arXiv subjects

Kan Liu

Publications and source records attributed to Kan Liu.

At least 19 recordsLinked to original sources

arXiv2026

GC-MoE: Overlapping Expert-Parallel Token Dispatch with FP8/FP4 Grouped GEMM via CUDA Green Contexts

Under expert parallelism, mixture-of-experts (MoE) layers pay all-to-all token dispatch, grouped GEMMs, and a combine step in every layer. In decode each rank holds only 1-256 tokens: dispatch latency and weight traffic dominate while tensor-core utilization is inherently low. In fused operators such as DeepGEMM's Mega MoE, dispatch and computation overlap only by sharing every SM: dispatch warps compete with TMA/MMA and epilogue warps for registers, shared memory, and issue slots; every phase shares one CTA resource budget, synchronization spans the whole SM array, and every replay ends with a workspace-clean pass and a cross-rank barrier. We present GC-MoE, an open-source FP8/FP4 MoE inference operator for NVIDIA Hopper and Blackwell GPUs that uses CUDA Green Contexts to split the device into a small communication partition and a large compute partition. Each role gets its own kernel, resource budget, and synchronization domain, so dispatch and GEMM tune independently. A three-node CUDA Graph (four on Hopper with fused shared experts) launches the dispatch and grouped-GEMM kernels as concurrent roots; a token-ring handshake starts computation on the first filled ring blocks, and double-buffered receive counters remove the trailing cross-rank barrier between replays. The compute partition runs operand-swapped tcgen05 block-scaled MMA on Blackwell and register-A WGMMA with in-register FP4 decoding on Hopper, with fused SwiGLU, requantization, and optional shared experts. On 8 H20 GPUs GC-MoE accelerates DeepSeek-V4.1-Flash decode by 1.188x on average on both the routed and shared-expert suites (peaks 1.296x and 1.353x); on 8 B300 GPUs it accelerates GLM-5.2 and DeepSeek-V4.1-Flash by 1.106x (1.123x with fused shared experts, peak 1.209x), beating Mega MoE at all 54 measured points at matched accuracy. GC-MoE reserves only 20 of 148 SMs on B300 for dispatch.

cs.DC↗

arXiv2026

When Shards Fly: Cross-Tensor In-Flight Conversion for Low-Latency, Low-Memory Weight Updates in Disaggregated RL

Disaggregated LLM reinforcement learning (RL) repeatedly synchronizes policy weights between training and rollout GPU pools. Existing systems reduce communication through shard delivery and replica forwarding, yet separate conversion passes and staging buffers still add latency and GPU-memory overheads to synchronous updates. We present ShardFly, which coordinates reusable transfer planning with streaming GPU execution for synchronous, in-place weight updates. ShardFly introduces two components. The Planner compiles native shard mappings, quantization dependencies, and replica routes into cross-tensor schedules, reserving payload buffers only for scheduled transfers. The Executor runs each rank's schedule in one persistent transfer kernel, fusing memory-layout conversion, quantization, GPU-initiated communication, and destination writes while forwarding transformed bytes across replicas. Plans and communication state are reused under stable configurations and replaced between updates for coordinated rollout scaling. On up to 32 H20 GPUs, ShardFly achieves complete BF16 weight-update speedups of 1.26--27.92$\times$ across five models and four baselines: veRL NCCL, Awex, veRL NIXL, and WeightBridge. In end-to-end Qwen3-30B-A3B GRPO across four rollout configurations, ShardFly improves response-token throughput by 5.49--6.71\% over veRL NCCL.

cs.DC↗

arXiv2026

RTP-LLM: Building an Evolvable and Reliable LLM Inference Engine at Alibaba Scale

Large language models and inference techniques evolve rapidly, making model enablement a recurring production task rather than a one-time integration effort. At Alibaba, a shared inference infrastructure must keep pace with new models and serving capabilities while remaining reliable across heterogeneous production environments. To reconcile rapid evolution with production reliability, we present RTP-LLM, an open-source production inference engine developed and deployed at Alibaba. Its architecture separates model-specific execution from shared request processing, state management, and resource coordination. Explicit interfaces localize compatible model changes while allowing models and deployment configurations to reuse common serving mechanisms. RTP-LLM has evolved from an exploratory system into a shared production engine that supports $\sim$100 models and processes hundreds of trillions of tokens per day. We describe how its architecture and production practices accommodate model evolution and contain the impact of failures, and share lessons for building evolvable and reliable inference engines at enterprise scale.

cs.OS↗

arXiv2026

SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation

Training high-quality diffusion models is computationally expensive, and inference-time guidance often adds further sampling cost. In this work, we introduce Swift, a simple yet effective self-guided training method that uses a frozen weak denoiser to improve a stronger model without adding inference overhead. For each noisy input, Swift extrapolates the clean regression target away from the weak denoiser's prediction, amplifying components that remain underfit. The weak guide can be an early training checkpoint, requires no backpropagation during main-model training, and is discarded afterward. The resulting model retains the original sampler and supports both conditional and unconditional generation without label dropout or an unconditional branch. We explain this mechanism through a spectral analysis of finite-time population gradient flow in an RKHS. With an early-stopped weak guide, we prove that a nonempty range of extrapolation strengths yields strictly lower posterior-mean error than matched standard training, despite shifting the regression target away from the posterior mean. Experiments on CIFAR-10 and ImageNet-256 demonstrate improvements across model scales and representations. In one ImageNet-256 configuration, Swift surpasses the baseline's best FID using roughly half the total training compute, including guide preparation and training overhead. On RAEv2, it achieves FID 1.15 without inference-time guidance and 0.99 with guidance. These results establish target extrapolation as an effective approach to weak-to-strong diffusion training with efficient single-model inference.

cs.LG↗

arXiv2026

Rethinking Cross-Layer Information Routing in Diffusion Transformers

Diffusion Transformers (DiTs) have become a de facto backbone of modern visual generation, and nearly every major axis of their design -- tokenization, attention, conditioning, objectives, and latent autoencoders -- has been extensively revisited. The residual stream that governs how information accumulates across layers, however, has been directly inherited from the original Transformer. In this paper, we present a systematic empirical analysis of cross-layer information flow in DiTs, jointly along depth and denoising timestep, and identify three concrete symptoms of traditional residual addition, namely monotonic forward magnitude inflation, sharp backward gradient decay, and pronounced block-wise redundancy. Motivated by this diagnosis, we propose Diffusion-Adaptive Routing (DAR), a drop-in residual replacement that performs learnable, timestep-adaptive, and non-incremental aggregation over the history of sublayer outputs. Moreover, the proposed DAR is compatible with many modern Transformer enhancement methods, such as REPA. On ImageNet $256\times256$, DAR improves SiT-XL/2 by $2.11$ FID ($7.56$ vs. $9.67$) and matches the baseline's converged quality with $8.75\times$ fewer training iterations. Stacked on top of REPA, it yields a $2\times$ training acceleration in the early stage, suggesting cross-layer information routing as an underexplored design axis in diffusion modeling, one that operates orthogonally to existing representation-alignment objectives. Beyond pretraining, DAR can also be applied during the fine-tuning stage of large-scale T2I models and preserves high-frequency details during Distribution Matching Distillation. Code is available at https://github.com/cg233666/attention-residuals-on-dit.

cs.CV↗

arXiv2026

Adversarial Gradient Forcing for Robust Few-Step Autoregressive Video Generation

Autoregressive video generation models generate videos sequentially, with each new block conditioned on previously generated blocks. By training on the model's own rollouts, Self Forcing and its follow-up methods substantially mitigate exposure bias caused by the mismatch between training- and inference-time history distributions. However, these models still suffer from error accumulation during long-horizon generation. We identify a key underlying limitation: the model remains highly sensitive to errors in its generated history, allowing small deviations to be amplified through subsequent generations. We refer to this underexplored problem as the local history robustness gap. To address this gap, we introduce Adversarial Gradient Forcing (AGF), a framework designed to improve the robustness of few-step autoregressive video models to perturbations and accumulated errors in their own generated histories. AGF uses distribution matching gradients from future block predictions to select adversarial perturbations of self-generated histories. It then jointly trains the generator on the original and perturbed histories to improve the robustness of future generation. Extensive experiments show that AGF achieves the highest VBench Total score on 60-second video generation and the best visual quality on four-minute video generation. Notably, AGF achieves such long-horizon robustness while requiring only 5-second rollouts for training.

cs.CV↗

arXiv2026

Super-Resolution in The Right Latent Space: A Frozen Vision-Foundation Substrate

In an image latent space, the embeddings of high-resolution, natural, and sharp images form a manifold. Degradation of high-resolution images pushes their embeddings off this manifold. Real-world super-resolution (SR) then becomes the task of mapping the degraded embedding back onto this manifold --- not anywhere on the manifold, but to the point that preserves what the input still carries, both its semantics and pixel details. Every published method implements this mapping in a reconstruction-oriented latent space or pixel space. We claim these spaces are the wrong substrates for SR. Low-resolution and degraded images are embedded far from the manifold, making the mapping difficult and expensive. The lack of semantic information in these substrates also makes it difficult to navigate to the faithful point on the manifold, causing severe hallucination when degradation is heavy. Thus, restoring in a suitable latent space is crucial to the SR task. We show that the latent space of 23 fused layers of a frozen DINOv3-L is one such space that makes the SR task easier. Degraded images are embedded near the manifold. Moreover, this substrate contains a hierarchy of information, from pixel record to degradation robust semantics, guiding the model to find the faithful point on the manifold. On this substrate, a 415M decoder is trained under reconstruction and adversarial objectives to map the degraded embeddings back and decode to pixel space in one pass. The resulting model, RAESR, attains the best fidelity--perception trade-off among state-of-the-art adversarial and diffusion-based restorers on RealSR, DRealSR, LSDIR and DIV2K-Val, at 37 ms per 512 by 512 image on a single H20 GPU. Swapping the substrate for a VAE latent under an identical recipe loses on every metric.

cs.CV↗

arXiv2026

Adaptive Context Parallelism for Production LLM Serving

As LLM context windows expand and input sequences grow longer, serving systems face increasing computational and memory demands. Context parallelism (CP), which partitions the input sequence across multiple ranks to parallelize the computation, has therefore become increasingly important for efficient LLM serving. However, existing CP-enabled systems either rely on static CP configurations or adjust the CP degree only for active requests or batches. In this paper, we present Vertumnus, an adaptive CP serving system designed for heterogeneous and evolving workloads. At the request level, Vertumnus routes requests among workers with different CP degrees using a placement cost that combines predicted queuing delay, cache-aware prefill time, and GPU-time cost. At the cluster level, Vertumnus adapts the worker composition through seconds-scale split and merge operations as workload demand changes. Vertumnus further introduces a global prefix-cache management policy that coordinates cache placement and replication among workers with the same or different CP degrees, preserving cache locality as request assignments and worker composition change. Experiments on a 64-GPU cluster with public and production workloads show that, under the highest evaluated loads, Vertumnus reduces mean TTFT by up to 28.1% and improves token-weighted SLO attainment by up to 13.3 percentage points over the strongest baseline.

cs.OS↗

arXiv2026

RT-Lynx: Putting GEMM Sparsity in the Right Place for Diffusion Models

Diffusion Transformers (DiT) achieve strong performance in image generation but incur substantial inference costs. While prior work has reduced this cost via quantization and distillation, semi-structured sparsity, which can nearly halve FLOPs, remains underexplored. A key reason is that most existing approaches focus on weight sparsification, and pruning 50% of the weights can remove critical model capacity and degrade generation quality. Our study, however, shows that DiT activations are intrinsically sparse and significantly more robust to N:M semi-structured sparsification than weights. Motivated by this observation, we advocate a paradigm shift from weight sparsification to activation sparsification. We propose RT-Lynx, which applies N:M sparsification to activations and incorporates error-compensation techniques to mitigate accuracy loss. We further implement highly optimized CUDA kernels tailored to this setting, achieving up to a 1.55x speedup on average in linear layers. Extensive experiments across multiple diffusion models demonstrate that our method preserves the generation quality of the original models while substantially accelerating inference.

cs.LG↗

arXiv2026

SCOPE: Subspace Clustering with Online Per-Head Top-K Estimation for Sparse Video Attention

Diffusion Transformers (DiTs) incur quadratic self-attention cost over spatiotemporal tokens. Existing training-free sparse attention methods often construct sparse masks from block-level or cluster-level proxy scores, which can obscure fine-grained differences among keys and miss high contribution keys under aggressive sparsity. Moreover, such proxy scores may yield overly concentrated softmax distributions, causing Top-$p$ to retain too few keys for some query clusters. Although a fixed Top-$k$ minimum alleviates this failure mode, a shared value cannot adapt to variations across heads and inputs. To address both limitations, we propose SCOPE, a training-free sparse attention framework that combines 3D-RoPE-aligned key subspace clustering with online per-head Top-$k$ estimation for efficient video-DiT inference. SCOPE partitions post-RoPE keys into temporal, height, and width subspaces, clusters them independently, and aggregates the corresponding centroid scores through lookup tables to obtain per key proxy scores for each query cluster. Building on existing hybrid Top-$p$/fixed Top-$k$ selection, SCOPE derives a head-specific Top-$k$ value online by averaging the initial retained key counts within each head, weighted by query cluster size, and selects additional keys only for query clusters whose initial retained key counts fall below this value. Sparse attention is then computed over the selected original keys and values. Across six model--task configurations, SCOPE consistently outperforms existing training-free baselines in both fidelity and latency, achieving up to a $1.99\times$ end-to-end speedup on 720p HunyuanVideo with $28.46$ dB PSNR relative to dense attention.

cs.CV↗

arXiv2026

Text Template Tokens Are Implicit Semantic Registers in Diffusion Transformers

Modern text-to-image diffusion transformers (DiTs) generate images through joint attention, in which text and image tokens interact directly within a single sequence. In large-scale DiTs, the conditioning input contains not only the user prompt but also chat-template tokens introduced by LLM-based text encoders. Yet how these tokens participate in the denoising computation remains poorly understood. To probe this, we introduce a causal interpretability framework. Using it to separate prompt-content tokens from chat-template tokens, we find that the template tokens carry little prompt-specific information at the encoder output. Yet surprisingly, they emerge as dominant image-to-text attention sinks and causally maintain object identity inside the DiT, acting as implicit semantic registers. We show that they acquire this identity indirectly. Rather than reading the prompt tokens, they draw the identity from the image latents into which the prompt semantics have already been injected at the very first layer. We further reveal a division of labor across heads and depth in DiTs, where distinct heads route semantics or render visual structure, and identity is committed in early blocks, carried by middle blocks, and refined in late ones. As a practical payoff, this analysis yields a training-free pruning rule that removes the causally inert prompt-reading heads and cuts $20\%$ of joint-attention FLOPs at a $1.4$-point cost in GenEval accuracy. Overall, our work not only reveals that the tokens encoding semantics at the input need not be those that maintain them during generation, but also provides a causal view of internal mechanisms in diffusion transformers.

cs.CV↗

arXiv2026

C$^2$KV: Compressed and Composable KV Cache Reuse for Efficient LLM Inference

Long-context inference is central to modern large language model (LLM) applications such as retrieval-augmented generation and multi-document reasoning. To mitigate the growing inference cost, recent work has explored key-value (KV) cache reuse to reduce redundant prefill computation. However, existing reuse methods primarily focus on computation savings and overlook a critical bottleneck in long-context LLM serving: the cost of storing and accessing large KV caches. While KV compression appears to be a natural complement, naively combining compression with non-prefix KV reuse often leads to severe accuracy degradation. In this work, we propose C$^2$KV, a unified framework for non-prefix KV reuse that jointly optimizes KV extraction and inference-time concatenation. C$^2$KV learns a composable and compressed KV cache manifold that is explicitly designed to be position-agnostic. Our approach introduces a lightweight sidecar Extractor with learnable compression tokens and a structured attention flow, enabling modular KV representations that can be flexibly reused and concatenated without modifying the frozen base model. We further employ a compression-concatenation co-training strategy to align extraction-time representations with their downstream reuse behavior. Extensive experiments across multiple long-context benchmarks and model families demonstrate that C$^2$KV significantly reduces KV cache storage and transfer costs, achieving up to 17$\times$ inference speedup under long contexts, while preserving generation quality.

cs.CL↗

arXiv2026

Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps

Long-context inference in large language models is bottlenecked by the quadratic cost of full attention. Existing efficient alternatives often rely either on native sparse training or on heuristic token eviction, creating an undesirable trade-off among efficiency, training cost, and accuracy. In this work, we show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation. Our approach is built on three observations: (1) only a small subset of attention heads truly requires full long-context processing; (2) long-range retrieval is governed primarily by a low-dimensional subspace, allowing relevant tokens to be retrieved efficiently with a 16-dimensional indexer; and (3) the useful token budget is strongly query-dependent, making dynamic top-$p$ selection more suitable than fixed top-$k$ sparsification. Based on these insights, we propose RTPurbo, which retains the full KV cache only for retrieval heads and introduces a lightweight token indexer for sparse attention. By exploiting the model's intrinsic sparsity, RTPurbo achieves sparsification with only a few hundred training steps. Experiments on long-context benchmarks and reasoning tasks show that RTPurbo preserves near-lossless accuracy while delivering substantial efficiency gains, including up to a 9.36$\times$ prefill speedup at 1M context and about a 2.01$\times$ decode speedup. These results suggest that strong sparse inference can be obtained from standard full-attention training without expensive native sparse pretraining.

cs.CL↗

arXiv2026

RTP-LLM: High-Performance Alibaba LLM Inference Engine

Large Language Models (LLMs) have revolutionized AI applications, but deploying them at scale presents significant challenges. We present RTP-LLM, a high-performance inference engine for industrial-scale LLM deployment, successfully deployed across Alibaba Group serving over 100 million users. RTP-LLM addresses fundamental bottlenecks through integrated design. It optimizes model loading via file-order-driven I/O and parallel I/O-communication overlapping. The Prefill-Decode Disaggregation architecture decouples compute-intensive prefill from memory-bound decode phases, combined with hierarchical multi-tiered KV cache management enabling efficient cache reuse. In addition, RTP-LLM incorporates modular speculative decoding supporting multiple algorithms, adaptive KV cache quantization, and decoupled multimodal processing, with support for multi-level parallelism. Comprehensive evaluations across diverse model architectures (8B-235B parameters) have been conducted, where both controlled benchmarks and real production workloads are used. The results demonstrate RTP-LLM's superior performance against vLLM and SGLang: 4.7x-6.3x model loading speedup, 35-37% TTFT P95 latency reduction with 215% cache reuse improvement in production traffic scheduling, 1.12x-2.48x and 1.86x-2.52x throughput improvements in speculative decoding and multimodal inference, respectively, and 35-40% batch latency reduction with 1.9x-3.0x TTFT improvement in quantized inference. RTP-LLM's production-proven architecture and open-source availability make it a comprehensive solution for industrial LLM deployment.

cs.OS↗

arXiv2026

Linear-DPO: Linear Direct Preference Optimization for Diffusion and Flow-Matching Generative Models

Direct Preference Optimization (DPO) is successful for alignment in LLMs but still faces challenges in text-to-image generation. Existing studies are confined to denoising diffusion models while overlooking flow-matching, and suffer from an objective mismatch when applying discrete NLP-based DPO to regression-based generative tasks.\ In this paper, we derive a generalized DPO objective that covers both diffusion and flow-matching via a unified reverse-time SDE framework, and point out from a gradient perspective that the standard DPO objective is suboptimal for text-to-image generation. Consequently, we propose Linear-DPO, which replaces the aggressive sigmoid-based utility function with a sustained linear utility and incorporates an EMA-updated reference model. Qualitative and quantitative experiments on diffusion models (SD1.5, SDXL) and flow-matching model (SD3-Medium) demonstrate the superiority of our approach over existing baselines.

cs.CV↗

arXiv2026

QCP: A Practical Separation Logic-based C Program Verification Tool

As software systems increase in size and complexity dramatically, ensuring their correctness, security, and reliability becomes an increasingly formidable challenge. Despite significant advancements in verification techniques and tools, their practical application to complex, real-world systems is often hindered by critical gaps in both automation and expressiveness. To address these difficulties, this paper presents \textbf{Qualified C Programming Verifier (QCP)}, a novel verification tool that integrates annotation-based automatic verification with interactive proving using Rocq. QCP employs symbolic execution and a separation logic entailment solver to automatically discharge many verification obligations, while deferring more complex obligations to Rocq for manual proof. Furthermore, QCP includes a VS Code extension designed to enhance proof efficiency and support a deeper understanding of both the program behavior and verification outcomes.

cs.PL↗

arXiv2026

LegoDiffusion: Micro-Serving Text-to-Image Diffusion Workflows

Text-to-image generation executes a diffusion workflow comprising multiple models centered on a base diffusion model. Existing serving systems treat each workflow as an opaque monolith, provisioning, placing, and scaling all constituent models together, which obscures internal dataflow, prevents model sharing, and enforces coarse-grained resource management. In this paper, we make a case for micro-serving diffusion workflows with LegoDiffusion, a system that decomposes a workflow into loosely coupled model-execution nodes that can be independently managed and scheduled. By explicitly managing individual model inference, LegoDiffusion unlocks cluster-scale optimizations, including per-model scaling, model sharing, and adaptive model parallelism. Collectively, LegoDiffusion outperforms existing diffusion workflow serving systems, sustaining up to 3x higher request rates and tolerating up to 8x higher burst traffic.

cs.DC↗

arXiv2026

Rethinking Retrieval-Augmentation as Synthesis: A Query-Aware Context Merging Approach

Retrieval-Augmented Generation (RAG) enables Large Language Models (LLMs) to extend their existing knowledge by dynamically incorporating external information. However, practical deployment is fundamentally constrained by the LLM's finite context window, forcing a trade-off between information sufficiency and token consumption. Standard pipelines address this via a retrieve-then-select strategy, typically retaining only the top-k chunks based on relevance. Nevertheless, this approach is suboptimal: it inherently truncates critical bridging evidence located in the long tail of the relevance distribution, while simultaneously wasting the token budget on semantically redundant high-ranking chunks. In this paper, we rethink retrieval-augmentation as a dynamic optimization problem aimed at maximizing information density. We propose MergeRAG, a novel framework that shifts the paradigm from static filtering to query-aware synthesis. MergeRAG employs a scoring agent to restructure retrieved contexts through a dual-pathway mechanism: 1) Symmetric Merging, which consolidates weak signals to recover lost bridging evidence; 2) Asymmetric Merging, which utilizes entropy-guided anchoring to eliminate redundancy without sacrificing semantic integrity. We further introduce a Hierarchical Parallel Merging strategy that mitigates information loss while maximizing computational parallelism. Extensive experiments on standard benchmarks demonstrate that MergeRAG significantly outperforms state-of-the-art RAG baselines, achieving up to 13.7 points improvement in F1 score and 11.5 points in Exact Match (EM), respectively.

cs.IR↗