Deploying heterogeneous AI models concurrently on a shared GPU introduces resource contention that complicates runtime scheduling. While surrogate models avoid costly online benchmarking, their profiling requirements typically grow combinatorially with the number of co-running models, limiting scalability. We propose a MeanField surrogate that predicts per-model performance from local configuration and aggregate GPU state rather than explicitly modeling all joint interactions. Experiments on concurrent LLM and vision workloads across $N \in \{2,3,4,5,6\}$ show high predictive accuracy ($R^2 \approx 0.96$) with an empirical sample budget that grows approximately linearly in $N$, in contrast to the combinatorial cost of fully joint profiling. Integrated into a genetic algorithm scheduler, the surrogate scales to an $N=5$ problem with 78,732 feasible joint configurations, remaining within 0.10% of the exhaustive search with zero SLA violations across eight dynamic workload scenarios, while complete online GA decisions take 26 ms median, about $5\times$ faster than exhaustive surrogate search.
Hybrid linear-attention architectures have recently scaled to large open-weight models, offering quality competitive with full attention while substantially reducing key/value (KV) cache growth. However, their in-place recurrent-state updates complicate cache management: prefix reuse requires state checkpoints alongside full-attention KV, while storing state checkpoints in full increases memory pressure, leading to more evictions and repeated prefill. By analyzing the decay structure of Gated DeltaNet (GDN) and Kimi Delta Attention (KDA), we find that different heads and channels retain prefix information over markedly different timescales, which we term \emph{retention horizons}. This variation suggests substantial compression potential in persistent state checkpoints. Building on this observation, we introduce \emph{Decay-Aware State Compression} (DASC), which derives retention horizons from model weights, selects long-horizon state units, and packs them into a ragged state checkpoint layout. To integrate efficiently with tensor-parallel inference engines, DASC furtherly balances compressed state checkpoints across TP ranks. On reuse, DASC either zero-fills omitted units or refreshes them from a bounded suffix with additional compute cost. Across retrieval and end-to-end reasoning benchmarks on Kimi-Linear, conservative DASC configurations remain close to full caching while compressing KDA recurrent state checkpoints by $2.63\times$. Under fixed state checkpoint memory budgets, the resulting capacity gains reduce mean Time to First Token (TTFT) by 42.6\% and improve input throughput by 68.4\%. At larger compression ratio, suffix refresh recovers much of the accuracy lost to more aggressive omission, at the cost of additional replay computation. Qwen with GDN exhibits a similar quality--efficiency trend, showing that DASC extends from channel-wise KDA to head-wise GDN.
Reported serving speedups from quantized kernels typically bundle the weight format, the kernel, and the inference runtime into one number. We present an attribution study on four NVIDIA RTX A5000 GPUs, 24 GiB each, on a single host with NVLink-bridged pairs. A matched intermediate stack that keeps the faster runtime without the quantized kernel splits the full speedup into a runtime part and a kernel and quantization part. Under matched greedy decoding the full stack reaches $2.58\times$ end to end, with the runtime change accounting for about two thirds of that gain on a logarithmic scale; across three similar model families the kernel and quantization part moves by at most 1.5%. Sharding one instance across all four cards falls well below doubling: a profiler trace attributes about 80% of the per token shortfall to coordination, and an NVLink versus PCIe control on the same hardware shows similar realized bandwidth on both links, pointing away from link bandwidth as the cause. Whether to run one sharded instance or several independent ones depends on the workload and the model, with the ranking reversing on the larger model: the smaller model splits between sharding and multiple instances by workload, while the larger model favors two paired instances on every workload. Quantization extends sustainable concurrent users roughly four times past a reproducible half precision memory cliff. Differences in sampling mode and prompt pool between the two stacks are documented as threats to validity.
Large Language Models are reshaping how software is developed and maintained. They are typically deployed in production using inference engines such as vLLM, which can efficiently serve pre-trained, highly configurable models. While prior work has focused on model architectures and hardware acceleration, the impact of inference engine configuration on energy consumption, performance, and output quality remains poorly understood. In this paper, we present a large-scale controlled study of three selected vLLM configuration options: attention kernel type, prefix caching, and chunked prefill. We evaluate all combinations of these configurations across 5 open-weight LLMs and 5 diverse inference tasks, totaling $9,000$ runs and $93,600$ measures. We analyze energy consumption, latency, and accuracy, and examine both main effects and interaction effects between configuration options and tasks. Our results show that the studied configuration options significantly impact energy and performance, mainly driven by attention type and prefix caching, while chunked prefill has a limited effect under the default vLLM serving configuration and evaluated workloads. These effects are highly model- and workload-dependent, and no configuration is universally optimal. We further show that model choice dominates global trade-offs, while configuration tuning provides local improvements along the Pareto frontier. Unexpectedly, inference options can also affect model accuracy.
Retrieval-augmented and agentic workloads repeatedly prefill recurring predictable structured inputs (which we call "spans") such as documents and code files. Yet, prefix caching in engines such as vLLM cannot reuse their KV entries unless they share identical prefixes with another request, while Position-Independent Caching (PIC) implementations within production-grade inference servers typically either require substantial server code changes or keep KV state outside the server, incurring host-to-device transfer overhead. We present Minimalistic PIC (MiniPIC): a minimal, flexible and fast vLLM design built from two ingredients: positional-encoding-free KV cache and user-controlled cache-reuse primitives. MiniPIC stores unrotated K vectors in the KV cache, applies RoPE to K tiles inside attention using per-request logical positions, and exposes three user-facing and token-level primitives: block-aligned padding, span separator (SSep), and prompt depend (PDep), that modify hashing behavior and effective block-level causal attention structure. With fewer than 100 lines of core-engine changes plus a custom attention backend, these primitives are sufficient to realize multiple PIC methods, including Block-Attention, EPIC, and Prompt Cache, within the same running vLLM instance, while natively integrating with KV cache CPU offload implementations. On 2WikiMultihopQA, MiniPIC with interleaved scheduling improves prefill throughput by 49% over baseline vLLM, reduces cached-span time-to-first-token by up to two orders of magnitude, preserves the linear prefill scaling of uncached spans, and incurs only 5.7% worst-case overhead.
Huzaifa Shaaban Kabakibo, Animesh Trivedi, Lin Wangcs.LG
As scalable inference services become popular, the cold start latency of an inference engine becomes important. Today, vLLM has evolved into the de-facto inference engine of choice for many inference workloads. Although popular, due to its complexity and rapid evolution, there has not been a systematic study on the startup latency of its engine. With major architectural innovations under it (e.g., the V1 API, introduction of torch.compile), in this paper, we present the first detailed performance characterization of vLLM startup latency. We break down the startup process into six foundational steps and demonstrate that this process is predominantly CPU-bound. Each step exhibits consistent and interpretable scaling trends with respect to model- and system-level parameters, enabling fine-grained attribution of latency sources. Building on these insights, we develop a lightweight analytical model that accurately predicts vLLM's startup latency for a given hardware configuration, providing actionable guidance for resource planning in large-scale inference environments. All our benchmarking datasets, analysis tools, and prediction scripts are open-sourced at https://github.com/upb-cn/vllm-startup-profiler
Machine learning (ML) inference serving systems host deep neural network (DNN) models and schedule incoming inference requests across deployed GPUs. However, limited support for task prioritization and insufficient latency estimation under concurrent execution may restrict their applicability in on-premises scenarios. We present \emph{Strait}, a serving system designed to enhance deadline satisfaction for dual-priority inference traffic under high GPU utilization. To improve latency estimation, Strait models potential contention during data transfer and accounts for kernel execution interference through an adaptive prediction model. By drawing on these predictions, it performs priority-aware scheduling to deliver differentiated handling. Evaluation results under intense workloads suggest that Strait reduces deadline violations for high-priority tasks by 1.02 to 11.18 percentage points while incurring acceptable costs on low-priority tasks. Compared to software-defined preemption approaches, Strait also exhibits more equitable performance.