X Liu

Huamin Chen, Xunzhuo Liu, Yuhan Liu et al.

Sizing a GPU fleet for LLM inference is harder than it looks. The obvious questions -- how many GPUs, which type, where to split a two-pool fleet -- have no closed-form answers. They depend on the full token-length distribution, the routing policy, and queueing dynamics that turn ugly under heavy-tailed workloads. Existing tools optimize per-engine configuration for a fixed GPU count; none of them address the upstream question of how many GPUs to buy and how to arrange them. inference-fleet-sim fills that gap. It combines analytical M/G/c queueing with discrete-event simulation (DES) to find the minimum-cost fleet configuration that empirically meets a P99 TTFT SLO. It includes a physics-informed GPU performance model covering A10G, A100, and H100 across monolithic, two-pool-routed, and disaggregated topologies, all without requiring access to real hardware. We run the tool on seven fleet-planning scenarios drawn from two public workload traces (LMSYS, Azure) and one synthetic agent-heavy trace. Each one surfaces a result that simple analysis gets wrong -- the right split threshold, the cheapest GPU type, whether an apparently idle fleet is actually broken -- and shows why joint simulation of queueing, routing, and hardware is necessary to find it.

Huamin Chen, Xunzhuo Liu, Yuhan Liu et al.

Modern LLM GPU fleets are provisioned for worst-case context lengths that the vast majority of requests never approach, wasting GPU capacity on idle KV-cache slots. We present FleetOpt, a framework that starts from first principles: given a workload's prompt-length CDF and a P99 TTFT target, derive the minimum-cost fleet analytically, then deploy it in practice. The analytical core models each pool as an M/G/c queue and derives that the minimum-cost fleet is a two-pool architecture -- a short-context pool and a long-context pool -- with an optimal boundary B* satisfying an equal marginal GPU cost condition across both pools. The fundamental barrier to achieving B* is the cost cliff: a hard routing step where requests just above B* consume 8x--42x more GPU capacity than requests just below it (depending on the context window ratio), creating a structural disincentive to lower the boundary. Compress-and-Route (C&R) is the implementation mechanism that resolves this barrier. Gateway-layer extractive compression trims borderline requests below B* before the engine ever sees them, converting the hard hardware boundary into a software parameter read from the workload CDF. The two components are unified in the FleetOpt offline planner: given a CDF and SLO, it returns the optimal (n_s*, n_l*, B*, gamma*) in under 1 ms. On three production traces, the combined framework reduces total GPU cost by 6--82% versus a homogeneous fleet, with C&R contributing 1--44 percentage points beyond plain pool routing depending on workload archetype. The analytical model is validated against a discrete-event simulator (inference-fleet-sim) with <= 3% error on predicted GPU utilization across all pools and workloads.

Huamin Chen, Xunzhuo Liu, Yuhan Liu et al.

How many tokens can a GPU inference cluster deliver per watt? Across deployments of identical hardware, the answer varies by 40x -- not because of software inefficiency, but because of the serving context window. We derive the 1/W law: tokens per watt halves every time the context window doubles. A larger context window shrinks the KV-cache concurrency limit while leaving GPU power draw roughly unchanged. At 64K context, an H100 holds 16 sequences in flight (tok/W = 1.5); at 4K context, the same H100 holds 256 sequences (tok/W = 17.6). Routing topology -- which determines the effective context window each GPU services -- is a more powerful energy lever than buying newer hardware. Working from published H100 power measurements, a calibrated logistic power model, and a roofline throughput model, we derive these results analytically using the inference-fleet-sim framework; no new hardware experiments were conducted. Two-pool context-length routing (FleetOpt) delivers roughly 2.5x better tok/W over a homogeneous fleet, while upgrading from H100 to B200 delivers roughly 1.7x. The gains are independent: combining FleetOpt with B200 yields 4.25x over the H100 homogeneous baseline. B200/H200 numbers are analytical projections (+-20% uncertainty); H100 results are calibrated to published measurements. For MoE models, active-parameter weight streaming adds a third lever. Qwen3-235B-A22B (22B active) reaches roughly 37.8 tok/W at 8K context on H100 -- 5.1x better than Llama-3.1-70B -- because decode time scales with activated weights, not total parameters. MoE dispatch overhead is excluded, so this is an upper bound.

Xunzhuo Liu, Bowei He, Xue Liu et al.

Retrieval-augmented generation (RAG) is increasingly deployed in enterprise search and document-centric assistants, where responses must be grounded in long and complex source materials. In practice, verifying that generated answers faithfully reflect retrieved documents is difficult: large language models can check long contexts but are too slow and costly for interactive services, while lightweight classifiers operate within strict context limits and frequently miss evidence outside truncated passages. We present the design of a real-time verification component integrated into a production RAG pipeline that enables full-document grounding under latency constraints. The system processes documents up to 32K tokens and employs adaptive inference strategies to balance response time and verification coverage across workloads. We describe the architectural decisions, operational trade-offs, and evaluation methodology used to deploy the verifier, and show that full-context verification substantially improves detection of unsupported responses compared with truncated validation. Our experience highlights when long-context verification is necessary, why chunk-based checking often fails in real documents, and how latency budgets shape model design. These findings provide practical guidance for practitioners building reliable large-scale retrieval-augmented applications. (Model, benchmark, and code: https://huggingface.co/llm-semantic-router)

Xunzhuo Liu, Bowei He, Xue Liu et al.

Computer-using agents (CUAs) act directly on graphical user interfaces, yet their perception of the screen is often unreliable. Existing work largely treats these failures as performance limitations, asking whether an action succeeds, rather than whether the agent is acting on the correct object at all. We argue that this is fundamentally a security problem. We formalize the visual confused deputy: a failure mode in which an agent authorizes an action based on a misperceived screen state, due to grounding errors, adversarial screenshot manipulation, or time-of-check-to-time-of-use (TOCTOU) races. This gap is practically exploitable: even simple screen-level manipulations can redirect routine clicks into privileged actions while remaining indistinguishable from ordinary agent mistakes. To mitigate this threat, we propose the first guardrail that operates outside the agent's perceptual loop. Our method, dual-channel contrastive classification, independently evaluates (1) the visual click target and (2) the agent's reasoning about the action against deployment-specific knowledge bases, and blocks execution if either channel indicates risk. The key insight is that these two channels capture complementary failure modes: visual evidence detects target-level mismatches, while textual reasoning reveals dangerous intent behind visually innocuous controls. Across controlled attacks, real GUI screenshots, and agent traces, the combined guardrail consistently outperforms either channel alone. Our results suggest that CUA safety requires not only better action generation, but independent verification of what the agent believes it is clicking and why. Materials are provided\footnote{Model, benchmark, and code: https://github.com/vllm-project/semantic-router}.

Chen Wang, Xunzhuo Liu, Yue Zhu et al.

LLM serving systems process heterogeneous query workloads where different categories exhibit different characteristics. Code queries cluster densely in embedding space while conversational queries distribute sparsely. Content staleness varies from minutes (stock data) to months (code patterns). Query repetition patterns range from power-law (code) to uniform (conversation), producing long tail cache hit rate distributions: high-repetition categories achieve 40-60% hit rates while low-repetition or volatile categories achieve 5-15% hit rates. Vector databases must exclude the long tail because remote search costs (30ms) require 15--20% hit rates to break even, leaving 20-30% of production traffic uncached. Uniform cache policies compound this problem: fixed thresholds cause false positives in dense spaces and miss valid paraphrases in sparse spaces; fixed TTLs waste memory or serve stale data. This paper presents category-aware semantic caching where similarity thresholds, TTLs, and quotas vary by query category. We present a hybrid architecture separating in-memory HNSW search from external document storage, reducing miss cost from 30ms to 2ms. This reduction makes low-hit-rate categories economically viable (break-even at 3-5% versus 15-20%), enabling cache coverage across the entire workload distribution. Adaptive load-based policies extend this framework to respond to downstream model load, dynamically adjusting thresholds and TTLs to reduce traffic to overloaded models by 9-17% in theoretical projections.

Xunzhuo Liu, Bowei He, Xue Liu et al.

System-level routers that intercept LLM requests for safety classification, domain routing, and PII detection must be both fast and operationally lightweight: they should add minimal latency to every request, yet not require a dedicated GPU -- an expensive resource better used for LLM inference itself. When the router co-locates on the same GPU as vLLM serving instances, standard attention's $O(n^2)$ memory makes long-context classification (8K--32K tokens) impossible: at 8K tokens, three concurrent classifiers need ${\sim}$4.5\,GB for attention masks alone, far exceeding the memory left by vLLM. We present three staged optimizations for the vLLM Semantic Router, benchmarked on AMD Instinct MI300X, that solve both the latency and the memory problem. \emph{Stage~1}: a custom CK Flash Attention operator for ONNX Runtime on ROCm reduces attention memory from $O(n^2)$ to $O(n)$ and end-to-end (E2E) latency from 4{,}918\,ms to 127\,ms (\textbf{38.7$\times$}), enabling 8K--32K tokens where SDPA OOMs. \emph{Stage~2}: classical NLP prompt compression (TextRank, position weighting, TF-IDF, and novelty scoring) reduces all inputs to ${\sim}$512 tokens without neural inference, capping both latency and GPU memory at a constant regardless of original prompt length (E2E 127$\to$62\,ms, \textbf{2.0$\times$}). \emph{Stage~3}: near-streaming body processing with adaptive chunking and zero-copy JSON eliminates serialization overhead (E2E 62$\to$50\,ms, \textbf{1.2$\times$}). Cumulatively: \textbf{98$\times$} improvement (4{,}918\,ms to 50\,ms), 16K-token routing in 108\,ms, and a total router GPU footprint under 800\,MB -- small enough to share a GPU with LLM serving and removing the need for a dedicated accelerator. Stage~1 targets AMD ROCm (NVIDIA GPUs already have FlashAttention via cuDNN); Stages~2 and~3 are hardware-agnostic.

Huamin Chen, Xunzhuo Liu, Junchen Jiang et al.

Semantic routers in LLM inference gateways select tools in the critical request path, where every millisecond of added latency compounds across millions of requests. We propose Outcome-Aware Tool Selection (OATS), which interpolates tool embeddings toward the centroid of queries where they historically succeed -- an offline process that adds no parameters, latency, or GPU cost at serving time. On MetaTool (199~tools, 4,287~queries), this improves NDCG@5 from 0.869 to 0.940; on ToolBench (2,413~APIs), from 0.834 to 0.848. We also evaluate two learned extensions: a 2,625-parameter MLP re-ranker and a 197K-parameter contrastive adapter. The MLP re-ranker hurts or matches baseline when outcome data is sparse relative to the tool set; the contrastive adapter provides comparable gains on MetaTool (NDCG@5: 0.931). All methods are evaluated on the same held-out 30\% test split. The practical takeaway is to start with the zero-cost refinement and add learned components only when data density warrants it. All mechanisms run within single-digit millisecond CPU budgets.

Xunzhuo Liu, Bowei He, Xue Liu et al.

Computer Use Agents (CUAs) translate natural-language instructions into Graphical User Interface (GUI) actions such as clicks, keystrokes, and scrolls by relying on a Vision-Language Model (VLM) to interpret screenshots and predict grounded tool calls. However, grounding accuracy varies dramatically across VLMs, while current CUA systems typically route every action to a single fixed model regardless of difficulty. We propose \textbf{Adaptive VLM Routing} (AVR), a framework that inserts a lightweight semantic routing layer between the CUA orchestrator and a pool of VLMs. For each tool call, AVR estimates action difficulty from multimodal embeddings, probes a small VLM to measure confidence, and routes the action to the cheapest model whose predicted accuracy satisfies a target reliability threshold. For \textit{warm} agents with memory of prior UI interactions, retrieved context further narrows the capability gap between small and large models, allowing many actions to be handled without escalation. We formalize routing as a cost--accuracy trade-off, derive a threshold-based policy for model selection, and evaluate AVR using ScreenSpot-Pro grounding data together with the OpenClaw agent routing benchmark. Across these settings, AVR projects inference cost reductions of up to 78\% while staying within 2 percentage points of an all-large-model baseline. When combined with the Visual Confused Deputy guardrail, AVR also escalates high-risk actions directly to the strongest available model, unifying efficiency and safety within a single routing framework. Materials are also provided Model, benchmark, and code: https://github.com/vllm-project/semantic-router.

Xunzhuo Liu, Hao Wu, Huamin Chen et al.

Conflict detection in policy languages is a solved problem -- as long as every rule condition is a crisp Boolean predicate. BDDs, SMT solvers, and NetKAT all exploit that assumption. But a growing class of routing and access-control systems base their decisions on probabilistic ML signals: embedding similarities, domain classifiers, complexity estimators. Two such signals, declared over categories the author intended to be disjoint, can both clear their thresholds on the same query and silently route it to the wrong model. Nothing in the compiler warns about this. We characterize the problem as a three-level decidability hierarchy -- crisp conflicts are decidable via SAT, embedding conflicts reduce to spherical cap intersection, and classifier conflicts are undecidable without distributional knowledge -- and show that for the embedding case, which dominates in practice, replacing independent thresholding with a temperature-scaled softmax partitions the embedding space into Voronoi regions where co-firing is impossible. No model retraining is needed. We implement the detection and prevention mechanisms in the Semantic Router DSL, a production routing language for LLM inference, and discuss how the same ideas apply to semantic RBAC and API gateway policy.

Xunzhuo Liu, Bowei He, Xue Liu et al.

Production AI agents frequently receive user-specific queries that are highly repetitive, with up to 47\% being semantically similar to prior interactions, yet each query is typically processed with the same computational cost. We argue that this redundancy can be exploited through conversational memory, transforming repetition from a cost burden into an efficiency advantage. We propose a memory-augmented inference framework in which a lightweight 8B-parameter model leverages retrieved conversational context to answer all queries via a low-cost inference path. Without any additional training or labeled data, this approach achieves 30.5\% F1, recovering 69\% of the performance of a full-context 235B model while reducing effective cost by 96\%. Notably, a 235B model without memory (13.7\% F1) underperforms even the standalone 8B model (15.4\% F1), indicating that for user-specific queries, access to relevant knowledge outweighs model scale. We further analyze the role of routing and confidence. At practical confidence thresholds, routing alone already directs 96\% of queries to the small model, but yields poor accuracy (13.0\% F1) due to confident hallucinations. Memory does not substantially alter routing decisions; instead, it improves correctness by grounding responses in retrieved user-specific information. As conversational memory accumulates over time, coverage of recurring topics increases, further narrowing the performance gap. We evaluate on 152 LoCoMo questions (Qwen3-8B/235B) and 500 LongMemEval questions. Incorporating hybrid retrieval (BM25 + cosine similarity) improves performance by an additional +7.7 F1, demonstrating that retrieval quality directly enhances end-to-end system performance. Overall, our results highlight that memory, rather than model size, is the primary driver of accuracy and efficiency in persistent AI agents.

Huamin Chen, Xunzhuo Liu, Bowei He et al.

The Semantic Router DSL is a non-Turing-complete policy language deployed in production for per-request LLM inference routing: content signals (embedding similarity, PII detection, jailbreak scoring) feed into weighted projections and priority-ordered decision trees that select a model, enforce privacy policies, and produce structured audit traces -- all from a single declarative source file. Prior work established conflict-free compilation for probabilistic predicates and positioned the DSL within the Workload-Router-Pool inference architecture. This paper extends the same language from stateless, per-request routing to multi-step agent workflows -- the full path from inference gateway to agent orchestration to infrastructure deployment. The DSL compiler emits verified decision nodes for orchestration frameworks (LangGraph, OpenClaw), Kubernetes artifacts (NetworkPolicy, Sandbox CRD, ConfigMap), YANG/NETCONF payloads, and protocol-boundary gates (MCP, A2A) -- all from the same source. Because the language is non-Turing-complete, the compiler guarantees exhaustive routing, conflict-free branching, referential integrity, and audit traces structurally coupled to the decision logic. Because signal definitions are shared across targets, a threshold change propagates from inference gateway to agent gate to infrastructure artifact in one compilation step -- eliminating cross-team coordination as the primary source of policy drift. We ground the approach in four pillars -- auditability, cost efficiency, verifiability, and tunability -- and identify the verification boundary at each layer.