one layer, they predict and load experts for subsequent layers. Under dLLM inference, block-wise routing expands the active expert working set within each iteration, making such prefetches difficult to complete in time and costly when mispredicted. Consequently, existing prefetch-based solutions often degenerate into on-demand expert loading with high decoding latency. We propose OLED-MoE, an expert offloading system that shifts the optimization target from intra-iteration prefetching to inter-iteration expert retention. Its key insight is that adjacent denoising iterations exhibit strong expert routing overlap, and token confidence indicates which experts are likely to be reused. OLED-MoE uses confidence-guided inter-iteration prediction to retain high-value experts in GPU memory without introducing extra prefetch traffic. It further compensates unavoidable cache misses through CPU-GPU cooperative execution, jointly considering dynamic expert computation load and predicted future reuse. Across diverse dLLM workloads, OLED-MoE reduces time per output token (TPOT) by 1.23x-7.93x and improves expert cache utilization by 1.44x-4.23x over state-of-the-art offloading systems. Notably, OLED-MoE approaches full-residency performance while using only 40% of the expert GPU memory, incurring merely 23% higher TPOT despite a 60% reduction in expert memory footprint. OLED-MoE's source code is publicly available at https://github.com/flashserve/OLED-MoE.
On policy training improves efficient transformer attention on long tasks
On-Policy Attention Linearization
Abstract: Hybrid transformer architectures that replace most softmax attention layers with linear attention offer transformer-level quality at a fraction of the memory cost. Rather than pretraining such models, a growing body of work distills them from already trained full-attention transformers. However, these distilled models often collapse on long-context retrieval and reasoning tasks, particularly when operating in thinking mode, where the efficiency gains of hybrid architectures matter most. Since linear attention layers must compress context into a fixed-size state, their errors compound over long sequences. As off-policy distillation never teaches the student model to recover from this drift, tasks that necessitate longer sequence lengths become especially challenging. We introduce On-Policy Attention Linearization (OPAL) in which the hybrid attention student samples its own long-context trajectories and receives dense supervision from the frozen full-attention teacher. Applying OPAL to Qwen3-4B and MiMo-7B-RL-0530, we recover $87$--$94\%$ of full-attention performance on commonsense reasoning, $100\%$ on needle-in-a-haystack (NIAH) retrieval, and $83$--$93\%$ on mathematical reasoning with only 3B training tokens. We achieve these results without supervised fine-tuning (SFT) or reinforcement learning with verifiable rewards (RLVR). Compared with the strongest prior linearization method, which recovers $68\%$ of its teacher's retrieval performance and $21.6\%$ absolute average mathematical reasoning accuracy, OPAL fully recovers retrieval and achieves $67.6$--$72.2\%$ on math reasoning.
Hybrid sparse attention improves long-context processing efficiency
HySparse2: Hybrid Sparse Attention with Two-Level KV Sharing
Abstract: Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context demands efficient prefill, compact KV-cache storage, and accurate long-context retrieval. To meet these demands, we introduce HySparse2, a hybrid sparse attention architecture with two-level KV sharing. At the outer level, KV Bridging adopts a YOCO-style self-decoder and cross-decoder structure, but bridges only full-attention layers. The self-decoder uses hybrid sliding-window attention (SWA), while the cross-decoder uses hybrid sparse attention. The KV caches for full-attention layers in the cross-decoder are generated from the hidden states of full-attention layers in the self-decoder. At the inner level, HySparse2 retains HySparse's core KV Reuse design with two refinements. First, it replaces block-level sparsity with token-level sparsity for finer long-context retrieval. Second, it removes the separate SWA branch from sparse layers and instead forces a sliding window of recent tokens into the sparse selection. This two-level KV sharing allows all cross-decoder KV caches to be constructed from self-decoder hidden states. Prefill can therefore exit after the self-decoder, skipping all cross-decoder layers. On an 80B-A3B MoE model, HySparse2 outperforms HySparse and Hybrid SWA on long-context retrieval and multi-turn agentic tasks, while substantially reducing prefill computation and KV-cache storage.