2025–2026 年 LLM 基础设施框架正经历从 "高吞吐 batch serving"到"agentic workload 全生命周期编排" 的范式转变。DeepSeek 的 DualPipe + PD 分离架构、vLLM/SGLang 的连续 batching 生态、以及 ByteDance veRL 的 RLHF 训练框架构成当前主流。新兴的 agentic serving 系统(ThunderAgent、Helium、Halo)、超低延迟 persistent kernel 引擎(TileRT/TokenSpeed)、和跨 DC PD 分离(PrfaaS/DualPath)正在快速扩展框架的覆盖边界。这一代框架的核心矛盾已从"怎么喂饱 GPU"转向"如何在动态多阶段工作负载下同时满足 SLO、吞吐和资源效率"。
| 维度 | 取值 | 定义 |
|---|---|---|
| Stage | training / serving / hybrid | 主要服务的推理/训练阶段 |
| Workload Pattern | batch / interactive / agentic / RL-rollout | 面向的工作负载形态 |
| Architectural Layer | scheduling / execution / communication / end-to-end | 干预的系统层次 |
| Scheduling | Execution/Runtime | Communication | End-to-End | |
|---|---|---|---|---|
| Serving-Interactive | DynaServe [2504.09285], JITServe [2504.20068], PPD [2603.13358], Justitia [ref:2510.17015] | TileRT [tile-ai-tilert], TokenSpeed [tokenspeed], Fleet [ref:2604.15379] | MFS [2603.17456], KVServe [kvserve] | vLLM Semantic Router [vllm-project-semantic-router] |
| Serving-Agentic | HexGen-Flow [ref:2505.05286], KVFlow [ref:2507.07400], Helium [ref:2603.16104], Halo [ref:2509.02121], Concur [ref:2601.22705] | ThunderAgent [ref:2602.13692], TokenDance [ref:2604.03143], Pancake [ref:2602.21477] | DualPath [2602.21548], PrfaaS [2604.15039] | Sutradhara [ref:2601.12967], Scepsy [ref:2604.15186], GLM [ref:2511.01633] |
| Serving-MoE | DynaServe [2504.09285], MoE-LB [ref:moe-lb-interai25] | ZeRO-Prefill [2605.02960] | MFS [2603.17456] | — |
| Training-RL | FlexRLHF [2312.11819], HybridFlow/veRL [2409.19256] | TensorHub [2604.09107] | — | — |
| Training-LLM | — | DeepSeek-V3/DualPipe [2412.19437] | — | — |
| Config/Auto-tune | AIConfigurator [ref:2601.06288] | Saguaro/SSD [ref:2603.03251] | — | — |
Leading: DualPath, PrfaaS, DynaServe, PPD, KVServe
PD (prefill-decode) 分离已从"单集群 RDMA"演进到跨 DC Ethernet。DualPath 用 dual-path loading + CNIC-centric flow isolation 聚合所有 SNIC 带宽 [2602.21548];PrfaaS 将 hybrid attention 降低的 KV 吞吐用跨 DC 架构变现 [2604.15039];DynaServe 用 micro-request 统一 colocation 和 disaggregation [2504.09285];PPD 按干扰差异动态路由 Turn 2+ 请求 [2603.13358];KVServe 用 service-aware 压缩解决带宽瓶颈 [kvserve]。
Leading: ThunderAgent, Helium, Halo, HexGen-Flow, KVFlow, Concur
Agentic workload(多轮、短追加、长上下文、DAG 依赖)正驱动框架从"请求级调度"转向"程序级编排"。ThunderAgent 引入 agentic program 作为一等调度单元 [ref:2602.13692];Helium 用 Templated Radix Tree 做 plan-level 优化 [ref:2603.16104];Halo 借数据库 query optimizer 思路做 batch DAG 整合 [ref:2509.02121];KVFlow 用 workflow-aware STE 替代 LRU [ref:2507.07400];Concur 用 AIMD 控制避免 KV thrashing [ref:2601.22705]。
Leading: TileRT, TokenSpeed, Fleet
近 BS=1 decode 下 kernel 间 idle 占主导。TileRT 将整个模型 AOT 编译为单个 persistent Engine Kernel,tile-level 调度实现 compute-IO-comm 持续重叠,GLM-5 达 500 tok/s [tile-ai-tilert]。TokenSpeed 用 C++ FSM + Placement 编译器 + Blackwell MLA 内核在 Kimi K2.5 上超越 TRT-LLM [tokenspeed]。Fleet 为 AMD MI350 多 die 设计 hierarchical megakernel [ref:2604.15379]。
Leading: HybridFlow/veRL, FlexRLHF, TensorHub, DeepSeek-V3
HybridFlow 用 hybrid single/multi-controller + 3D-HybridEngine 实现 1.53–20.57× 加速 [2409.19256];FlexRLHF 的 Disaggregated 策略达 11× [2312.11819];TensorHub 用 Reference-Oriented Storage 消除权重传输瓶颈 [2604.09107];DeepSeek-V3 的 DualPipe 实现跨节点 all-to-all 完全隐藏 [2412.19437]。
JITServe 的 GMAX 算法提供 1/8.55 常数竞争比 [2504.20068];MFS 用 RMLQ 做 MoE 三阶段通信隔离 [2603.17456];Justitia 用 KV token-time 做公平调度 [ref:2510.17015]。
ZeRO-Prefill 反转 EP 数据流方向,用 AsyncEP 替代 AllToAll [2605.02960];MoE-LB 用 ILP 做联合复制/分配 [ref:moe-lb-interai25]。
| 系统 | Stage | 核心指标 | Baseline 对比 | 硬件 | 代码开源 | 可复现性 |
|---|---|---|---|---|---|---|
| DualPath [2602.21548] | Serving | 1.87× offline JCT, 1.96× online APS | vs Basic (same infra) | H800+IB+3FS | 未公开 | 不可 |
| PPD [2603.13358] | Serving | Turn2+ TTFT −48–73% | vs vLLM PD | 4×H100 | 未公开 | 部分 |
| MFS [2603.17456] | Serving | SLO attainment 1.2–2.4× | vs FS/SJF/EDF/Karuna | 32 GPU testbed | 未公开 | 不可 |
| TensorHub [2604.09107] | Training | 6.7× stall reduction (standalone) | vs NCCL/UCX | 1024 GPU Hopper | 未公开 | 不可 |
| PrfaaS [2604.15039] | Serving | +54% 吞吐, −64% P90 TTFT | vs 同构 PD | 32H200+64H20 | 未公开 | 不可 |
| ZeRO-Prefill [2605.02960] | Serving | 1.35–1.37× throughput | vs DP×EP/TP/PP | 8×A100/H100/H200 | 开源 (vLLM) | 复现 |
| TileRT [tile-ai-tilert] | Serving | 600 tok/s (DSV3.2) | vs 传统框架 ~几十 tok/s | 8×B200 | 部分 | 部分 |
| TokenSpeed [tokenspeed] | Serving | Pareto 优于 TRT-LLM ~9–11% | vs TRT-LLM | B200 | 开源 (MIT) | 部分 |
| KVServe [kvserve] | Serving | up to 10× KV 压缩, 8.5× 延迟 | vs 无压缩 | A100 | 开源 (Apache) | 复现 |
| DynaServe [2504.09285] | Serving | 1.15–3.07× serving capacity | vs coloc./disagg. | A100 | 未公开 | 部分 |
| JITServe [2504.20068] | Serving | 1.4–6.3× goodput | vs Sarathi/vLLM/Autellix | 16×A100 | 未公开 | 部分 |
| HexGen-Flow [ref:2505.05286] | Serving | P95 latency 1.42–1.56× lower | vs vLLM/VTC/QLM | 未报告 | 开源 | 复现 |
| KVFlow [ref:2507.07400] | Serving | 1.83–2.19× speedup | vs SGLang+HiCache | A100 | 未公开 | 部分 |
| Halo [ref:2509.02121] | Serving | >400× vs vLLM, 1.03–3.6× vs agent | vs vLLM/agent baselines | 未报告 | 部分 | 部分 |
| Justitia [ref:2510.17015] | Serving | 57.5% JCT reduction | vs VTC/Parrot | Llama 7B/13B | 未公开 | 不可 |
| GLM [ref:2511.01633] | Serving | 15.1× throughput, +38% accuracy | vs Graph-CoT baseline | 未报告 | 未公开 | 不可 |
| AIConfigurator [ref:2601.06288] | Serving | TPOT MAPE 6–12%, search 170K× faster | vs GPU benchmarking | 多种 | 未公开 | 不可 |
| Sutradhara [ref:2601.12967] | Serving | FTR −15% median | vs vanilla vLLM | A100 | 未公开 | 不可 |
| Concur [ref:2601.22705] | Serving | 4.09× throughput | vs SGLang | A100 | 未公开 | 不可 |
| ThunderAgent [ref:2602.13692] | Serving | 1.48–3.58× serving throughput | vs vLLM+K8s/Autellix | 未报告 | 开源 | 复现 |
| Pancake [ref:2602.21477] | Serving | E2E throughput 4.29× avg | vs baseline memory ops | 未报告 | 未公开 | 不可 |
| Saguaro/SSD [ref:2603.03251] | Serving | ~30% faster than best SD | vs standard speculative | A100 | 开源 (MIT) | 复现 |
| Helium [ref:2603.16104] | Serving | 1.56× vs KVFlow | vs KVFlow/vLLM | 未报告 | 开源 | 复现 |
| TokenDance [ref:2604.03143] | Serving | 2.7× concurrent agents | vs baseline | 未报告 | 未公开 | 不可 |
| Scepsy [ref:2604.15186] | Serving | 2.4× throughput | vs K8s autoscaler | 16 GPU | 开源 | 复现 |
| Fleet [ref:2604.15379] | Kernel/Serving | 1.3–1.56× decode speedup | vs vLLM on MI350 | MI350X | 未公开 | 不可 |
| MoE-LB [ref:moe-lb-interai25] | Serving | 12.5% MoE latency reduction | vs prior heuristics | 未报告 | 未公开 | 不可 |
| FlexRLHF [2312.11819] | Training | 11× throughput | vs DeepSpeed-Chat | 128×A100 | 未公开 | 不可 |
| HybridFlow/veRL [2409.19256] | Training | 1.53–20.57× throughput | vs DSChat/OpenRLHF/NeMo | 16–64×A100 | 开源 | 复现 |
| DeepSeek-V3 [2412.19437] | Training | bubble −60%+, $5.576M total | vs 1F1B | 2048×H800 | 部分 | 不可 |
| Semantic Router [vllm-project-semantic-router] | Routing | 4.2K stars, 20+ signal types | vs simple routers | — | 开源 | 复现 |
| TileRT blog [tilert-speed-scaling-law] | Serving | ~10× gap 弥合(理论) | vs traditional frameworks | 8×H200 | 部分 | 部分 |
| 系统 | Strength | Weakness | Best-for |
|---|---|---|---|
| DualPath | 聚合所有 SNIC 带宽,近线性扩展至 1152 GPU | 深度绑定 DeepSeek 内部栈,需双网隔离 | DeepSeek 级 agentic RL rollout |
| PPD | 零运行时路由开销,稳定性恢复不可用配置 | 仅 vLLM prototype,offline profiling 需一次性成本 | 多轮对话 PD serving |
| DynaServe | 统一 colocation/disaggregation,SLO-aware | 只支持两路分割,需 RDMA | 动态不平衡在线服务 |
| JITServe | 可证明竞争比,支持 compound SLO | QRF 冷启动风险,未开源 | 混合 SLO 多应用部署 |
| TileRT/TokenSpeed | BS=1 极致低延迟,生产验证 | 不支持 batching,硬件绑定 | Agent/TTS/real-time 场景 |
| ZeRO-Prefill | 消除 MoE EP 通信瓶颈,1 GPU 即可部署 235B | 仅 prefill-only,绑定 NVLink | 大规模判别式推理 |
| HybridFlow/veRL | 算法灵活 (PPO/ReMax/Safe-RLHF),zero-redundancy resharding | 仅同构 A100,auto-mapping 搜索慢 | RLHF 研究和生产训练 |
| TensorHub | 消除 trainer stall,线性扩展 pipeline replication | 未开源,绑定 Mooncake RDMA | 大规模 RL weight distribution |
| ThunderAgent | 程序级调度,lifecycle-aware eviction | 架构复杂度高 | Agentic batch rollout |
| Helium | DB-style plan optimization,Templated Radix Tree | 需要模板化 workflow | 结构化 agent 流水线 |
| KVServe | 模块化压缩 + service-aware controller,zero-fork vLLM | 不支持 MLA,依赖 nvCOMP | 跨 DC 低带宽 PD serving |
| DeepSeek-V3 DualPipe | all-to-all 完全隐藏,$5.576M 训 671B | 2× 参数内存,硬件高度耦合 | 超大 MoE 预训练 |
DynaServe [2504.09285] 的 micro-request 精确量化了此 trade-off: colocation 追求吞吐但 P99 TBT > 300ms,disaggregation 满足 SLO 但 MFU 低至 0.2%。DynaServe 的 Pareto 推进方式是在任意 token 边界动态分割请求。TileRT [tile-ai-tilert] 在 BS=1 走另一极端——彻底放弃 batching 换取 600 tok/s 单请求延迟。
DeepSeek-V3 DualPipe [2412.19437] 以 2× 参数内存换取 60%+ bubble 减少和 all-to-all 完全隐藏。DualPath [2602.21548] 以 80GB/node DRAM buffer 换取 storage I/O 瓶颈消除。ZeRO-Prefill [2605.02960] 以每 GPU 保留当前层全部 expert weights 换取消除 AllToAll。三者都遵循"以空间换时间隐藏"范式。
HybridFlow [2409.19256] 的 hybrid paradigm (single-controller 编排 + multi-controller 计算) vs FlexRLHF [2312.11819] 的静态 placement ratio;TokenSpeed [tokenspeed] 的 Placement 编译器 vs TileRT [tile-ai-tilert] 的 AOT 静态编译。更灵活的系统更容易适配新模型/算法但性能上限稍低。
DeepSeek (DualPath/DualPipe), ByteDance (TensorHub), Moonshot (PrfaaS) 都是内部系统,展示最佳数字但不可复现。vLLM 生态 (KVServe, ZeRO-Prefill, HexGen-Flow) 和 veRL 提供完整复现路径但性能通常落后闭源系统 10–30%。
PPD [2603.13358] 测量 append-prefill 仅造成 2% TPOT 劣化(vs full prefill 的 48%),从而论证 Turn 2+ 可以安全 co-locate。而 DualPath [2602.21548] 和 PrfaaS [2604.15039] 则假设 prefill 必须物理分离到不同节点。
矛盾根源: 规模和模型不同。PPD 在 Llama-3.1-8B + 4×H100 上测量 batch=200 的 append-prefill 干扰;DualPath 在 DeepSeek-660B + 千卡集群上面对 98.7% KV 命中率的 cache-compute ratio = 22 GB/PFLOP。小模型上 append-prefill 确实低干扰,但 MoE 超大模型的 I/O 密集 prefill 干扰模式完全不同——此时瓶颈不是计算干扰而是 storage I/O 带宽。两者在各自实验范围内都正确。
JITServe [2504.20068] 在 Appendix D 严格证明 SJF 和 EDF 对 goodput 无常数竞争比。但 DynaServe [2504.09285] 的 local scheduler 默认用 FCFS。
矛盾根源: 优化目标不同。JITServe 优化 SLO-aware goodput(per-request 级别),DynaServe 优化 serving capacity(系统级别,P99 TBT < 100ms 约束下的最大 QPS)。DynaServe 的 local scheduler 是 SLO-aware batch composition(不是简单 FCFS),其"FCFS"只描述请求入队顺序,实际 batch 组合由 profile table 动态决定。
ZeRO-Prefill [2605.02960] 论证传统 EP 的 AllToAll 是 prefill 瓶颈,用 AsyncEP(按 weight 聚集)完全消除。DeepSeek-V3 [2412.19437] 的 DualPipe 和 MFS [2603.17456] 的 RMLQ 则花大量精力优化和调度 AllToAll。
矛盾根源: Workload 差异。ZeRO-Prefill 仅面向 prefill-only(大 batch、长 compute window),此时 weight AllGather 可被完全隐藏。DeepSeek-V3 的 DualPipe 面向训练(forward+backward 双向),MFS 面向 online serving(低延迟约束)。在 decode 或训练时,per-token compute window 太短无法隐藏 weight AllGather——AllToAll 仍是更优选择。ZeRO-Prefill 自己也承认对 decode 不适用。
基于 taxonomy 的维度组合分析,以下空白点在技术上可解但目前无人做:
| 场景 | 推荐系统 | 原因 |
|---|---|---|
| DeepSeek 级 MoE 推理 (百卡 RDMA 集群) | DualPath [2602.21548] 的思路 + SGLang EP | DualPath 的 dual-path loading 和 CNIC-centric design 是当前最优,但需要自研。替代方案用 SGLang + Mooncake distributed KV |
| 多轮对话低延迟 serving (几十卡) | PPD [2603.13358] 路由策略 + vLLM PD | Turn 2+ append-prefill 干扰极低,可安全 co-locate 于 decode 节点,减少 75% 网络传输 |
| Agentic batch inference (offline RL rollout) | ThunderAgent [ref:2602.13692] 或 Concur [ref:2601.22705] + SGLang | ThunderAgent 的 program-aware 调度避免 KV thrashing,Concur 的 AIMD admission control 在高并发下更稳健 |
| 单请求超低延迟 (AI 编程/TTS/agent 链) | TileRT [tile-ai-tilert] (GLM-5) 或 TokenSpeed [tokenspeed] (Kimi K2.5) | BS=1 tile-level 调度消除所有 kernel 间 idle,600 tok/s 是当前最快 |
| MoE prefill-only 判别式推理 | ZeRO-Prefill [2605.02960] | AsyncEP 消除 AllToAll、支持单卡部署 235B,vLLM v0.11 单 flag 启用 |
| RLHF/GRPO 训练 (研究团队) | veRL (HybridFlow) [2409.19256] | 开源、算法灵活、auto-mapping、EuroSys 验证 |
| 跨 DC 异构推理 (H200 prefill + H20 decode) | PrfaaS [2604.15039] 思路 | hybrid attention 模型的 KV 吞吐降 4–13×,跨 DC Ethernet 可行 |
| 结构化 multi-agent workflow | Helium [ref:2603.16104] + KVFlow [ref:2507.07400] | Helium 做 plan-level 优化,KVFlow 做 cache 优化,两者互补 |
| 混合 SLO 生产集群 | JITServe [2504.20068] 中间件 + vLLM | GMAX 提供可证明 goodput 保证,仅需几行 API 修改 |
| 跨框架配置优化 | AIConfigurator [ref:2601.06288] 思路 | 不跑 GPU 即可预测 TPOT 并搜索最优并行配置 |
| ID | Title | Date | Venue/Status |
|---|---|---|---|
| 2602.21548 | DualPath: Breaking the Storage Bandwidth Bottleneck in Agentic LLM Inference | 2026-02 | arXiv |
| 2603.13358 | Not All Prefills Are Equal: PPD Disaggregation for Multi-turn LLM Serving | 2025-03 | arXiv |
| 2603.17456 | MFS: Multi-stage Flow Scheduling for LLM Serving | 2026-03 | arXiv |
| 2604.09107 | TensorHub: Scalable and Elastic Weight Transfer for LLM RL Training | 2026-04 | arXiv |
| 2604.15039 | PrfaaS: KVCache of Next-Generation Models Could Go Cross-Datacenter | 2026-04 | arXiv |
| 2605.02960 | ZeRO-Prefill: Zero Redundancy Overheads in MoE Prefill Serving | 2026-05 | arXiv |
| tilert-speed-scaling-law | Speed as the Next Scaling Law (TileRT blog) | 2026-05 | Blog |
| kvserve | KVServe: Service-aware KV-cache Compression | 2026-05 | GitHub |
| tile-ai-tilert | TileRT: Tile-Based Runtime for Ultra-Low-Latency LLM Inference | 2026-05 | GitHub |
| tokenspeed | TokenSpeed: Speed-of-Light LLM Inference Engine | 2026-05 | GitHub |
| vllm-project-semantic-router | vLLM Semantic Router | 2026-05 | GitHub |
| 2312.11819 | FlexRLHF: Adaptive Placement and Parallelism for RLHF Training | 2023-12 | arXiv |
| 2409.19256 | HybridFlow (veRL): A Flexible and Efficient RLHF Framework | 2024-09 | EuroSys'25 |
| 2412.19437 | DeepSeek-V3 Technical Report | 2024-12 | arXiv |
| 2504.09285 | DynaServe: Unified and Elastic Execution for LLM Serving | 2025-04 | arXiv |
| 2504.20068 | JITServe: SLO-aware LLM Serving with Imprecise Request Information | 2025-04 | arXiv |
| 2505.05286 | HEXGEN-FLOW: Optimizing Scheduling for Agentic Text-to-SQL | 2025-05 | arXiv |
| 2507.07400 | KVFlow: Efficient Prefix Caching for Multi-Agent Workflows | 2025-07 | arXiv |
| 2509.02121 | Halo: Batch Query Processing for Agentic Workflows | 2025-09 | arXiv |
| 2510.17015 | Justitia: Fair and Efficient Scheduling for LLM Applications | 2025-10 | arXiv |
| 2511.01633 | GLM: Graph Chain-of-Thought with Efficient LLM Serving | 2025-11 | PVLDB'26 |
| 2601.06288 | AIConfigurator: Lightning-Fast Configuration Optimization | 2026-01 | arXiv |
| 2601.12967 | Sutradhara: Orchestrator-Engine Co-design for Agentic Inference | 2026-01 | arXiv |
| 2601.22705 | Concur: Proactive Agent-Level Admission Control | 2026-01 | arXiv |
| 2602.13692 | ThunderAgent: Program-Aware Agentic Inference System | 2026-02 | arXiv |
| 2602.21477 | Pancake: Hierarchical Memory System for Multi-Agent Serving | 2026-02 | arXiv |
| 2603.03251 | Saguaro: Speculative Speculative Decoding | 2026-03 | arXiv |
| 2603.16104 | Helium: Efficient LLM Serving for Agentic Workflows | 2026-03 | arXiv |
| 2604.03143 | TokenDance: Scaling Multi-Agent via Collective KV Cache Sharing | 2026-04 | arXiv |
| 2604.15186 | Scepsy: Serving Agentic Workflows Using Aggregate LLM Pipelines | 2026-04 | arXiv |
| 2604.15379 | Fleet: Hierarchical Task-based Megakernels on Multi-Die GPUs | 2026-04 | arXiv |
| moe-lb-interai25 | Latency-Optimal Load Balancing for Distributed MoE Inference | 2025 | InterAI'25 Workshop |