Pith. sign in

REVIEW 3 major objections 5 minor 54 references

Xema: Efficient Diffusion Serving through Fine-Grained Memory Management and Auto-Configuration

T0 review · 3 major / 5 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Diffusion serving can fit high-resolution images and long videos on limited GPUs by cutting memory only in short predicted pressure windows and auto-picking the plan that maximizes goodput.

desk verdict Solid systems paper: interval-local mitigation plus lifetime layout and joint planning deliver real SLO and planning-cost gains on production diffusion pipelines; main risks are scope and the template-static hinge, not internal inconsistency. read the letter →

arxiv 2607.11136 v1 pith:FVFHMEXP submitted 2026-07-13 cs.DC

Xueze Kang , Guangyu Xiang , Suyi Li , Yuxin Wang , Shaohuai Shi , Lin Zhang , Xiaowen Chu This is my paper · ORCID
classification cs.DC
keywords diffusionservingGPUmemorymanagementtrace-guidedmitigationstaticlayoutauto-configurationgoodputsequenceparallelismSLOattainment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Production image and video diffusion often fails on common GPUs because peak memory from weights, DiT activations, VAE activations, and allocator fragmentation exceeds capacity even at batch size one. Coarse global tricks such as full weight offloading, sharding, or VAE slicing fix the fit but add large latency that wrecks service-level objectives. Xema starts from a simple fact: for a fixed request template the operator sequence and tensor lifetimes are known offline, and near-peak memory appears in only a few short intervals. It builds an offline memory trace, mitigates only inside those intervals and only by the amount needed to stay under budget, and lays out predictable tensors statically so reserved memory tracks live tensors. An offline planner then jointly chooses parallelism, batch size, and these fine-grained controls, stores the best plan, and the online runtime just looks it up. On Flux.2, CogVideoX-5B, and LTX-2 the result is up to 3.7× higher SLO attainment and planning that finishes in minutes instead of hours.

What carries the argument

Trace-guided fine-grained memory mitigation plus a static lifetime-based layout: for each request template an offline shape-only run produces tensor lifetimes and a reserved-memory trace; short intervals that exceed the budget receive the minimum intensity of offloading, HSDP, chunking, fusion, or temporary serialization; non-overlapping tensors share address ranges so reserved memory closely matches active memory and runtime matches the offline plan.

What would settle it

Run the same request templates with content that triggers data-dependent control flow, dynamic shapes, or large unplanned runtime buffers; if reserved peaks then diverge from the offline plan or the static allocator OOMs while active memory still fits, the central predictability claim fails.

Watch

Extended reading notes

Core claim

Because diffusion inference for a fixed request template is template-static and its memory pressure is localized, a system can derive an offline memory trace, apply mitigation only where and by the amount needed to fit a GPU budget, pin predictable tensors into a static layout that removes fragmentation, and then automatically select the highest-goodput combination of parallelism, concurrency, and memory control under memory and SLO constraints—yielding large gains in feasible request range and SLO attainment over coarse global memory controls and grid search.

Load-bearing premise

For a fixed model, resolution, frame count, and denoising-step count, the full operator sequence, tensor shapes, and allocation and free events are fixed by architecture and do not depend on the actual prompt content, so a shape-only offline run predicts the real reserved-memory peaks.

Editorial extensions

If this is right

  • High-resolution image and long-video templates that currently OOM under global offloading or VAE slicing become serviceable on the same GPUs.
  • Serving systems can replace multi-hour grid search with a plan table filled in minutes per template and hardware budget.
  • Goodput under tight SLOs rises because mitigation cost is paid only in the few percent of intervals that actually threaten the budget.
  • Allocator fragmentation ceases to be a silent source of OOM: reserved memory tracks active memory closely enough that offline reasoning is trustworthy at runtime.
  • The same planner can retarget different GPU counts and memory budgets without manual retuning of SP, DP, batch size, and offload depth.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same lifetime-trace idea could extend to other template-static generative pipelines (e.g., multi-stage image editing or 3D diffusion) where peaks are brief and shapes are known offline.
  • If future models introduce heavy data-dependent branching, hybrid plans that re-verify only the dynamic regions while keeping static layout for the rest may still preserve most of the gain.
  • Operators may start publishing request-template catalogs with precomputed plans rather than only model weights, turning memory fit into a first-class deployment artifact.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. Xema is a diffusion serving system that treats request templates (model, resolution, frame count) as template-static so that offline shape-only runs can produce memory traces and tensor lifetimes. It applies memory mitigation only inside short pressure intervals and only at the intensity needed to fit a GPU budget, builds a static 2D best-fit layout for graph-visible tensors to cut fragmentation, and uses an offline planner that jointly chooses parallelism, batch size, and fine-grained memory controls under memory and SLO constraints, storing the result in a plan table for online lookup. Implemented on Diffusers/xDiT and evaluated on Flux.2, CogVideoX-5B, and LTX-2 against vLLM-Omni, globally extended primitives, and a greedy heuristic, the paper reports up to 3.7× higher SLO attainment and planning cost reduced from 6.3 hours to 197 seconds versus grid search.

Significance. If the results hold, the paper is a solid systems contribution for production diffusion serving: it shows that coarse global offloading/slicing is often the wrong granularity, that localized mitigation plus static layout can make extreme image/video templates feasible on 48 GB GPUs, and that joint planning of SP/DP, batch size, and interval-local controls measurably improves goodput. Strengths include multi-model end-to-end SLO curves (rate, GPU count, SLO scale, skewed mix), extreme single-GPU feasibility, allocator vs. expandable_segments comparison, planning-cost breakdown, and configuration ablations. The work is empirical rather than formal, but the evaluation design is appropriate for the claim and the planning-cost reduction is concrete.

major comments (3)
  1. The central claim rests on the template-static contract (§3.3, §6.1): for fixed (model, H, W, T) and denoising steps, a shape-only run yields operator sequence, tensor shapes, and allocation/free events that match runtime for managed tensors, with dynamic memory covered by a fixed ~5% margin. The paper asserts data-independent control flow but does not stress-test content-dependent paths, adaptive step counts, or runtime buffers that could break offline traces and the static allocator. A short sensitivity study or explicit failure modes would make the load-bearing assumption falsifiable rather than assumed.
  2. §9.1 defines the base SLO as 2.5× the latency of the best feasible strategy in the shared candidate space. Because Xema expands that space with fine-grained controls, the relative SLO can favor configurations that only Xema can reach and makes the 3.7× attainment figure harder to interpret against systems that cannot serve the same templates. Reporting absolute latency/goodput or an external fixed-latency SLO would strengthen the comparison.
  3. Evaluation is limited to a single 8×A6000 PCIe node with synthetic Poisson mixes (§9.1). Multi-node interconnects, production traffic, and other GPU generations are not shown; the planner’s communication/latency tradeoffs (SP, HSDP) may change under different topologies. This does not invalidate the single-node results but bounds how far the goodput claims can be generalized without additional evidence.
minor comments (5)
  1. Figure 1 caption and bars are useful but the exact configuration strings (e.g., “offload 3 blocks + fusion”) are dense; a short legend or table of the winning plans would help.
  2. Algorithm 1 is clear; stating the complexity or typical |T| for Flux.2/CogVideoX would help readers judge offline layout cost.
  3. §5.1 objective minimizes sum of latency overheads under per-segment memory constraints; clarifying whether L_i is measured end-to-end or per-segment and how overlap is credited would reduce ambiguity.
  4. Related work (§10) is appropriate; a one-sentence contrast with training-oriented rematerialization/swapping systems on why diffusion’s short pressure intervals change the design would sharpen the positioning.
  5. Minor polish: “V AE” spacing in §2.2; ensure consistent use of “goodput” vs. “SLO attainment” when citing the 3.7× claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical systems evaluation; SLO/goodput and planning-time claims are measured, not derived from quantities defined as the targets.

full rationale

Xema’s load-bearing chain is engineering plus measurement, not a first-principles derivation that collapses into its inputs. Offline memory traces come from shape-only pipeline runs under the stated template-static premise (§3.3, §6.1); the planner then selects configurations by constrained goodput maximization using those traces and profiled latencies (§7.1–7.3), and the headline 3.7× SLO attainment and 6.3 h→197 s planning numbers are end-to-end experimental comparisons against baselines (abstract, §9.2, Table 1). Goodput is the standard throughput-of-SLO-meeting-requests metric (citing DistServe), not a quantity fitted from the same data being “predicted.” Static layout (Algorithm 1) is best-fit packing of observed lifetimes, not a tautological redefinition of reserved memory. Related-work self-citations (e.g., prior serving systems by overlapping authors) are background and not used as uniqueness theorems or load-bearing premises that force the result. The template-static assumption is an empirical hinge that could fail, but failure would falsify the system rather than make the claims true by construction. No self-definitional loop, fitted-input-as-prediction, or ansatz-via-self-citation reduction is present.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

Load-bearing content is systems assumptions about diffusion execution predictability and a few operational constants (memory margin, SLO multiplier, skew parameter), not free physical constants. Invented items are engineering constructs (plan table, static layout, mitigation library interface), not new physical entities.

free parameters (3)
  • Dynamic-region memory margin = 5%
    Xema reserves an additional 5% beyond the static reserved peak for CUDA contexts, NCCL, and other non-graph memory (§6.2); chosen operationally, not derived.
  • Base SLO multiplier = 2.5× best feasible latency
    Base SLO per template is set to 2.5× the latency of the best feasible strategy in the shared candidate space, following prior work (§9.1); this scales all attainment curves.
  • Skew mix concentration α = 1.0
    Skewed workload samples templates with p_i ∝ exp(α L_i / L_max) and sets α=1.0 (§9.2); affects stress-test results.
assumptions (5)
  • domain assumption Template-static diffusion execution: for fixed (model, H, W, T) and denoising steps, operator sequence, tensor shapes, and alloc/free events are content-independent and offline-derivable.
    Stated in §3.3 and used for memory analyzer, static layout, and planner pruning; if false, offline traces and plans diverge from runtime.
  • domain assumption Memory pressure is highly localized: only a small fraction of allocation events approach phase peak, so global mitigation is unnecessary.
    Supported by CogVideoX CDFs in Fig. 2 and percentages in §5.2; motivates interval-limited control.
  • domain assumption Dynamic/runtime-invisible memory is small relative to static tensors in diffusion serving (hundreds of MB), so a fixed margin suffices.
    §6.1–6.2; underpins reliability of offline reserved-memory reasoning.
  • ad hoc to paper Larger batch sizes under fixed (SP, DP) increase both memory pressure and latency, so pruning a batch size allows stopping the batch search for that layout.
    Stage-1 pruning rule in §7.3; usually true but not formally proved for all operators.
  • standard math Standard GPU allocator and diffusion pipeline semantics (PyTorch caching allocator, DiT/VAE structure, SP/DP/HSDP costs).
    Background systems assumptions throughout §§2–7.
invented entities (2)
  • Xema plan table (per-template execution plan)
    purpose: Store offline-selected parallelism, batch size, mitigation actions, and tensor layout for online lookup.
    Engineering artifact of the planner; not an independent scientific entity.
  • Offline 2D best-fit static memory layout (Algorithm 1) independent evidence
    purpose: Assign address offsets from tensor lifetimes to minimize reserved peak and match offline traces at runtime.
    New allocator policy for this system; evaluated empirically vs PyTorch/expandable_segments, not a postulated physical object.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Xema: Efficient Diffusion Serving through Fine-Grained Memory Management and Auto-Configuration." pith.science (2026). https://pith.science/paper/FVFHMEXP

@misc{pith2026260711136,
  author       = {Pith},
  title        = {Pith review of: Xema: Efficient Diffusion Serving through Fine-Grained Memory Management and Auto-Configuration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FVFHMEXP}},
  note         = {Machine review of arXiv:2607.11136}
}
read the original abstract

Diffusion models are increasingly deployed as production visual-generation services, where serving high-resolution image and long video generation is often limited by GPU memory. Popular memory-saving techniques such as weight offloading, sharding, and VAE slicing are often not practical because they tend to introduce significant performance overhead. In this paper, we present Xema, a memory-efficient diffusion serving system that exploits predictable tensor lifetimes for trace-guided memory optimization. For each request template, Xema derives an offline memory trace to identify short memory-pressure intervals and applies memory mitigation only within these intervals and only by the amount needed to fit the target GPU budget. Xema further constructs a static memory layout for tensors with predictable lifetimes, reducing fragmentation-induced reserved memory and making offline memory reasoning reliable at runtime. Built on this memory optimization layer, Xema introduces an offline planner that jointly selects parallelism, concurrency, and memory control under GPU memory and SLO constraints. The selected plan is stored in a plan table and directly used by the online serving runtime. We implement Xema on production diffusion pipelines and evaluate it with Flux.2, CogVideoX-5B, and LTX-2. Compared with existing serving configurations, Xema improves SLO attainment by up to 3.7x and reduces planning cost from 6.3 hours to 197 seconds compared with grid search.

Figures

Figures reproduced from arXiv: 2607.11136 by the authors.

Figure 1
Figure 1. Optimal execution configurations vary across LTX￾2 request templates. For each template, the first bar is selected by Xema, and the next two bars are the best and second-best configurations searched by vLLM-Omni. By expanding the search space with fine-grained memory controls, Xema finds higher-goodput configurations than vLLM-Omni. 1 Introduction Diffusion models are now widely used in visual generation, including … view at source ↗
Figure 3
Figure 3. System architecture. a fixed rule or manually tuned configuration cannot consis￾tently achieve high goodput across templates. It motivates an offline planner that selects, for each request template, the execution configuration that satisfies the memory budget and SLO while maximizing goodput. 4 System Overview Xema is a diffusion serving system that exploits template￾static tensor lifetimes to support low-overhead m… view at source ↗
Figure 5
Figure 5. Selective mitigation increases the overlap window for weight transfers. mitigation as needed by selecting 𝜋 ∗ = arg min 𝜋 ∑︁𝑛 𝑖=1 (𝐿𝑖(u𝑖(𝜋)) − 𝐿𝑖(0)) s.t. 𝑀𝑖(u𝑖(𝜋)) ≤ 𝐵, ∀𝑖 ∈ {1, . . . , 𝑛}. Here, 0 denotes no mitigation, so the objective minimizes the latency overhead added by mitigation while satisfying the memory budget. For transfer-based mitigation, this cost appears as com￾munication on the critical path. Laye… view at source ↗
Figures from the paper (7 more)
Figure 6
Figure 6. Figure 6: Temporary serialization lowers concurrency only within memory-pressure intervals. mitigation setting for the entire request lifecycle, so mitiga￾tion also affects segments whose baseline memory already fits the budget. For mitigations that can be applied locally, this …
Figure 7
Figure 7. Figure 7: Xema reuses address ranges across non￾overlapping tensor lifetimes [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Xema planning workflow. Latency profiler. Given a request template and an execu￾tion configuration, the Latency Profiler measures the ex￾ecution time of the request. Profiling a complete request can take minutes or even hours, so Xema profiles the la￾tency of the opera…
Figure 9
Figure 9. Figure 9: End-to-end SLO attainment under different arrival rates, GPU counts, and SLO scales. • Heuristic. Uses the extended memory-control prim￾itives and fine-grained execution support, but greed￾ily searches parallelism, memory control, and concur￾rency. Xema uses the extend…
Figure 10
Figure 10. Figure 10: SLO attainment under a skewed request mix. gap comes from the combination of batching and selective memory mitigation: Flux.2 benefits from concurrency, but coarse or poorly planned mitigation quickly turns the extra latency into SLO violations. For video workloads, X…
Figure 12
Figure 12. Figure 12: Memory usage and per-step latency of different allocators [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 14
Figure 14. Figure 14: Configuration ablation on LTX-2 and Flux.2. infeasible or unnecessary configurations, taking 6.3 hours in total. Xema first performs lifecycle analysis to obtain the memory trace, then prunes candidates that are clearly infea￾sible or already violate the SLO. It furth…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 3 canonical work pages

  1. [1]

    Sitara- man, and Hui Guan

    Sohaib Ahmad, Qizheng Yang, Haoliang Wang, Ramesh K. Sitara- man, and Hui Guan. 2024. DiffServe: Efficiently Serving Text- to-Image Diffusion Models with Query-Aware Model Scaling. arXiv:2411.15381 [cs.DC] doi:10.48550/arXiv.2411.15381

    arXiv doi:10.48550/arxiv.2411.15381 2024

  2. [2]

    Apache TVM Contributors. 2021. Unified Static Memory Plan- ning.https://discuss.tvm.apache.org/t/rfc-unified-static-memory- planning/10099

    2021

  3. [3]

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros. 2022. InstructPix2Pix: Learning to Follow Image Editing Instructions. arXiv:2211.09800 [cs.CV] doi:10.48550/arXiv.2211.09800

    arXiv doi:10.48550/arxiv.2211.09800 2022

  4. [4]

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhen- guo Li. 2023. PixArt- 𝛼: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. arXiv:2310.00426 [cs.CV] doi:10.48550/arXiv.2310.00426

    arXiv doi:10.48550/arxiv.2310.00426 2023

  5. [5]

    Chin-Yi Cheng, Forrest Huang, Gang Li, and Yang Li. 2023. PLay: Parametrically Conditioned Layout Generation using Latent Diffusion. arXiv:2301.11529 [cs.LG] doi:10.48550/arXiv.2301.11529

    work page Pith review arXiv doi:10.48550/arxiv.2301.11529 2023

  6. [6]

    Jiarui Fang, Jinzhe Pan, Jiannan Wang, Aoyu Li, and Xibo Sun. 2024. PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transform- ers Inference. arXiv:2405.14430 [cs.CV] doi:10.48550/arXiv.2405.14430

    arXiv doi:10.48550/arxiv.2405.14430 2024

  7. [7]

    Cong Guo, Rui Zhang, Jiale Xu, Jingwen Leng, Zihan Liu, Ziyu Huang, Minyi Guo, Hao Wu, Shouren Zhao, Junping Zhao, and Ke Zhang. 2024. GMLake: Efficient and Transparent GPU Memory Defragmentation for Large-scale DNN Training with Virtual Memory Stitching. InProceed- ings of the 29th ACM International Conference on Architectural Support for Programming Lan...

    doi:10.1145/3620665.3640423 2024

  8. [8]

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. 2024. LTX-Video: Realtime Video Latent Diffusion. arXiv:2501.00103 [cs.CV] doi:10.48550/arXiv.2501.00103

    arXiv doi:10.48550/arxiv.2501.00103 2024

Show all 54 references
  1. [9]

    Kingma, Ben Poole, Mohammad Norouzi, David J

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. 2022. Imagen Video: High Definition Video Generation with Diffusion Models. arXiv:2210.02303 [cs.CV] doi:10.48...

    2022 doi arXiv

  2. [10]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. InAdvances in Neural Information Processing Systems.https://arxiv.org/abs/2006.11239

    2020 arXiv

  3. [11]

    Zixiao Huang, Junhao Hu, Hao Lin, Chunyang Zhu, Yueran Tang, Quanlu Zhang, Zhen Guo, Zhenhua Li, Shengen Yan, Zhenhua Zhu, Guohao Dai, and Yu Wang. 2025. STAlloc: Enhancing Memory Effi- ciency in Large-Scale Model Training with Spatio-Temporal Planning. arXiv:2507.16274 [cs.LG...

    2025 doi

  4. [12]

    Gonzalez, and Ion Stoica

    Paras Jain, Ajay Jain, Tianjun Zhang, Pieter Abbeel, Joseph E. Gonzalez, and Ion Stoica. 2020. Checkmate: Breaking the Memory Wall with Optimal Tensor Rematerialization. InProceedings of Machine Learning and Systems.https://arxiv.org/abs/1910.02653

    2020 arXiv

  5. [13]

    Xiaoxiao Jiang, Suyi Li, Lingyun Yang, Tianyu Feng, Zhipeng Di, Weiyi Lu, Guoxuan Zhu, Xiu Lin, Kan Liu, Yinghao Yu, Tao Lan, Guodong Yang, Lin Qu, Liping Zhang, and Wei Wang. 2026. FlashPS: Efficient Generative Image Editing with Mask-aware Caching and Scheduling. InProceedin...

    2026

  6. [14]

    Ryoo, and Tian Xie

    Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S. Ryoo, and Tian Xie. 2024. Adaptive Caching for Faster Video Generation with Diffusion Transformers. arXiv:2411.02397 [cs.CV] doi:10.48550/arXiv.2411.02397

    2024 doi arXiv

  7. [15]

    Marisa Kirisame, Steven Lyubomirsky, Altan Haan, Jennifer Bren- nan, Mike He, Jared Roesch, Tianqi Chen, and Zachary Tatlock. 2021. 13 Conference’17, July 2017, Washington, DC, USA Kang et al. Dynamic Tensor Rematerialization. InInternational Conference on Learning Representat...

    2021 arXiv

  8. [16]

    Ioannis Lamprou, Zhen Zhang, Javier de Juan, Hang Yang, Yongqiang Lai, Etienne Filhol, and Cédric Bastoul. 2023. Safe Optimized Static Memory Allocation for Parallel Deep Learning. InProceedings of Machine Learning and Systems. https://proceedings.mlsys.org/paper_files/paper/2...

    2023

  9. [17]

    Maksim Levental. 2022. Memory Planning for Deep Neural Networks. arXiv:2203.00448 [cs.LG] doi:10.48550/arXiv.2203.00448

    2022 doi arXiv

  10. [18]

    Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Ming-Yu Liu, Kai Li, and Song Han. 2024. DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models. arXiv:2402.19481 [cs.CV] doi:10.48550/arXiv.2402.19481

    2024 doi arXiv

  11. [19]

    Suyi Li, Lingyun Yang, Xiaoxiao Jiang, Hanfeng Lu, Dakai An, Zhipeng Di, Weiyi Lu, Jiawei Chen, Kan Liu, Yinghao Yu, Tao Lan, Guodong Yang, Lin Qu, Liping Zhang, and Wei Wang. 2024. Swift- Diffusion: Efficient Diffusion Model Serving with Add-on Modules. arXiv:2407.02031 [cs.D...

    2024 doi arXiv

  12. [20]

    Suyi Li, Lingyun Yang, Xiaoxiao Jiang, Hanfeng Lu, Dakai An, Zhipeng Di, Weiyi Lu, Jiawei Chen, Kan Liu, Yinghao Yu, Tao Lan, Guodong Yang, Lin Qu, Liping Zhang, and Wei Wang. 2025. Katz: Efficient Workflow Serving for Diffusion Models with Many Adapters. In2025 USENIX Annual ...

    2025

  13. [21]

    Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung- Yi Lin. 2022. Magic3D: High-Resolution Text-to-3D Content Creation. arXiv:2211.10440 [cs.CV] doi:10.48550/arXiv.2211.10440

    2022 doi arXiv

  14. [22]

    Dong Liu, Yanxuan Yu, Jiayi Zhang, Yifan Li, Ben Lengerich, and Ying Nian Wu. 2025. FastCache: Fast Caching for Dif- fusion Transformer Through Learnable Linear Approximation. arXiv:2505.20353 [cs.LG] doi:10.48550/arXiv.2505.20353

    2025 doi

  15. [23]

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. 2025. Timestep Embedding Tells: It’s Time to Cache for Video Diffusion Model. InIEEE/CVF Conference on Computer Vision and Pattern Recog- nition.https://arxiv.or...

    2025 arXiv

  16. [24]

    Xuewen Liu, Zhikai Li, and Qingyi Gu. 2025. CacheQuant: Comprehen- sively Accelerated Diffusion Models. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition.https://arxiv.org/abs/2503.01323

    2025 arXiv

  17. [25]

    Ma, Ang Chen, and Mosharaf Chowdhury

    Runyu Lu, Shiqi He, Wenxuan Tan, Shenggui Li, Ruofan Wu, Jeff J. Ma, Ang Chen, and Mosharaf Chowdhury. 2026. TetriS- erve: Efficient DiT Serving for Heterogeneous Image Generation. arXiv:2510.01565 [cs.LG] doi:10.48550/arXiv.2510.01565

    2026 doi arXiv

  18. [26]

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. 2024. DeepCache: Accelerating Diffusion Models for Free. InIEEE/CVF Conference on Computer Vision and Pattern Recognition.https://arxiv.org/abs/2312. 00858

    2024

  19. [27]

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. 2021. SDEdit: Guided Im- age Synthesis and Editing with Stochastic Differential Equations. arXiv:2108.01073 [cs.CV] doi:10.48550/arXiv.2108.01073

    2021 doi arXiv

  20. [28]

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. 2021. GLIDE: Towards Photorealistic Image Generation and Editing with Text- Guided Diffusion Models. arXiv:2112.10741 [cs.CV] doi:10.48550/arXiv. 2112.10741

    2021 doi arXiv

  21. [29]

    OpenXLA Contributors. 2026. XLA: Optimizing Compiler for Machine Learning.https://openxla.org/xla

    2026

  22. [30]

    William Peebles and Saining Xie. 2023. Scalable Diffusion Models with Transformers. InIEEE/CVF International Conference on Computer Vision. 4195–4205.https://arxiv.org/abs/2212.09748

    2023 arXiv

  23. [31]

    Barron, and Ben Milden- hall

    Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Milden- hall. 2022. DreamFusion: Text-to-3D using 2D Diffusion. arXiv:2209.14988 [cs.CV] doi:10.48550/arXiv.2209.14988

    2022 doi arXiv

  24. [32]

    Ramya Prabhu, Ajay Nayak, Jayashree Mohan, Ramachandran Ramjee, and Ashish Panwar. 2025. vAttention: Dynamic Memory Management for Serving LLMs without PagedAttention. InProceedings of the 30th ACM International Conference on Architectural Support for Program- ming Languages a...

    2025

  25. [33]

    PyTorch Contributors. 2026. CUDA Semantics: Memory Manage- ment.https://docs.pytorch.org/docs/main/notes/cuda.html#cuda- memory-management

    2026

  26. [34]

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He

  27. [35]

    InInternational Conference for High Performance Comput- ing, Networking, Storage and Analysis

    ZeRO: Memory Optimizations Toward Training Trillion Param- eter Models. InInternational Conference for High Performance Comput- ing, Networking, Storage and Analysis. 1–16. doi:10.1109/SC41405.2020. 00024

    2020 doi arXiv

  28. [36]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. 2022. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv:2204.06125 [cs.CV] doi:10.48550/arXiv.2204.06125

    2022 doi arXiv

  29. [37]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. 10684–10695. doi:10.1109/CVPR52688.2022.01042

    2022 doi

  30. [38]

    Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. 2022. Photorealistic Text-to-Image Diffus...

    2022 doi arXiv

  31. [39]

    SGLang Team. 2026. SGLang Diffusion Models.https://docs.sglang. io/supported_models/diffusion_models.html

    2026

  32. [40]

    Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wen- qing Zhang, Vincent Y. F. Tan, and Song Bai. 2023. DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing. arXiv:2306.14435 [cs.CV] doi:10.48550/arXiv.2306.14435

    2023 doi arXiv

  33. [41]

    Andy Shih, Suneel Belkhale, Stefano Ermon, and Dorsa Sadigh. 2023. Parallel Sampling of Diffusion Models. InAdvances in Neural Informa- tion Processing Systems.https://arxiv.org/abs/2305.16317

    2023 arXiv

  34. [42]

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. 2022. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv:2209.14792 [cs.CV] doi:10. 48550/ar...

    2022 arXiv

  35. [43]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2021. Denoising Diffusion Implicit Models. InInternational Conference on Learning Representations.https://arxiv.org/abs/2010.02502

    2021 arXiv

  36. [44]

    Desen Sun, Zepeng Zhao, and Yuke Wang. 2026. MixFusion: A Patch- Level Parallel Serving System for Mixed-Resolution Diffusion Models. InProceedings of the 31st ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP ’26). ACM. doi:10.1145/ 3774934.3786420

    2026

  37. [45]

    Yu Tang, Chenyu Wang, Yufan Zhang, Yuliang Liu, Xingcheng Zhang, Linbo Qiao, Zhiquan Lai, and Dongsheng Li. 2022. DELTA: Dy- namically Optimizing GPU Memory beyond Tensor Recomputation. arXiv:2203.15980 [cs.LG] doi:10.48550/arXiv.2203.15980

    2022 doi arXiv

  38. [46]

    vLLM-Omni Team. 2026. CPU Offloading for Diffusion Mod- els.https://docs.vllm.ai/projects/vllm-omni/en/stable/user_guide/ diffusion/cpu_offload_diffusion/

    2026

  39. [47]

    vLLM-Omni Team. 2026. HSDP.https://docs.vllm.ai/projects/vllm- omni/en/latest/design/feature/hsdp/

    2026

  40. [48]

    vLLM-Omni Team. 2026. vllm-omni serve.https://docs.vllm.ai/ projects/vllm-omni/en/latest/cli/serve/ 14 Xema: Efficient Diffusion Serving through Fine-Grained Memory Management and Auto-Configuration Conference’17, July 2017, Washington, DC, USA

    2026

  41. [49]

    Linnan Wang, Jiarui Fang, Xijie Zhang, Xupeng Miao, Yang Yu, Xi- aowen Chu, and Bin Cui. 2020. SwapAdvisor: Pushing Deep Learn- ing Beyond the GPU Memory Limit via Smart Swapping. InProceed- ings of the Twenty-Fifth International Conference on Architectural Sup- port for Progr...

    2020 doi

  42. [50]

    Yifei Xia, Fangcheng Fu, Hao Yuan, Hanke Zhang, Xupeng Miao, Yijun Liu, Suhan Ling, Jie Jiang, and Bin Cui. 2025. TridentServe: A Stage- level Serving System for Diffusion Pipelines. arXiv:2510.02838 [cs.DC] doi:10.48550/arXiv.2510.02838

    2025 doi

  43. [51]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Bin Xu, Yuxiao Dong, and Jie Tang. 2024. CogVideoX: Text-to-Video Diffusion Models with An Expert T...

    2024 arXiv

  44. [52]

    Peiqi Yin, Jiangyun Zhu, Han Gao, Chenguang Zheng, Yongxiang Huang, Taichang Zhou, Ruirui Yang, Weizhi Liu, Weiqing Chen, Canlin Guo, Didan Deng, Zifeng Mo, Cong Wang, James Cheng, Roger Wang, and Hongsheng Liu. 2026. vLLM-Omni: Fully Disaggregated Serving for Any-to-Any Multi...

    2026

  45. [53]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding Conditional Control to Text-to-Image Diffusion Models. arXiv:2302.05543 [cs.CV] doi:10.48550/arXiv.2302.05543

    2023 doi arXiv

  46. [54]

    Yinmin Zhong, Shengyu Liu, Junda Chen, Jianbo Hu, Yibo Zhu, Xu- anzhe Liu, Xin Jin, and Hao Zhang. 2024. DistServe: Disaggregat- ing Prefill and Decoding for Goodput-optimized Large Language Model Serving. In18th USENIX Symposium on Operating Systems Design and Implementation ...

    2024

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.