arXiv subjects

Tianyuan Wu

Publications and source records attributed to Tianyuan Wu.

14 recordsLinked to original sources

arXiv2026

KREX: Concurrent Kernel Benchmarking on Shared GPUs via Region-Granular Exclusivity

LLM agents automate GPU kernel optimization by repeatedly composing candidates and measuring their duration on real GPUs. Existing systems preserve measurement fidelity by reserving a GPU for an entire agent session or benchmarking command. However, this results in poor utilization because only a small fraction of command execution requires exclusive GPU access. Sharing GPUs could recover this idle capacity, but introduces contention that compromises measurement fidelity and misdirects the agent's search. We present KREX, a runtime for concurrent kernel agent benchmarking with region-granular exclusivity. KREX lets agents mark critical regions involving timing-sensitive operations within a benchmarking command. The runtime then enforces exclusivity within marked regions and allows concurrent execution outside them, achieving high throughput while preserving measurement fidelity. To enforce in-region exclusivity, KREX blocks new competing GPU submissions and drains outstanding work before freezing sibling processes and isolating CPU cores, protecting both GPU execution and the host threads that drive measurements. To maximize off-region concurrency, KREX reuses GPU contexts in persistent context processes to avoid repeated, node-wide serialized context creation. We evaluate KREX on NVIDIA and AMD GPUs. Compared with command-granular exclusivity baselines, KREX delivers up to $3.4\times$ the benchmarking throughput with a negligible p95 timing inflation of $0.30\%$, $1.58\%$, and $3.90\%$ for kernels longer than 10 ms, 1 ms, and 0.1 ms, respectively.

cs.DC↗

arXiv2026

AgentPProf: Semantic Profiler for Long Horizon AI Agents

AI agents increasingly orchestrate long-running activities with users, tools, and system resources for days and weeks. To improve agent quality, safety, and cost efficiency, developers need to determine where failures happen, what triggers unsafe effects, and which tasks consume the most budget, then optimize those tasks. In systems software, profiling answers similar questions by aggregating resource consumption and attributing it to responsible code paths to identify hotspots. Yet existing agent observability tools focus on per-execution debugging and tracing rather than cross-run, long term profiling, making these questions difficult to answer at scale. Agent observability needs profiling, not only debugging, but profiling agents is challenging: the responsible entities are task intent like diagnose authentication, compare branches rather than code paths, and lack stable identifiers for aggregation. We propose a semantic operation stack model that adapts profiling to agent trajectories. Uniform operations represent all activities, and operation stacks replace the runtime call stack, enabling hierarchical attribution at different granularities. We observe that an agent's task occupies a contiguous span and decomposes into subtasks, so we introduce recursive operation segmentation, which recursively splits trajectories at task boundaries. AgentPProf is a profiler that aggregates agent trajectories into pprof-compatible profiles, enabling flame graph visualization and analysis. AgentPProf reaches 0.764 $B^3$ F1 against human annotations on CodeTraceBench. On three problem-localization benchmarks, the profile raises MAP by up to 56%, demonstrating that it effectively attributes resources, locates problems, and helps optimize token cost at practical profiling cost. AgentPProf is available at https://github.com/eunomia-bpf/agentsight.

cs.AI↗

arXiv2026

ActPlane: Programmable OS-Level Policy Enforcement for Agent Harnesses

AI agents increasingly run in production through harnesses, the software around the LLM, including an engine that enforces safety and effectiveness policies, e.g., 'run tests before committing.' Enforcing these policies requires bridging a semantic gap: policy intent is expressed in underspecified natural language, while enforcement must act on concrete system actions, e.g., which test to run. Many policies also define event ordering or data flow actions. Yet existing approaches fall short. Tool-call guardrails miss system actions that bypass the tool layer, while OS sandboxes control resource access instead of actions, returning opaque errors that confuse the agent. Our key insight is that policy context lives within the agent closest to the task, while enforcement must happen at the OS to cover all execution paths. We introduce ActPlane, a policy engine that lets agents declare policies and enforces them in the OS kernel with semantic feedback and isolation. ActPlane uses a simple information-flow control (IFC) DSL to support cross-event policies. We implement ActPlane with eBPF and evaluate it on policies from the empirical study, coding-task benchmarks, and safety benchmarks. ActPlane improves policy compliance, including on indirect execution paths that tool-call interception cannot observe, with 1.9%-8.4% overhead. ActPlane is at https://github.com/eunomia-bpf/ActPlane

cs.OS↗

arXiv2026

RollArt: Disaggregated Multi-Task Agentic RL Training at Scale

Agentic Reinforcement Learning (RL) trains LLMs through multi-turn interactions with environments, producing workloads that mix compute-bound prefill, bandwidth-bound decoding, CPU-heavy environment execution, and bursty reward evaluation. Existing systems either colocate all stages on a single GPU cluster or decouple them only at a coarse granularity, overlooking hardware heterogeneity and incurring substantial synchronization overhead across stages. We present ROLLART, a system for multi-task agentic RL on disaggregated infrastructure. ROLLART maps each pipeline stage to best-fit hardware, routing prefill-heavy tasks to compute-optimized GPUs, decode-heavy tasks to bandwidth-optimized GPUs, and environments to CPU clusters. It decouples rollout at the trajectory level, allowing generation, environment interaction, and reward scoring to proceed independently, so that slow or failed environments never block the others. ROLLART offloads stateless reward computation to serverless infrastructure and overlaps rollout with training via staleness-bounded asynchronous weight synchronization. Our results demonstrate that ROLLART effectively improves training throughput and achieves 1.31--2.05 $\times$ training time reduction compared to various RL systems. We also evaluated ROLLART by training a hundreds-of-billions-parameter MoE model for Qoder product on an Alibaba cluster with above 3,000 GPUs, demonstrating its stability and scalability.

cs.DC↗

arXiv2026

ROSE: Rollout On Serving GPUs via Cooperative Elasticity for Agentic RL

Agentic reinforcement learning (RL) is reshaping LLM post-training, but end-to-end training time is dominated by compute-intensive, multi-turn rollouts whose resource demand varies significantly across training steps. Resource-fixed systems cannot adapt to this variation, while resource-elastic approaches that provision external GPUs on demand suffer from high allocation overhead and limited availability. We observe that serving clusters leave substantial GPU compute and memory idle, and propose cooperative elasticity: sharing already-deployed serving GPUs with rollout workloads to provide on-demand elastic capacity. Realizing this is non-trivial, as it must preserve serving SLOs under bursty traffic while minimizing cross-cluster communication overhead. We present ROSE, a system that realizes cooperative elasticity for agentic RL post-training, comprising three components: (1) an SLO-safe co-serving executor that co-locates heterogeneous serving and rollout models on the same GPUs, dynamically sharing memory and compute while preserving serving SLOs; (2) a cross-cluster weight transfer engine that leverages shard-aware routing and weight sparsity for fast synchronization; and (3) an elastic rollout scheduler that dynamically routes rollouts across dedicated and opportunistic serving GPUs. Experiments across multiple model sizes and cluster scales show that ROSE improves end-to-end throughput by 1.3 - 3.3 x over resource-fixed baselines and reduces rollout time by 1.2 - 1.5 x over resource-elastic baselines, with no serving SLO violations.

cs.DC↗

arXiv2026

Crab: A Semantics-Aware Checkpoint/Restore Runtime for Agent Sandboxes

Autonomous agents act through sandboxed containers and microVMs whose state spans filesystems, processes, and runtime artifacts. Checkpoint and restore (C/R) of this state is needed for fault tolerance, spot execution, RL rollout branching, and safe rollback-yet existing approaches fall into two extremes: application-level recovery preserves chat history but misses OS-side effects, while full per-turn checkpointing is correct but too expensive under dense co-location. The root cause is an agent-OS semantic gap: agent frameworks see tool calls but not their OS effects; the OS sees state changes but lacks turn-level context to judge recovery relevance. This gap hides massive sparsity: over 75% of agent turns produce no recovery-relevant state, so most checkpoints are unnecessary. Crab (Checkpoint-and-Restore for Agent SandBoxes) is a transparent host-side runtime that bridges this gap without modifying agents or C/R backends. An eBPF-based inspector classifies each turn's OS-visible effects to decide checkpoint granularity; a coordinator aligns checkpoints with turn boundaries and overlaps C/R with LLM wait time; and a host-scoped engine schedules checkpoint traffic across co-located sandboxes. On shell-intensive and code-repair workloads, Crab raises recovery correctness from 8% (chat-only) to 100%, cuts checkpoint traffic by up to 87%, and stays within 1.9% of fault-free execution time.

cs.OS↗

arXiv2025

RollMux: Phase-Level Multiplexing for Disaggregated RL Post-Training

Rollout-training disaggregation is emerging as the standard architecture for Reinforcement Learning (RL) post-training, where memory-bound rollout and compute-bound training are physically disaggregated onto purpose-built clusters to maximize hardware efficiency. However, the strict synchronization required by on-policy algorithms introduces severe dependency bubbles, forcing one cluster to idle while the dependent phase is running on the other. We present RollMux, a cluster scheduling framework that reclaims these bubbles through cross-cluster orchestration. RollMux is built on the insight that the structural idleness of one job can be effectively utilized by the active phase of another. To realize this, we introduce the co-execution group abstraction, which partitions the cluster into isolated locality domains. This abstraction enables a two-tier scheduling architecture: an inter-group scheduler that optimizes job placement using conservative stochastic planning, and an intra-group scheduler that orchestrates a provably optimal round-robin schedule. The group abstraction also imposes a residency constraint, ensuring that massive model states remain cached in host memory to enable "warm-star" context switching. We evaluate RollMux on a production-scale testbed with 328 H20 and 328 H800 GPUs. RollMux improves cost efficiency by 1.84x over standard disaggregation and 1.38x over state-of-the-art co-located baselines, all while achieving 100% SLO attainment.

cs.DC↗

arXiv2025

AReaL-Hex: Accommodating Asynchronous RL Training over Heterogeneous GPUs

Maximizing training throughput and cost-efficiency of RL for LLMs is essential to democratize this advanced technique. One promising but challenging approach is to deploy such a computational workflow over heterogeneous GPUs. Unlike conventional large-scale LLM pretraining, RL training generally decomposes into three coupled stages, i.e., rollout generation, reward computation, and policy/value updates, which exhibit markedly different compute intensities, memory footprints, and communication patterns. Recent research shows that fully asynchronous RL training can disaggregate these stages across disjoint hardware pools without sacrificing training stability, creating a great opportunity for real-world heterogeneous deployment. To this end, we present AReaL-Hex, a heterogeneity-aware asynchronous RL training system that effectively schedules how to execute rollout generation and policy model training over heterogeneous GPUs while enforcing data staleness bounds. Concretely, we use a two-phase scheduler: (i) a constrained search with MILP to select per-stage parallelization strategies and workload assignments given a resource budget, and (ii) a graph-partitioning step that allocates heterogeneous GPUs and interconnects to maximize end-to-end throughput. Built atop a fully asynchronous RL architecture, AReaL-Hex maps HBM-I/O-bound generation and compute-bound optimization to more cost-efficient resources and balances their producer-consumer interactions to avoid both idleness and stale rollout trajectories. On the mathematical reasoning task with various model scales (1.5B, 7B, and 14B), compared to homogeneous deployments of state-of-the-art asynchronous RL systems: (i) When maintaining the same total budgets, AReaL-Hex delivers up to 1.50x higher training throughput; (ii) When achieving the same training throughput, AReaL-Hex results in up to 1.46x reduction in training cost.

cs.DC↗

arXiv2025

RollPacker: Mitigating Long-Tail Rollouts for Fast, Synchronous RL Post-Training

Reinforcement Learning (RL) is a pivotal post-training technique for enhancing the reasoning capabilities of Large Language Models (LLMs). However, synchronous RL post-training often suffers from significant GPU underutilization, referred to as bubbles, caused by imbalanced response lengths within rollout steps. Many RL systems attempt to alleviate this problem by relaxing synchronization, but this can compromise training accuracy. In this paper, we introduce tail batching, a novel rollout scheduling strategy for synchronous RL that systematically consolidates prompts leading to long-tail responses into a small subset of rollout steps (long rounds), while ensuring that the majority of steps (short rounds) involve only balanced, short rollouts. By excluding long responses from short rounds and rescheduling them into a few designated long rounds, tail batching effectively reduces GPU idle time during rollouts and significantly accelerates RL training without sacrificing accuracy. We present RollPacker, a system that fully harnesses the benefits of tail batching through holistic optimizations across all three RL stages: elastic parallelism adaptation for rollout, dynamic resource allocation and scheduling for reward, and stream-based training. Empirical results show that RollPacker achieves a 2.03x-2.56x end-to-end training time reduction compared to veRL and up to 2.24x speedup compared to RLHFuse for the Qwen2.5 family of LLMs on up to 128 H800 GPUs.

cs.DC↗

arXiv2025

Reinforcement Learning Optimization for Large-Scale Learning: An Efficient and User-Friendly Scaling Library

We introduce ROLL, an efficient, scalable, and user-friendly library designed for Reinforcement Learning Optimization for Large-scale Learning. ROLL caters to three primary user groups: tech pioneers aiming for cost-effective, fault-tolerant large-scale training, developers requiring flexible control over training workflows, and researchers seeking agile experimentation. ROLL is built upon several key modules to serve these user groups effectively. First, a single-controller architecture combined with an abstraction of the parallel worker simplifies the development of the training pipeline. Second, the parallel strategy and data transfer modules enable efficient and scalable training. Third, the rollout scheduler offers fine-grained management of each sample's lifecycle during the rollout stage. Fourth, the environment worker and reward worker support rapid and flexible experimentation with agentic RL algorithms and reward designs. Finally, AutoDeviceMapping allows users to assign resources to different models flexibly across various stages.

cs.LG↗

arXiv2025

Adaptra: Straggler-Resilient Hybrid-Parallel Training with Pipeline Adaptation

Training large Deep Neural Network (DNN) models at scale often encounters straggler issues, mostly in communications due to network congestion, RNIC/switch defects, or topological asymmetry. Under advanced pipeline parallelism, even minor communication delays can induce significant training slowdowns. This occurs because (1) slow communication disrupts the pipeline schedule, creating cascading "bubbles" in a domino effect, and (2) current GPU kernel scheduling is susceptible to head-of-line blocking, where slow communication blocks subsequent computations, further adding to these bubbles. To address these challenges, we present ADAPTRA, a straggler-resilient training system with two key optimizations. First, it optimally adapts the pipeline schedule in the presence of stragglers to absorb communication delays without inducing cascading bubbles, using a simple yet effective algorithm guided by an analytical model. Second, upon detecting slow communication, ADAPTRA offloads communication operations from GPU to host memory and utilizes CPU-side RDMA for data transfer. This eliminates head-of-line blocking as subsequent computation kernels can be scheduled immediately on GPUs. Together, these optimizations effectively reduce pipeline stalls in the presence of communication stragglers, improving the training iteration time by 1.2-3.5x in our experiments under various settings.

cs.DC↗

arXiv2024

Echo: Simulating Distributed Training At Scale

Simulation offers unique values for both enumeration and extrapolation purposes, and is becoming increasingly important for managing the massive machine learning (ML) clusters and large-scale distributed training jobs. In this paper, we build Echo to tackle three key challenges in large-scale training simulation: (1) tracing the runtime training workloads at each device in an ex-situ fashion so we can use a single device to obtain the actual execution graphs of 1K-GPU training, (2) accurately estimating the collective communication without high overheads of discrete-event based network simulation, and (3) accounting for the interference-induced computation slowdown from overlapping communication and computation kernels on the same device. Echo delivers on average 8% error in training step -- roughly 3x lower than state-of-the-art simulators -- for GPT-175B on a 96-GPU H800 cluster with 3D parallelism on Megatron-LM under 2 minutes.

cs.LG↗

arXiv2024

FALCON: Pinpointing and Mitigating Stragglers for Large-Scale Hybrid-Parallel Training

Fail-slows, or stragglers, are common but largely unheeded problems in large-scale hybrid-parallel training that spans thousands of GPU servers and runs for weeks to months. Yet, these problems are not well studied, nor can they be quickly detected and effectively mitigated. In this paper, we first present a characterization study on a shared production cluster with over 10,000 GPUs1. We find that fail-slows are caused by various CPU/GPU computation and cross-node networking issues, lasting from tens of seconds to nearly ten hours, and collectively delaying the average job completion time by 1.34%. The current practice is to manually detect these fail-slows and simply treat them as fail-stops using a checkpoint-and-restart failover approach, which are labor-intensive and time-consuming. In this paper, we propose FALCON, a framework that rapidly identifies fail-slowed GPUs and/or communication links, and effectively tackles them with a novel multi-level mitigation mechanism, all without human intervention. We have applied FALCON to detect human-labeled fail-slows in a production cluster with over 99% accuracy. Cluster deployment further demonstrates that FALCON effectively handles manually injected fail-slows, mitigating the training slowdown by 60.1%.

cs.DC↗

arXiv2024

CaraServe: CPU-Assisted and Rank-Aware LoRA Serving for Generative LLM Inference

Pre-trained large language models (LLMs) often need specialization for domain-specific tasks. Low-Rank Adaptation (LoRA) is a popular approach that adapts a base model to multiple tasks by adding lightweight trainable adapters. In this paper, we present CaraServe, a system that efficiently serves many LoRA adapters derived from a common base model. CaraServe maintains the base model on GPUs and dynamically loads activated LoRA adapters from main memory. As GPU loading results in a cold-start that substantially delays token generation, CaraServe employs a CPU-assisted approach. It early starts the activated adapters on CPUs for prefilling as they are being loaded onto GPUs; after loading completes, it then switches to the GPUs for generative LoRA inference. CaraServe develops a highly optimized synchronization mechanism to efficiently coordinate LoRA computation on the CPU and GPU. Moreover, CaraServe employs a rank-aware scheduling algorithm to optimally schedule heterogeneous LoRA requests for maximum service-level objective (SLO) attainment. We have implemented CaraServe and evaluated it against state-of-the-art LoRA serving systems. Our results demonstrate that CaraServe can speed up the average request serving latency by up to 1.4$\times$ and achieve an SLO attainment of up to 99%.

cs.DC↗