Xingda Wei

Yifan Peng, Jiafei Fan, Xingda Wei et al.

Learned top-K search is a promising approach for serving vector queries with both high accuracy and performance. However, current models trained for a specific K value fail to generalize to real-world multi-K queries: they suffer from accuracy degradation (for larger Ks) and performance loss (for smaller Ks). Training the model to generalize on different Ks requires orders of magnitude more preprocessing time and is not suitable for serving vector queries in the wild. We present OMEGA, a K-generalizable learned top-K search method that simultaneously achieves high accuracy, high performance, and low preprocessing cost for multi-K vector queries. The key idea is that a base model properly trained on K=1 with our trajectory-based features can be used to accurately predict larger Ks with a dynamic refinement procedure and smaller Ks with minimal performance loss. To make our refinements efficient, we further leverage the statistical properties of top-K searches to reduce excessive model invocations. Extensive evaluations on multiple public and production datasets show that, under the same preprocessing budgets, OMEGA achieves 6-33% lower average latency compared to state-of-the-art learned search methods, while all systems achieve the same recall target. With only 16-30% of the preprocessing time, OMEGA attains 1.01-1.28x of the optimal average latency of these baselines.

Dingyan Zhang, Jinbo Han, Kaixi Zhang et al.

High-quality LLM request scheduling requires achieving two key objectives: whether the routed instance has KV$ to accelerate the request execution and whether the workload is balanced across instances. Achieving both objectives is challenging because pursuing one objective may compromise the other. Current approaches adopt various combinators (e.g., linear combinations) to compute a scheduling score combining indicators for the two objectives, which are complex in that they either require significant workload-specific hyperparameter tuning or model-hardware-aware simulator development, and could still lead to suboptimal performance. In this paper, we show that using a simple multiplication of two carefully chosen indicators-one for KV$-aware (new prefill tokens if routed to an instance) and one for load balancing-aware (current batch size of the instance)-as the scheduling score can simultaneously achieve both objectives well without any hyperparameter tuning. The key idea is that the multiplied score considers both objectives in a manner similar to a linear combination, with a nice property that the original hyperparameters are canceled out during comparison so we don't need tuning to find the best parameters. The two indicators are chosen based on our analysis of LLM characteristics, and our extensive experiments show that this simple approach can reduce TTFT by 92% and 52%, and TPOT by 21% and 20%, compared to vLLM-v1 and a production scheduler on real-world workloads covering chatbots, API calls, and coding agents. We also mathematically derive the conditions under which multiplication may fail, and find that such conditions are extremely rare in practice and can be detected (and mitigated) beforehand.

Jiahao Wang, Kaizhan Lin, Kaixi Zhang et al.

LLM scheduling is critical to serving, yet it remains unclear how well existing designs fit agentic serving--with LLM requests issued by agents instead of humans. This shifts the workload in two ways: (1) agents act only on complete responses, making the cluster's tokens per second (TPS) the primary goal and relaxing--not eliminating--per-token latency requirements; and (2) requests share much of their KV\$-reuse exceeds 80% of request tokens in a production trace from BAILIAN, versus 54-62% in chat. This paper first contributes a systematic study of request scheduling for agents on two real-world traces. We find that to increase KV\$ reuse, existing schedulers overly prioritize routing requests to instances caching their KV\$, overloading a few while leaving the rest idle, capping TPS. We thus present two key insights: (1) load balance need not sacrifice all KV\$ reuse, thanks to the global-tier KV\$ store and (2) by utilizing the workload's intra-session locality, balancing a small fraction of requests--the first request in each agent session--suffices to balance the cluster without sacrificing most KV\$ reuse on local instances. SMETRIC realizes these insights with balanced session-centric scheduling: it routes each session's first request purely for load balance and its follow-up requests in a cache-aware manner, preserving load balance and local reuse while keeping demand on the global tier low. Using the session turn information as the scheduling metric is deliberate: it is derived efficiently and accurately from the user inputs alone, so the scheduler stays clean and stateless. SMETRIC improves cluster TPS by 10-16% under prefill-decode colocation with a global store and prefill TPS by 2-34% under disaggregation over state-of-the-art schedulers, also with a better per-token latency.

Jiahao Wang, Jinbo Han, Xingda Wei et al.

Serving large language models (LLMs) is important for cloud providers, and caching intermediate results (KV\$) after processing each request substantially improves serving throughput and latency. However, there is limited understanding of how LLM serving benefits from KV\$ caching, where system design decisions like cache eviction policies are highly workload-dependent. In this paper, we present the first systematic characterization of the KV\$ workload patterns from one of the leading LLM service providers. We draw observations that were not covered by previous studies focusing on synthetic workloads, including: KV\$ reuses are skewed across requests, where reuses between single-turn requests are equally important as multi-turn requests; the reuse time and probability are diverse considering all requests, but for a specific request category, the pattern tends to be predictable; and the overall cache size required for an ideal cache hit ratio is moderate. Based on the characterization, we further propose a workload-aware cache eviction policy that improves the serving performance under real-world traces, especially with limited cache capacity.

Hongtao Lyu, Dingyan Zhang, Mingyu Wu et al.

Multi-agent LLM systems -- coding agents, devops agents, document agents -- now routinely run several agents in parallel against the same git tree, Kubernetes cluster, or document. As soon as two of them mutate shared state, they enter the regime classical concurrency control has studied for decades, but classical mechanisms fit LLM agents poorly. A single agent transaction spans minutes of inference, read sets are broad and opaque rather than statically inferable, and the live state agents act on admits neither fork nor buffer, so writes take effect the moment they execute. Locks block long inference intervals; OCC abort-and-retry discards minutes of work on every conflict. This paper builds concurrency control on a capability classical transactions lack: the LLM inside each agent can judge whether a conflicting write invalidates its plan, and can repair exactly the operations that depended on it. Control therefore turns advisory: the runtime informs, the agent repairs. Our protocol, MTPO (Monotonic Trajectory Pre-Order), fixes a serialization order at launch, serves each read the order-filtered value, and applies writes speculatively in place; a one-way notification asks an affected reader to re-judge and patch its plan, while the framework mechanically undoes and reorders misplaced writes through the saga-style inverse each tool registers in advance. At quiescence the run is serializable in the pre-decided order. We realize MTPO as CoAgent, toolcall middleware whose privileged ToolSmith grows footprint-declared, undoable tools online. On ten contended workloads, CoAgent stays within 5\% of serial correctness at a $1.4\times$ speedup and near-serial token cost, where 2PL and OCC surrender nearly all concurrency gains; on a bash-only target system, it grows a 25-tool library online and lifts the task pass rate from 45/71 to 63/71 at $0.80\times$ the time and $0.86\times$ the cost.

Rongxin Cheng, Yuxin Lai, Xingda Wei et al.

Serving LLMs with a cluster of GPUs is common nowadays, where the serving system must meet strict latency SLOs required by applications. However, the stateful nature of LLM serving requires maintaining huge states (i.e., KVCache) in limited GPU memory. Under spikes in real-world workloads, GPU memory can be easily throttled, leading to orders of magnitude higher response latency due to queuing introduced by waiting for KVCache to be reclaimed. Prior KVCache-centric approaches handle load throttling by dropping, migrating, or swapping KVCache. These methods fail to release sufficient memory quickly with requests still queued. This paper proposes the first parameter-centric approach to handling throttling by selectively dropping replicated parameters to instantly free memory for requests, based on an unnoticed observation that model parameters are commonly replicated across GPUs for serving LLMs. With additional memory, all requests can be served with a larger batch without queuing. To make the parameter-centric approach correct and efficient, we cooperatively execute requests on GPUs with a complete copy of parameters using pipeline parallelism, and derive an appropriate drop plan without unnecessary cooperation. We also design techniques to minimize the performance overhead due to pipeline parallelism with the execution patterns of requests under drop. Evaluations show that {\sys} reduces the tail TTFT of requests under throttling by up to 72.2 times compared to the state-of-the-art systems including Llumnix, vLLM and InferCept.

Rongxin Cheng, Kai Zhou, Xingda Wei et al.

Rollout dominates the training time in large language model (LLM) post-training, where the trained model is used to generate tokens given a batch of prompts. SpecActor achieves fast rollout with speculative decoding that deploys a fast path (e.g., a smaller model) to accelerate the unparallelizable generation, while the correctness is guaranteed by fast parallel verification of the outputs with the original model. SpecActor addresses two foundational challenges in speculative rollout by (1) a \emph{dynamic decoupled speculation} execution method that maximizes the GPU computational efficiency to realize speedup for large-batch execution -- a configuration common in training but unfriendly to speculative execution and (2) a \emph{dynamic Best-of-N speculation} method that selects and combines different drafting methods according to the rollout progress. It substantially improves the speculation accuracy even when the best drafting method is unknown a priori, meanwhile without requiring adding extra computation resources. {\sys} is {1.3--1.7}\,$\times$ faster than common post-training baselines, and is {1.3--1.5}\,$\times$ faster compared to naively adopting speculative decoding for rollout.