Pith. sign in

REVIEW 4 major objections 6 minor 27 references

A generic coding agent, steered by a structured workflow and a verification loop, turns single-GPU reference code into multi-GPU deployments with ~70x lower latency and up to 3.6x higher throughput — and beats an expert baseline by 65% on o

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 15:45 UTC pith:LEWOFZRB

load-bearing objection Useful systems paper with a solid formal core; the empirical claims need verification, especially the KV-cache independence assumption behind the 8-GPU LiveAvatar number. the 4 major comments →

arxiv 2607.18171 v1 pith:LEWOFZRB submitted 2026-07-20 cs.LG

FlashRT: Agent Harness for Guiding Agents to Deploy Real-Time Multimodal Applications

Krish Agarwal , Zhuoming Chen , Yanyuan Qin , Zhenyu Gu , Atri Rudra , Beidi Chen This is my paper
classification cs.LG MSC 68M2090B35
keywords real-time multimodal servingcoding agentschain-of-programintermediate representationpipeline parallelismlatency-throughput tradeoffmulti-GPU deploymentagentic deployment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that hand-engineering a serving system for every new real-time multimodal application is unnecessary. A generic coding agent, steered by a structured workflow it calls chain-of-program, can take an ordinary single-GPU reference implementation and turn it into an efficient multi-GPU deployment. The workflow first forces the agent to encode the pipeline as a hierarchical graph IR — marking which edges can stream and which nodes share persistent state — then validates the IR, then runs a self-directed loop of implement, verify, and benchmark. Reported results include ~70x latency reduction, up to 3.6x throughput gain across two GPU platforms, and, for one multimodal-LLM workload, 65% lower latency than an expert hand-built serving implementation. If true, deploying voice agents, avatars, and video models stops being manual systems engineering.

Core claim

The central claim is that the bottleneck in serving real-time multimodal applications is discovering a good deployment — placement, streaming, and intra-model parallelism — and that a coding agent can do this if the workflow forces an explicit program-analysis stage. A naive agent, told to optimize directly, misses model-level pipeline parallelism and fails to compose strategies; FlashRT's IR, with persistent-state and streaming annotations, surfaces these opportunities, and a measurement-gated loop validates each hypothesis against real buffers. In the conversational-agent case study, the agent composes streaming, disaggregation, and diffusion-step pipeline parallelism to cut latency ~70x a

What carries the argument

The carrying object is the chain-of-program workflow: the agent first lifts a reference implementation into a hierarchical graph IR — nodes are self-contained operations, edges are data dependencies marked blocking or streaming, node annotations record persistent-state reads/writes (the cross-batch, λ=1 edges of the formal model) — then runs a self-driven loop of hypothesis, implementation, element-wise correctness check, and benchmark, directed by a re-ranked variant queue. The formal backbone is the task-graph model G=(V,E) with placement ρ and non-preemptive schedules; two bounds drive everything: the path bound ties latency to the critical path (co-location wins), the resource-load bound

Load-bearing premise

The headline 8-GPU frame-rate result depends on the assumption, asserted but not measured in the paper, that each LiveAvatar diffusion step uses an independent KV cache, so the four denoising steps can be pipelined across four GPUs with no synchronization; if denoising steps share state, that deployment is semantically invalid and the 173.67 FPS claim collapses.

What would settle it

Run the 8-GPU pipeline-parallel deployment and the single-GPU reference on identical spoken prompts and compare outputs; any divergence on any prompt fails the deployment's correctness claim. More directly, instrument LiveAvatar to log KV-cache accesses per diffusion step: if step i at batch j+1 reads state written by step i−1 of batch j, the cross-batch edges (Qi, Qi, 1) carry a nonzero synchronization lag that Appendix A.4 sets to zero.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A developer can hand over a plain single-GPU reference implementation and receive a working multi-GPU deployment, with the same workflow producing latency-optimized or throughput-optimized variants across 1–8 GPU budgets.
  • The workflow transfers across hardware: re-run from scratch on AMD MI355X with the same references and budgets, it recovers the same deployment families and shows larger relative gains (3.6x vs 2.8x throughput) where expert optimizations are less mature.
  • For Qwen3-Omni text-to-audio, the agent rediscovers the disaggregation and streaming structure of the hand-engineered vLLM-Omni serving implementation and reduces response latency by 65% via lighter-weight inter-component transfer.
  • The formal analysis explains why the agent's choices are principled: co-location minimizes the latency-binding critical path while disaggregation minimizes the sustainable period, so each discovered deployment is optimal for its stated target.
  • Because the deployment problem is NP-hard even without dependencies or communication costs, per-application search by an agent is not a convenience but a necessity — no fixed policy or exact solver can replace it.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension of the same harness would point it at embodied visuomotor policies or live audio pipelines, which share the streaming-stage structure; whether the agent rediscovers the same disaggregation patterns would probe how general the IR annotations are.
  • The paper does not ablate its own recipe: giving the agent identical prompts but stripped-down IR annotations (no state or streaming marks) would measure how much of the ~70x comes from the workflow versus the agent's priors.
  • The 8-GPU result leans on an architectural claim the paper asserts but does not measure — that LiveAvatar's diffusion steps keep independent KV caches. Directly probing that cache independence would show whether diffusion-step pipeline parallelism generalizes to other DiT-based generators or is specific to this checkpoint.
  • The variant queue's re-ranking hints at a longer-term direction: a learned prior over which deployment axes matter per application class could replace the loop's cold start and cut the agent's compute budget per application.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. FlashRT proposes an agent harness that takes a synchronous, single-GPU reference implementation of a multimodal application and guides a generic coding agent through a structured workflow — IR construction, static dependency/state analysis, candidate transformation, and a measurement-gated validation loop — to produce multi-GPU deployments optimized for latency or throughput. The paper contributes a formal pipeline-graph framework for reasoning about co-location, disaggregation, streaming, and pipeline/sequence parallelism, an NP-hardness argument for the general placement problem, and experiments on five applications (LiveAvatar conversational agent, Qwen3-Omni, Krea-Realtime+SAM 3, WorldPlay, LongLive) on NVIDIA B200 and AMD MI355X. Headline claims include up to ~70× latency reduction, up to 3.6× throughput gain, and a 65% latency reduction over vLLM-Omni for Qwen3-Omni on MI355X.

Significance. If the claims hold, FlashRT would be a valuable step toward automating application-specific multimodal serving, reducing the manual systems effort currently required for each new interactive pipeline. The formal Appendix A framework is a genuine strength: it gives clean latency and throughput characterizations for the DiT–VAE and sequence-vs-pipeline-parallelism tradeoffs, and the NP-hardness reduction is sound. The paper is also transparent in places: it labels the 8-GPU frame rate as theoretical, provides a GitHub repository and blog, and acknowledges in Section 6 that kernel-level agent optimization is not integrated and only one agent configuration was tested. However, the empirical evidence is currently too thin to establish the generality claimed: agent runs are stochastic and unrepeated, the showcase 8-GPU deployment rests on an unverified architectural assumption, and the only expert-system baseline is vLLM-Omni. These issues are fixable, but they are load-bearing for the central claims.

major comments (4)
  1. [§4.1 and Appendix A.4; Tables 1–2] The 8-GPU LiveAvatar deployment is legal only under the asserted property that "each diffusion step uses an independent KV cache." This is not cited to the LiveAvatar paper and is not verified by any code excerpt, measurement, or reference in the manuscript. If denoising steps share state across steps or chunks, placing DiT1–DiT4 on separate GPUs changes semantics and the 173.67/78.5 FPS rows collapse. The agent's output-equivalence check on sample inputs cannot detect timing-dependent cross-batch state leakage. Please supply source-level evidence for the per-step independent-cache property, or remove/qualify the 8-GPU rows.
  2. [§5.1–§5.4, Tables 1, 3, 9] The empirical methodology is too thin for the strength of the claims. Each application is reported as a single agent session; coding agents are stochastic, so without repeated runs, seeds, or variance statistics the reported ~70×, 3.6×, and 65% numbers are single draws. Table 1's frame rate is described as "theoretical" in Section 5.2; if it is not an end-to-end measured frame rate, the throughput claims are not established. In addition, there is no quantitative ablation: Section 4.1 gives a qualitative failure mode for a naive agent, but no head-to-head success rates, iteration counts, or cost statistics that demonstrate the chain-of-program and validation-loop components are responsible for the reported gains.
  3. [§5.3.1, Table 3; §6] The comparison against expert/hand-engineered systems consists solely of vLLM-Omni on Qwen3-Omni. The abstract and conclusion claim that FlashRT deployments are "on par or more performant" than other serving systems and expert-written systems, but LiveAvatar, WorldPlay, LongLive, and Krea-Realtime have no expert baseline. The paper's own Section 6 limitation — one agent configuration and no kernel optimization agents — further limits the generality of the claim. Either add expert baselines for additional applications or temper the general claim to what is actually measured.
  4. [§4.3] Verification of agent-generated deployments is underspecified for generative and streaming pipelines. "Element-wise output equivalence" is not defined for stochastic generative outputs: what tolerance, what inputs, how many samples, and how are streaming/ordering properties checked? Because the correctness claim is central to trusting a deployment that uses disaggregation and pipelining, the verification protocol needs to be concrete enough to be reproduced and to rule out state leakage across chunks, not just output similarity on a few sample inputs.
minor comments (6)
  1. [§4.2] Typo: "more more robust" should read "more robust" or "far more robust."
  2. [§5.2] Typo: "pipline parallism" likely should be "pipeline parallelism."
  3. [§3 and §4.2] The paper says the user reference defines persistent state scopes, but later the agent is required to annotate node-level persistent-state reads/writes. Clarify who is responsible for identifying the state scopes: the user, the IR interpreter, or the agent's static analysis.
  4. [Abstract and §5.3.2] The abstract states 2.8× throughput on B200, while Table 4 shows 2.8× for one deployment and other tables show different gains. Make the "up to" qualifiers consistent with the specific rows they refer to.
  5. [§5.4 and Table 9] The MI355X baseline latencies differ from the B200 baselines (e.g., 101.17 s vs. 107.92 s for LiveAvatar). If the references are identical, this is expected hardware variability, but reporting the same baseline under the same protocol would help readers separate hardware effects from deployment effects.
  6. [Appendix B] Phrases such as "1.6× lower latency" are ambiguous: specify whether the factor is a reduction (new/old) or an improvement (old/new) consistently.

Circularity Check

0 steps flagged

No significant circularity: the performance claims are externally benchmarked, and the load-bearing LiveAvatar premise is an explicit external assumption rather than a fitted or self-cited input.

full rationale

FlashRT's central contributions are empirical and architectural, not deductive predictions from its own fitted inputs. The headline numbers (~70x latency reduction, 3.6x throughput, 65% Qwen3-Omni latency improvement) are measured against external baselines—the sequential reference implementation and vLLM-Omni—so they are not constructed to equal their inputs. The IR design does encode persistent-state scopes, and the provided static tools surface the parallelization/streaming opportunities that follow from those annotations, which means part of the 'agent discovers X' narrative is shaped by the harness; however, this is an attribution/claim-validity concern, not a circular derivation. The agent must still generate correct annotations from code, implement each variant, and pass an output-equivalence check, and the benchmark gains are externally measured rather than derived from the annotations themselves. The formal propositions in Appendices A.3 and A.4 are transparent algebraic consequences of explicitly stated model assumptions, not hidden fits. The load-bearing LiveAvatar assumption that 'each diffusion step uses an independent KV cache' is explicitly stated in Section 4.1 and formalized in Appendix A.4 as inter-batch edges (Qi, Qi, 1) with zero synchronization lag; the paper itself notes that without that structure the PP schedule is infeasible ('For DiTs without the per-step inter-batch dependency structure above, the PP schedule analyzed here is not feasible under the stated framework'). This is an unverified external premise bearing correctness risk, but it is not a circular step: it is neither derived from nor equivalent to the paper's own outputs. There are no fitted parameters renamed as predictions, no uniqueness theorem imported from the same authors, and no ansatz smuggled in via self-citation. Therefore the derivation chain is self-contained, and the correct circularity finding is score 0.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No free parameters are fitted to data. The axioms are standard scheduling/domain assumptions plus two methodological assumptions about the agent's verification and benchmark process. The per-step KV-cache assumption is the most fragile and is structurally necessary for the strongest 8-GPU result. No new physical entities are introduced.

axioms (5)
  • domain assumption Per-step independent KV cache in LiveAvatar-style DiTs.
    Section 4.1 and Appendix A.4 assume each diffusion step has an independent KV cache, enabling cross-GPU pipeline parallelism. Not independently verified.
  • domain assumption Earliest-start schedule achieves the path lower bound whenever T is sustainable.
    Used to turn Lemma A.2 bounds into equalities in Propositions A.4 and A.6; not proven for general schedules.
  • domain assumption Sustainable throughput is governed only by the most loaded resource, ignoring communication fabric as a bottleneck.
    Lemma A.3 models throughput via per-resource load; communication costs affect latency but not the sustainable-period bound except when modeled as an additional resource.
  • domain assumption The coding agent can faithfully implement and benchmark arbitrary pipeline variants.
    Section 4.3 and Section 5.1 assume the agent's generated test harness accurately measures user-facing latency and throughput.
  • domain assumption Element-wise output equivalence is a valid correctness check for generative model outputs.
    Section 4.3 requires verifying implementations against a baseline, but generative models are stochastic and exact equivalence is not well-defined; tolerance criteria are not specified.

pith-pipeline@v1.3.0-alltime-deepseek · 26383 in / 11779 out tokens · 116460 ms · 2026-08-01T15:45:35.835127+00:00 · methodology

0 comments
read the original abstract

Real-time multimodal applications, including voice agents and interactive video generation, compose heterogeneous models into pipelines whose efficient deployment requires application-specific decisions about placement, streaming, and intra-model parallelism. Existing serving systems and auto-parallelism compilers commit to limited transformations and fixed workload assumptions, so achieving high performance on a new application requires hand-crafting an efficient implementation. We present FlashRT, an agent harness that guides coding agents to lift simple developer-written reference implementations into optimized multi-GPU deployments that flexibly weigh target metrics like latency and throughput. Using a new chain-of-program paradigm, FlashRT directs a generic coding agent through a multi-pass transformation process where an agent transforms the reference into an intermediate representation (IR) to capture data dependencies and persistent-state scopes, validates this IR via a sequential interpreter, and performs static analyses to identify candidate transformations. Then, the agent iteratively implements, verifies, and benchmarks each candidate under a measurement-gated optimization loop to produce effective deployments that span different hardware budgets. Across various applications, including video world models and multimodal LLMs, FlashRT converts reference implementations into highly efficient deployments, delivering up to ~70x latency reduction and 2.8x throughput improvement on NVIDIA B200 GPUs. On AMD MI355X GPUs, FlashRT matches the peak latency reduction while increasing peak throughput improvement to 3.6x, demonstrating that agent-driven optimization can be more scalable on platforms with less mature expert optimization. In fact, for Qwen3-Omni text-to-audio inference, FlashRT reduces response latency by 65% compared to the expert vLLM-Omni implementation on AMD MI355X.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 19 linked inside Pith

  1. [1]

    Claude code overview

    Anthropic. Claude code overview. https://code.claude.com/docs/en/overview, 2026a. Documentation for Claude Code; accessed 2026-05-04. Anthropic. Claude opus 4.7. https://www.anthropic.com/news/claude-opus-4-7 , April 2026b. Model ID: claude-opus- 4-7; accessed 2026-05-04. Phil Ball, Jakob Bauer, Frank Belletti, Bethanie Brownfield, Ariel Ephrat, Shlomi Fr...

    2026

  2. [3]

    π0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164,

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky. π0: A visi...

    Pith/arXiv arXiv

  3. [6]

    Advancing open-source world models.arXiv preprint arXiv:2601.20540,

    Zelin Gao, Qiuyu Wang, Yanhong Zeng, Jiapeng Zhu, Ka Leong Cheng, Yixuan Li, Hanlin Wang, Yinghao Xu, Shuailei Ma, Yihang Chen, Jie Liu, Yansong Cheng, Yao Yao, Jiayi Zhu, Yihao Meng, Kecheng Zheng, Qingyan Bai, Jingye Chen, Zehong Shen, Yue Yu, Xing Zhu, Yujun Shen, and Hao Ouyang. Advancing open-source world models.arXiv preprint arXiv:2601.20540,

    Pith/arXiv arXiv

  4. [7]

    Qwen3-tts technical report, 2026.https://arxiv.org/abs/2601.15621

    Hangrui Hu, Xinfa Zhu, Ting He, Dake Guo, Bin Zhang, Xiong Wang, Zhifang Guo, Ziyue Jiang, Hongkun Hao, Zishan Guo, Xinyu Zhang, Pei Zhang, Baosong Yang, Jin Xu, Jingren Zhou, and Junyang Lin. Qwen3-tts technical report, 2026.https://arxiv.org/abs/2601.15621. Yubo Huang, Hailong Guo, Fangtai Wu, Shifeng Zhang, Shijie Huang, Qijun Gan, Lin Liu, Sirui Zhao,...

    Pith/arXiv arXiv 2026

  5. [8]

    Zhihao Jia, Oded Padon, James Thomas, Todd Warszawski, Matei Zaharia, and Alex Aiken

    https://arxiv.org/abs/1807.05358. Zhihao Jia, Oded Padon, James Thomas, Todd Warszawski, Matei Zaharia, and Alex Aiken. TASO: Optimizing deep learning computation with automatic generation of graph substitutions. InProceedings of the 27th ACM Symposium on Operating Systems Principles (SOSP),

    Pith/arXiv arXiv

  6. [12]

    Ma, Jae-Won Chung, Jisang Ahn, Yizhuo Liang, Akshay Jajoo, Myungjin Lee, and Mosharaf Chowdhury

    Jeff J. Ma, Jae-Won Chung, Jisang Ahn, Yizhuo Liang, Akshay Jajoo, Myungjin Lee, and Mosharaf Chowdhury. Cornserve: Efficiently serving any-to-any multimodal models.arXiv preprint arXiv:2512.14098,

    Pith/arXiv arXiv

  7. [13]

    Krea realtime 14b: Real-time video generation, 2025.https://github.com/krea-ai/realtime-video

    Erwann Millon. Krea realtime 14b: Real-time video generation, 2025.https://github.com/krea-ai/realtime-video. Deepak Narayanan, Aaron Harlap, Amar Phanishayee, Vivek Seshadri, Nikhil R. Devanur, Gregory R. Ganger, Phillip B. Gibbons, and Matei Zaharia. PipeDream: Generalized pipeline parallelism for DNN training. InProceedings of the 27th ACM Symposium on...

    2025

  8. [14]

    Alexander Novikov, Ngan Vu, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergei Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, et al. AlphaEvolve: A coding agent for scientific and algorithmic discovery.arXiv preprint arXiv:2506.13131,

    Pith/arXiv arXiv

  9. [15]

    Zhang, William Hu, Christopher Re, and Azalia Mirhoseini

    Anne Ouyang, Simon Guo, Simran Arora, Alex L. Zhang, William Hu, Christopher Re, and Azalia Mirhoseini. KernelBench: Can LLMs write efficient GPU kernels?arXiv preprint arXiv:2502.10517,

    Pith/arXiv arXiv

  10. [16]

    ModServe: Scalable and resource-efficient large multimodal model serving.arXiv preprint arXiv:2502.00937,

    Haoran Qiu, Anish Biswas, Zihan Zhao, Kunal Mahajan, Inigo Goiri Han, Chetan Bansal, Rodrigo Fonseca, Ramachan- dran Iyer, and Ricardo Bianchini. ModServe: Scalable and resource-efficient large multimodal model serving.arXiv preprint arXiv:2502.00937,

    arXiv

  11. [17]

    Seamless: Multilingual expressive and streaming speech translation.arXiv preprint arXiv:2312.05187,

    Seamless Communication, Lo ¨ ıc Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Mark Duppenthaler, Paul-Ambroise Duquenne, Brian Ellis, Hady Elsahar, Justin Haaheim, John Hoffman, Min-Jae Hwang, Hirofumi Inaguma, Christopher Klaiber, Ilia Kulikov, Pengwei Li, Daniel Licht, Jean Maillard, Ruslan Mavlyutov, Alice Rakotoarison, et al. S...

    Pith/arXiv arXiv

  12. [19]

    WorldPlay: Towards long-term geometric consistency for real-time interactive world modeling.arXiv preprint arXiv:2512.14614,

    Wenqiang Sun, Haiyu Zhang, Haoyuan Wang, Junta Wu, Zehan Wang, Zhenwei Wang, Yunhong Wang, Jun Zhang, Tengfei Wang, and Chunchao Guo. WorldPlay: Towards long-term geometric consistency for real-time interactive world modeling.arXiv preprint arXiv:2512.14614,

    Pith/arXiv arXiv

  13. [20]

    AwareCompiler: Agentic context-aware compiler optimization via a synergistic knowledge-data driven framework.arXiv preprint arXiv:2510.11759,

    Hongyu Tan, Jiawei Wang, Yuxiang Liu, Wenliang Sun, and Tianyi Zhao. AwareCompiler: Agentic context-aware compiler optimization via a synergistic knowledge-data driven framework.arXiv preprint arXiv:2510.11759,

    arXiv

  14. [22]

    Astra: A multi-agent system for GPU kernel performance optimization.arXiv preprint arXiv:2509.07506,

    Anjiang Wei, Tianran Sun, Yogesh Seenichamy, Hang Song, Anne Ouyang, Azalia Mirhoseini, Ke Wang, and Alex Aiken. Astra: A multi-agent system for GPU kernel performance optimization.arXiv preprint arXiv:2509.07506,

    arXiv

  15. [23]

    Chain-of-thought prompting elicits reasoning in large language models, 2023.https://arxiv.org/abs/2201.11903

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023.https://arxiv.org/abs/2201.11903. Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Y...

    Pith/arXiv arXiv 2023

  16. [25]

    CudaForge: An agent framework with hardware feedback for CUDA kernel optimization.arXiv preprint arXiv:2511.01884, 2025b

    Zijian Yang, Junfeng Liu, Yifan Liu, Yuxuan Sun, Hao Wang, and Tianlong Chen. CudaForge: An agent framework with hardware feedback for CUDA kernel optimization.arXiv preprint arXiv:2511.01884, 2025b. Peiqi Yin, Jiangyun Zhu, Han Gao, Chenguang Zheng, Yongxiang Huang, Taichang Zhou, Ruirui Yang, Weizhi Liu, Weiqing Chen, Canlin Guo, Didan Deng, Zifeng Mo, ...

    arXiv

  17. [26]

    MuseTalk: Real-time high quality lip synchronization with latent space inpainting.arXiv preprint arXiv:2410.10122,

    Yue Zhang, Minhao Liu, Zhaokang Chen, Bin Wu, Yubin Zeng, Chao Zhan, Yingjie He, Junxin Huang, and Wenjiang Zhou. MuseTalk: Real-time high quality lip synchronization with latent space inpainting.arXiv preprint arXiv:2410.10122,

    Pith/arXiv arXiv

  18. [27]

    Its κ captures any cost of moving that state between resources

    edge encodes a logical inter-batch dependency, for instance a shared KV cache that batch j + 1 inherits from batch j. Its κ captures any cost of moving that state between resources. A.1.3 Latency metric Chunk (j, k) becomes display-ready atY j,k :=C j,o(k), the completion time ofo(k) in batchj. Lemma A.1(Display-offset characterization).Fix a feasible dep...

    1977

  19. [1977]

    Shangzhan Li, Zefan Wang, Ye Tian, Yuhan Sun, Yiyi Wu, Bowen Liu, Xiao Yu, Tianyu Liu, and Yuxiang Cheng

    doi: 10.1016/S0167-5060(08)70743-X. Shangzhan Li, Zefan Wang, Ye Tian, Yuhan Sun, Yiyi Wu, Bowen Liu, Xiao Yu, Tianyu Liu, and Yuxiang Cheng. AutoTriton: Automatic Triton programming with reinforcement learning in LLMs.arXiv preprint arXiv:2507.05687, 2025a. Xiaoya Li, Xiaofei Sun, Albert Wang, Jiwei Wang, Fei Wu, Xianzhi Han, and Tianwei Yu. CUDA-L1: Imp...

    Pith/arXiv arXiv

  20. [2018]

    Longlive-2.0: An nvfp4 parallel infrastructure for long video generation, 2026.https://arxiv.org/abs/2605.18739

    Yukang Chen, Luozhou Wang, Wei Huang, Shuai Yang, Bohan Zhang, Yicheng Xiao, Ruihang Chu, Weian Mao, Qixin Hu, Shaoteng Liu, Yuyang Zhao, Huizi Mao, Ying-Cong Chen, Enze Xie, Xiaojuan Qi, and Song Han. Longlive-2.0: An nvfp4 parallel infrastructure for long video generation, 2026.https://arxiv.org/abs/2605.18739. Alexandre D´ efossez, Laurent Mazare, Manu...

    Pith/arXiv arXiv 2026

  21. [2019]

    StreamDiffusion: A pipeline-level solution for real-time interactive generation.arXiv preprint arXiv:2312.12491,

    14 Akio Kodaira, Chenfeng Xu, Toshiki Hazama, Takanori Yoshimoto, Kohei Ohno, Shogo Mitsuhori, Soichi Sugano, Hanying Cho, Zhijian Liu, Masayoshi Tomizuka, and Kurt Keutzer. StreamDiffusion: A pipeline-level solution for real-time interactive generation.arXiv preprint arXiv:2312.12491,

    Pith/arXiv arXiv

  22. [2021]

    https: //arxiv.org/abs/2105.04663. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin...

    Pith/arXiv arXiv

  23. [2022]

    ISBN 978-1-939133-28-1

    USENIX Association. ISBN 978-1-939133-28-1. https://www.usenix.org/ conference/osdi22/presentation/unger. Xingyao Wang, Boxuan Li, Yufan Song, Frank F. Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H. Tran, Fuqiang Li, Ren Ma, Mingzhang Zheng, Bill Qian, Yanjun Shao, Niklas Muennighoff, Yizhe Zhang, Binyuan Hui, ...

    Pith/arXiv arXiv

  24. [2023]

    Towards robust agentic CUDA kernel benchmarking, verification, and optimization.arXiv preprint arXiv:2509.14279,

    Robert Tjarko Lange, Aaditya Prasad, Qi Sun, Maxence Faldor, Yujin Tang, and David Ha. Towards robust agentic CUDA kernel benchmarking, verification, and optimization.arXiv preprint arXiv:2509.14279,

    arXiv

  25. [2024]

    Sam 3: Segment anything with concepts, 2026.https://arxiv.org/abs/2511.16719

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman R¨ adle, Triantafyllos Afouras, Effrosyni Mavroudi, Katherine Xu, Tsung-Han Wu, Yu Zhou, Lilian...

    Pith/arXiv arXiv 2026

  26. [2025]

    Carlo Baronio, Pietro Marsella, Ben Pan, Simon Cherniavskii, Janice Wang, and Fotis Iliopoulos

    https: //deepmind.google/blog/genie-3-a-new-frontier-for-world-models/. Carlo Baronio, Pietro Marsella, Ben Pan, Simon Cherniavskii, Janice Wang, and Fotis Iliopoulos. Kevin: Multi-turn RL for generating CUDA kernels.arXiv preprint arXiv:2507.11948,

    Pith/arXiv arXiv

  27. [2026]

    Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao

    https://arxiv.org/abs/2601.21337. Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. InAdvances in Neural Information Processing Systems (NeurIPS),

    Pith/arXiv arXiv